2025 Remote Work Conference

- Well hello everyone. Thank you so much for inviting me. So I'm excited to present this project to this room of experts on the work from home. So today I'm going to present my work that I started during my PhD at the University of Pittsburgh and that I continuing now at the University of Lauan. So I guess like all of us in this room, I was initially motivated by the persistent post pandemic shift to work from home that happened in the US and across the world. And in my project, I'm interested in looking, interested in looking at the implications of the shift to, for policy in particular looking at the policies like national progressive income tax and the restrictive housing regulations. And these are the policies that have been documented to hurt the aggregate productivity because they disincentivize workers from working for the most productive firms in the most productive cities. Restrictive housing regulations just physically limit the amount of space that is available for workers in those high productivity cities and national progressive income tax taxes. Those workers that earn higher wages in those most productive cities at a higher rate, again disincentivizing them for working in those productive cities. So in this project, I want to look how this shift, this dramatic shift to work from home affects those policy distortions. So in principle it can exacerbate or alleviate those distortions. And the reason why it can help with the aggregate productivity, I guess it's more straightforward, firms now have this work from home option that allows them to substitute expensive in person workers for cheaper workers that work remotely because those workers don't require office space. And also they can probably accept some wage cut because they can view the remote option as an amenity for them. In addition, the firms in most productive places now do not, are not constrained by the local labor market they can hire from all over the US economy. So they get access to this national pool of remote workers that effectively sustains this winner takes all effect where the labor is reallocated from the less productive firms to the more productive firms in high productivity cities. And that increases the a realized output in the economy. However, the same policies can also suppress the supply of remote workers and in-person in most productive cities. Thus putting a break on the winner takes all effect. So whereas the out the realized output goes up because of the winner takes all effect, the potential output that take economy can achieve in the work from home world can increase by even more than the realized output. That's increasing the gap between the potential and the realized output, increasing the size of the distortions. So it can go either way and this is why I want to build the spatial equilibrium model that will allow me to run the policy experiments calibrated to the US data. And after running those experiments, I find that in the long run, the shift to work from home increases the productivity loss due to tax exclusivity. But the effect of housing regulations is ambiguous depending on the specification of the model. So to give you some concrete numbers, I find that moving to flat income tax schedule from the current tax schedule in the US increased output by an extra one percentage point in the world with work from home compared to 2019, which in relative numbers means that 57% increase in output. But because we know that work from home access is not even across different sectors of the economy, for example, workers in sectors like information or finance are more likely to work remotely than workers in sectors like warehousing or transportation due to the physical nature of those sectors. There's also this inequality component to the work from home shift. And what my model shows is that basically in the world of work from home, this equity efficiency trade-off is exacerbated compared to the 2019 economy, meaning that the size of productivity loss increases. But if it will try to remove that productivity loss will increase in equality by more. So if you want to decrease in equality, you will have to more productivity and vice versa in the work with work from home. So I contribute to mainly three different strands of literature. Of course there has been already a lot of papers written on the spatial implications of work from home, many of which have been presented here. There is also a vast literature in macroeconomics talking about the spatial misallocation due to tax productivity or housing supply restrictions. And I also contribute to the discussion of the superstar cities and firms where this winner takes all effect here is basically facilitated by the work from home. So it's another channel after which the women takes all effect can operate. So this data I think all of us have seen already, but just two main points is that share of remote workers have drastically increased since 2019 and that this share has stabilized since then despite all the recent pushback from some firms and the US administration even in the most recent data from October, 2025. And the second point that you can see here is that the access to remote work has not been equal for workers, say with university degree and without one. The second fact that is important for my paper is that while the share of remote workers and their wages have been stable for a few years now the fact that there you, you can, you can think that the transition to this, to this new work from home equilibrium has not been completed yet. And one margin along which I can see that is the changes in residences. Basically how much people move. And you can see in this graph that people who work remotely are much, much more likely to move than those people who don't. And while before the 2019, before, before the pandemic, this, these shares have been relatively the same after the pandemic is a huge jump and the seriouses have not converged back to being equal yet. So there is still some transition going on in the economy and therefore it is important to model this work from home shift as opposed to just taking the new equilibrium from the data. So I guess for the interest of time, for the sake of time, I will not go through the full model, but let me just give you a brief overview on the building pieces. So as I mentioned because there is this inequality component in access to work from home, I'm going to have two worker types, exogenous worker types in my model, lower skilled and higher skilled workers, I'm going to have many different locations characterized by production and residential amenities, which are going to have the exogenous part and the endogenous part or the agglomeration effects. Each location is going to be characterized by its own housing supply function, which is going to be subject to the local housing supply regulations. And there's going to be a matrix of commuting costs between physical locations. So if you choose work one and live in one in another, you're going to pay commuting cost. And in addition, in each location you're also going to have the location specific cost of working remotely. I'm going to model the market for remote labor. This is the crucial PMA model as the national market for remote labor. So basically firms in San Francisco do not necessarily care if the worker is in San Francisco or Atlanta in Pittsburgh or New York. And then I'm also going to have standard local markets for in-person labor households and a model are going to choose in which location to live and in which location to work. Where I code the remote work as option zero, like this national market for remote work and the physical locations are all the other allocations. There's going to be production of the consumption good in my model, which is going to combine the floor space and the exogen, sorry, and the remote and in-person labor of both types. And this good is going to be freely traded across the economy. And then finally, and crucially, I'm going to have distortions in my model, the endogenous supply of housing constrained by local regulations. So basically this is going to be captured by the housing supply elasticity parameters at the local level. And I'm going to have progressive income tax at the national level that funds redistribution between households with different wage levels as well as some national government spending. Any questions about the model? So just to be precise, the two aggregate, I'm going to concentrate on the aggregate outcomes in my model. And the two things I'm looking at in particular is the GDP, which in my model is the total wage bill plus the earnings of absentee landlords and the exempt expected welfare of workers of which type. So I'm going to have different, I'm going to do different kinds of experiments in my model. The first experiment I do is the work from home shock and motivated by the facts that the share of remote workers has increased and stabilized and the workers who work remotely have been experiencing STA stagnant wages. I'm going to calibrate the work from home shock to match these facts by increasing the productivity of remote work for firms and simultaneously decreasing the cost of remote work for households. So to generate this increase in quantity of remote work while keeping the wages of remote work the same and after I obtained my counterfactual work from home economy are going to run two types of policy experiments with the original economy in 2019 and the counterfactual economy that fully adjusted to the work from home shock, I can change the degree of textbook relativity and compare the two economies, or I can change local housing supply elasticity and again, compare the two economies. So I take some elasticities from the existing literature and then I calibrate the economy to the 2019 data from the A CS, which gives me the residents workplace metrics, wages of remote workers and in-person workers by location and local rents. Each location is going to re to correspond to the lowest level at which the wage data is available. So one workplace puma, which is roughly equivalent to one county in the census data. And then to calibrate the CRU home shock, I'm going to take the 2022 national shares of wages, sorry, national shares of remote workers and their wages. And the final piece is to take the housing supply elasticities at the local level from null. So before I preview the results of my policy experiments, I want to say that just the work from home adjustment itself, so the shift to work from home increases the output in the economy in my model by roughly three percentage points and also increases the inequality by roughly five percentage points where most of this comes from the increase in the exempt expected welfare of higher skilled workers. So lower skilled workers do not lose much in my model from the shift to work from home, but the high skilled workers gain. And then on top of these changes, I'm going to run the tax productivity and housing supply elasticity experiments. So on this graph I'm going to plot the results of the progressivity experiments where on the Y axis you have the change in aggregate output or GDP on the X axis you have the degree of tax productivity where point 18 corresponds to the current degree of progressivity in the us The blue line is the experiment run with the baseline 2019 economy. And the red line is the experiment run with the work from home economy that has fully adjusted to the work from home shift. So you can see that I normalize the change in output to one, despite that the work from home itself has brought some initial increase in output. And then I'm going to look at the changes in output on top of the initial jump in GDP. So as I move the economy from the initial degree of tax exclusivity to the flat income, flat flat tax income schedule, which corresponds to the tax pro safety parameter equal to zero, you can see how output increases in both economies, but it increases by much more in the work from home economy. So there's this additional gain in output in the world with work from home, basically meaning that tax exclusivity in the world with work from home has become more costly. So next I'm trying to decompose this gap into different channels. Basically the gap, the product, the output can increase when the most productive firms are going to hire more in-person workers as well as they're going to hire more remote workers. So to decompose the full effect, I'm going to fix some outcomes in my model and rerun the experiment after fixing some outcomes. So in the to to extract the direct effect on an of in person work on the full effect, I'm going to fix the number of remote workers in the economy and only allow the number of in-person workers to change. Whereas to decompose the part of the full effect that this driven by the work from home workers, I'm going to fix a number of in-person workers in the most productive cities while allowing the number of remote workers to change. And so you can see that this one percentage point gap between the 2019 economy and the 2019 economy with the work from home is generated by this additional direct effect of work from home on productivity. I, I can also talk a little bit about, I don't think I have enough time, but I can talk about the result of the experiment where I changed the housing supply elasticities in the most constrained places to be at the medium level. And for those experiments I also have some gap, but as I said, that will not be robust to different specifications of the model. And the reason for that is because whereas tax progressivity directly affects both the supply of in-person and remote workers, the housing regulations only directly affect the supply of in-person workers. They do not imp impact the supply of remote workers. So this is why this result will be ambiguous, but this result is going to be robust different specifications because this part is always going to add additional productivity gains. And then I'm only have a few minutes I'm going to, I want to discuss the welfare implications of the progressivity experiments. So once again, I have the degree of progressivity on the X axis, the change in exam to welfare on the Y axis and the blue lines are corresponding to the 2019 economy and the red lines corresponding to the welcome from home economy. So this is the change in welfare for the college educated workers and this is the change in welfare for non-college educated workers. And as we reduce tax aggressivity to flat income taxation, of course we're going to have less redistribution. So less skilled workers are going to suffer because of that. But then you can also see that there's an additional gap that arises in the world with work from home. So this effectively means that if you want to reduce the inequality in the world with work from home, you will have to hurt the productivity more if you're going to do it through tax progressivity for progressive taxation. And I'm also trying to decompose where this difference in welfare is coming from. So you can have different channels for this. You can have the effect of housing prices, you can have the effect of wages. And in my experiments I mainly find that the wage channel is the most prominent. So basically what happens in my model is that with the access to work from home, the the higher skilled workers in less productive places can switch from local production to the most productive firms in the most productive cities, thus hurting the production at the less productive places. As that happens, there is a reduction in the demand for lower skilled work as well because the high skilled and low skilled work are low skilled labor are complimentary to an extent. So there is a fall in demand for remote workers and that pushes their wages down. And where the work from home is important here is that less skilled workers cannot escape this reduction wages and employment opportunities because they have less access to remote work compared to higher skilled workers. And that is the main reason for this additional gap in inequality. So to conclude, I want to, so I'm, I'm in this paper, I'm studying the dramatic shift to work from home post the pandemic. And I find that it exacerbates the spatial distortions due to tax productivity by suppressing the winner Taxol effect. But the effect of spatial distortions due to housing regulations is ambiguous. However, if we simply flatten the tax schedule by going to generate a mechanical increase in equality, which is now exacerbated by the fact that low skilled workers have less, less access to work from home. So altogether the work fromm home intensifies the equity efficiency trade off in the spatial setting. So to get more productivity, you have to increase the equality more and vice versa. And this essentially underscores the need from one nuanced tax policy. Yeah, that's all I have. Thank you.

- Great, thanks. Thanks very much. So first question would be, I think there are th I could see three channels on increasing inequality, which I guess one or maybe two I thought of, but one I totally hadn't. So one is that basically higher educated people are more able to work from home so they get an amen. I think you mentioned it earlier, it's just so that comes up all the time. The second is I hadn't thought of as much, which is that they also get better access to labor markets. So if you're higher educated, you're already earning more, but you can also access a much bigger labor market, which on average is better for them. The third thing I guess I hadn't thought of at all, which is it also makes the kind of progressive progressivity, it makes it more costly to have high taxes 'cause labor's come more mobile.

- Yes. - And therefore that's gonna push down tax rates, which again, if you think of part of that as redistribution. And so I dunno, somehow tax has to be raised, but you raise it in different ways less than labor income and more say on sales taxes or something. But all three of them push in the same direction, I think, in your model. So

- Yes, I think the, the fact that basically, this is another implication I want to, I can mention in my papers that yes, high skilled workers now become more sensitive to taxation. I wanted to have, I wanted to look at this from the local perspective, from the local taxation, but the data for that is a bit harder to get. So I'm looking at it from the national perspective. I, I dunno how to like tease out these effects in, in, in, in, in my model because I can fix certain outcomes in the model I like if you, I have to think a little bit more about how I can tease out it in the, in the model.

- There's some evidence on the, on the final one, by the way, already places like Tulsa remote,

- Yes.

- Quite tax, yes. But you're basically paying a subsidy. It's not that oddly enough, it's not Tulsa, it's some charity, but you're paying a subsidy to attract very footloose, high skilled labor and there's a bunch of places that are now having these special zones for work from home workers. So it's kind of ha it's not like the state of California isn't really changing it as of yet, but it is starting to happen. Yeah,

- Thanks. There's already a lot going on in your model, but there's, as I understand it, there's a lot left out too that affects the productivity calculation, the output calculations in particular. So let me just mention a couple three, one Nick already alluded to, but in which is that the scope for better matching in the labor market is improved among those workers who can work remotely essentially because the market's now national not localized, that reduces matching frictions, but more importantly it untethered location from workplace. And so each firm that hires remote workers is now can search in a national market and find the best suited workers. So that's one second you don't talk about this, but it seems to me important cities also differ in the, in the efficiency with which they transform tax revenues into publicly generated consumer amenities. Okay, so some, we, we live near a city that is famously suffered through a period of poor public governance. That's an, an extreme example of a city that doesn't effectively translate tax revenues into consumer amenities. Once you free up the ability of people to detach location, residential location from work workplace location, there's a, there's a welfare benefit from that immediately because people can now reallocate to those cities that are more efficient in transforming tax collections into consumer amenities. But second, there's likely there there's much greater spatial intensity of the competition among cities for workers, for residents, excuse me, as a consequence. And if you think that comp spatial competition among cities leads them to improve their transformation of tax collections into consumer amenities, you'd get another benefit from that. So I'm, I see all this not as a criticism of your paper, but just to note that the full analysis of those two big questions you set out to ask the output effects and the welfare effects involves many elements, some of which are in your model and some of which aren't.

- This is me. Alright. So yeah, great paper. I had a lot of interesting, it, it triggered a lot of interesting ideas. I was wondering, so if I think about your model more generally, this sounds sort of like the Enrico Moretti idea of like you have the like coders downtown and they have to buy lunch and they have to buy haircuts and that's bills over into the low, into the low wage workers who now get high wages because they get to hang around these coders who need to get their lunches locally. And, and I think the, if I understand correctly, the welfare negative welfare effects here come from the fact that if the coders now all move to Boulder, Colorado, no one is in downtown San Francisco buying haircuts and sandwiches. And that's sort of a bad, that's bad for the, the local low wage workers who can't move to those places. I feel like in this setting, a lot of the, the benefits are gonna come from how you model agglomeration densities. Right. I I think the way I understand, I think you mostly focus on density as a driver of conglomeration in the paper. Is that, is that right?

- Yes. - Because it seems like, it seems like one thing that's gonna be crucial here is to understand the impact of work from home is whether or not physical location still remains the most important driver of agglomeration effects, right? Like, so somehow I can project all my work productivity across the globe, but my ideas all like fall into the forest outside my house and I can't actually talk to anyone about them. I think that's sort of how, how you're currently modeling it. But it would be interesting to see what happens if you think of agglomeration also as being, having a remote component in some ways and no longer being anchored to the physical location as much. Because I think that might alleviate some of the negative effects because if I don't need to be physically next to someone to benefit from their ideas, that makes it less negative that that not everyone is hanging out at the same coffee shops downtown.

- Yeah, I think I just wanted to emphasize I guess is that in, in my model, elimination effects are even increase in the work with work from home because the expansion of most productive cities is going to drive in more in-person workers in there. So of course this will depend on the parameters that I have, but effectively in my model, work from home increases those most productive cities to the detriment of the less productive places, even in terms of density.

- Thanks so much. I'm gonna keep us on time. Thank you for the questions.

- Thank you so much for having me here. I'm just waiting for the slides to load. So my name is ysu, I'm currently at Southern Methodist University and so this is a joint work with Franklin Tian at UNC University of North Carolina. I feel like this is a little bit more formal than I expected, so I feel underdressed definitely, but I'll try to manage my nervousness. Okay, so I mean this city is really sorry, this paper is really about the future, like the, our view about the future of cities, right? So as we all know, COVID-19 pandemic had triggered a massive remote work revolution and you know, during the peak of the pandemic, right, about two third of all of the work hours supplied were, you know, supplied remotely. And since then that number has come down to like, you know, about 30% it has since converged, right? And so I think it, it's pretty, you know, and I think it's becoming more and more obvious to people that we are now sort of seeing a, a sort of more or less permanent shift in remote work, you know, regime. And so with the permanent adoption of remote work, you know, there's obviously a lot of labor market implications, but for me, an urban economist and a city lover, this triggered a lot of, you know, worry, right? I mean, including real estate developers, right? The people who hold these assets, they're very worried about, you know, the future of cities, right? Because, you know, based on the traditional view of city centers or urban centers, at least in the United States, urban centers traditionally viewed as a place of, you know, a destination of commuters, right? The people coming to work these, these huge office buildings. And so a side effect of remote work is that there's gonna be reduced flow of commuters into the city center and that generates a lot of worry, right? So a lot of people are very, it's like, oh, you know, the city future of the city is doomed and, and, and it's not just economists, right? If you turn on the news, you see that all the time. And so what we try to do in our paper is to kind of counter, I mean, in a very cautioned way, counter that ar our, that view. And so we're going to try to argue that instead of killing cities, remote work seems to be transforming cities more, transforming cities from a destination of commuters more into a destination of leisure travelers and or, you know, in a lose slightly looser term, a looser term, it's gonna transform cities into more of a, from more of a producer city and into a consumer city. So that's sort of the, the idea of what we're trying to show here. And then this is the empirical paper. So we're gonna present our conceptual argument. I'm gonna show you a lot of empirical evidence to support our argument. Yeah, there's a headline, tons of them, you know, and so before I, we talk about our, you know, conceptual framework, why we think remote work could actually generate a positive effect on urban economic activity. We wanna first start with the conceptual framework behind the urban Doom meal, right? What, what is actually going on within this framework. And then we're gonna try to say, okay, this framework is a little bit simplistic. So almost all of the model with remote work has some kind of a flavor of this. So basically city center is a place of, you know, tradable service production, right? These office buildings. And, and so this is where a lot of commuters coming in come in. And because of that, they also attract a lot of residents, right? To save commuting time and therefore you have the presence of both commuters and residents. And because of that, you, it generates a huge amount of urban amenities like restaurant, retail, et cetera, all of that. So it's sort of this, all of that ultimately depends on the presence of commuters, right? And then from there, you know, you build all these other economic activity and then what remote workshop does is that it exogenously reduced both the commuters and the residents that live in the city because of the need to commute. And because all of that other stuff is ultimately built on the need to commute, once you remove that, that's gonna, you know, really severely damage the economic activity in the city center. And because remote work is like a permanent shift, that decline in economic activity is likely to be permanent, right? So that is basically like the, the, the doom worldview. And so we want to show that the worldview actually depends on some very sneaky implicit assumption that we almost never think about, but they're actually quite important. And once you remove those assumptions, the effect of remote work on city centers is actually quite ambiguous and actually be positive, right? So the first assumption is that city center is inherently a center of production or tradable service production. And so, so you know, and, and, and then the amenity provision and all the other activity are sort of the byproduct andogenous byproduct. And we, as urban economist, even before the pandemic, we already have seen so many papers that have shown that city centers, especially those with very high density, seems to have inherent advantage of providing urban amenity. They're just very valuable place to go, right? People really want to go there, right? Like New York City, Chicago, downtown and stuff, places like that. And so even if the con the, the, the commuters go away, these places probably are still very, very attractive, right? And so this is what the first assumption that we relax but relaxing this one assumption is actually not enough. Why? Because there's another assumption that's very sneaky in most models, which is that amen demand come from local residents and commuters, right? So in other words, the, the, the people who visits urban amenities are the urban residents and the commuters, right? And why is this important? Because if remote work is adopted permanently, then you have a lot of people moving to the suburbs away from the, from the urban center. And that's gonna mechanically take away the foot traffic right in from the city center to the, to the, to the suburbs. And that's gonna lead to a decline in, in, in the city. A doom, a doom in the city. But this assumption is actually not, it's pretty shaky actually. If you look at the data, a lot of residents travel for urban amenities from from the suburbs. I mean even though they live in the suburbs, they travel into the city to enjoy the amenities. And so the economic, the foot traffic occurs in the city even though they live somewhere else, right? And this is actually becoming more true with remote work. And I'm gonna talk, talk a little bit more about this. And once you relax this assumption, you can see that if city center, if city center has very high exogenous amenity value, they're just very valuable and people travel for amenity, then you can see that even as people ize, right, they changed their housing location to the suburb, this could still anchor foot traffic in the city center, right? Once you relax those two assumptions, however, this is still not enough because this, even if you relax, these are two assumptions, the effect is still unambiguously negative remote works effect, right? Because even though some people wanna travel into the city center, you know, if, you know, if people are still izing, there's still gonna be some negative effect, right? And so that brings out, that's, that brings to the third sneaky assumption, which is that consumers overall pool of time for amenity is like fixed, right? And so this is, so if, if, if, if people spend a fixed amount of time for amenities, then if people move away, yeah, that's gonna drive down foot traffic in urban center. But that's also a kind of a crazy assumption, right? With if you don't have to commute and you also don't have to go into the office, I have a lot of free time, right? If I have a lot of free time, that really changes how I manage, you know, whether or not I decide to, to travel for fun, right? And so there are two effects. One is a mechanical effect. Now I spend more time for leisure. So that could boost urban amenities for traffic, right? So that's one effect. But the more important effect is that now I don't have to commute and I don't have to go into the office. I might be more tolerant to travel far 'cause I, I don't mind traveling far, right? 'cause now I don't have to commute. And what that does is that all these mediocre places I used to settle for, okay, forget about it. We're going to the real deal, we're going to the place that I really enjoy and those tend to be in these big dense amenity center, right? And so as a result, you can actually get more traffic in these high density amenity hotspot at the expense of these more mediocre ones. So that's sort of to, so you can see that the, once you relax all three assumptions under the right condition, it could actually completely change the conclusion of working from homes effect on the economic, the spatial, the spatial distribution of economic activity. So that's sort of the central thing. And so the other thing I want to mention just before we jump into the empiric stuff is that, you know, we find that one relaxing, these assumptions doesn't do very much on residential, like the donut effect, right? The donut effect is still gonna happen regardless. The biggest impact is on the prediction amenity foot traffic, right? So we're gonna predict that the humidity foot traffic is not gonna do very much at, I mean to say the least. And if the reduced commuting time actually relax people's time constraint, it's really gonna, it's gonna increase the demand for urban amenities. So that's sort of the, the, the model, okay? Just the preview of the evidence. So, and residential, you know, if you look at the residential population don't affect, you know, people are staying suburbanized, right? There is no evidence for re urbanization. I'm gonna show you more of that later. And we also, oops, oh no. And, and we also saw that these urban centers, especially these high humidity urban centers are seeing very strong recovery and humidity, foot traffic and spending activity. And more importantly, you can see that the visitors who visit these amenities, they're increasing, coming from further away. So it's like they're increasing coming from the suburbs, you as people move to the suburbs, right? And then, yeah, that's kind of the, the preview, the, the only other thing I wanna say about the, the conceptual framework, trust me this is empirical favor, is that the pandemic released two shocks, not just one shock. The, the, especially when we analyze foot traffic, it's very important to realize that. So the first shock is the remote workshop, which is more or less, oops, which is more or less a permanent shock, right? But then there's a temporary shock, which is an amenity aversion shock, right? Because there's COVID floating around, right? You don't wanna go out, right? So if people don't want to go out, that has a huge impact on where people visit, right? And so why is it important? Because during the pandemic with this amenity aversion shock, the central city, the places with very high amenity is gonna have, is gonna suffer disproportionately because of that. Why? Because ordinarily these places receive a lot of inbound visitors, right? And when people are not going out though, they are gonna suffer disproportionately. And if we erroneously think that that's because of remote work, that's, that's wrong, right? So we really have to look at what happens after the pandemic is fully over and then look at what hap the recovery trajectory and what happens there. So that kind of motivates our empirical finding. Okay? Very briefly. So we, the, the main data set that we're gonna show you today come from the, the advanced safeguard traffic data. So this is like mobile phone based data reported at by establishment at the establishment level. And we also know the na code of the establishment. So this allows us to see the commuter traffic and the, and the amenity visiting traffic, right? And this is reported monthly. I know that there's a lot of weird kinks. So we in the data, so if some of you have worked with this data, we've spent a lot of time trying to bridge those weird kinks. So if you see the smoothness, if you're weirded out, we can talk offline. And so the, the next thing is that we are also complimenting it with safe graph spending data. So this is like credit card, debit card transaction at the establishment level, but the sample is much smaller and we have some other data we can talk about it as we go. So the first thing I wanna say before looking at the traffic data itself is I wanna convince you that there is really no reverse, there's no evidence that there is systematic reverse migration back to the urban center as of recently. So this is coming from the New York Fed consumer credit panel. And this is actually coming from another paper of mine. So here we're plotting the change in the distance to downtown for all of these samples in the, in the, in the micro data. So essentially every quarter you can see where people are. And then we just take the difference between the distance to downtown that you're this, this quarter minus the distance to downtown you're of you in the previous quarter, right? So if you're didn't move, that number is gonna be zero, right? So here you can see that during the peak of the pandemic, this movement outward is really kind of going through the roof, right? So this is the donut effect and then since then that has come down. But there's no evidence that this, you know, for there to be reversal people moving back to the city, this has to go below zero, right? And this is far from zero. So this donut effect is actually still going on. So there's no evidence of that reversing itself. And you can sort of see the same thing with rent growth. So here we are doing, we, we we're plotting rent, normalized rent to 2019 level for urban center, which is five mile radius between within downtown and the suburbs. And you can see that rent trajectory in the suburbs far outpaces this, the, the, the urban center. And that gap is not closing, you know, if people are moving back, we should see that closing and that gap is actually widening and you can see the same thing with home value as well. So I hope I can, I have already convinced you that the residential is, is in terms of residents, right? They're not sort of moving back to the, to the city. But I'll show you that foot traffic pattern, very, very different. Let's just start with like the overall picture. This is American time use survey. So we're plotting the time that people spend at workplace versus time people spend at amenities. So you can see that at, in 2020 the both of them saw a huge drop, right? So a huge drop. But for, you know, people, the onsite time, people who work on the, the time people spend on time on site, on on their work site, that has pretty much been stable since then, right? This is kind of, we kind of already know that, but if you look at the, the, the amenity people spend in restaurants, you know, stores and, you know, entertainment venues that had dropped out even bigger cliff during the pandemic, but it has since, you know, been on a very fast, you know, recovery trajectory, right? So this is sort of overall picture. So here, this is from the safe graph data. So this is mobile phone visits. So this is all, all of the country, right? So we are plotting the trajectory of or amenity for traffic and commuting trips for in establishment in suburbs and the urban center. So this is like the, the, so the solid line are the amenity for traffic. And one, two things I want you to take away, right? So he, the first is that amenity foot traffic to urban center really suffered the most, but then the recovery is also the fastest, right? And this is also faster, not only relative to the amenity foot traffic in suburbs, it's also faster than the commuting trips, right? So this is sort of the first graph I wanna show you. And by the way, this is the least dramatic graph. So everything is gonna be more dramatic because if you plot the same graph with a very well known mono centric city, like New York metropolitan area, the, the, the pattern is like much, much more dramatic, right? So you can see that amenity for traffic to urban center of New York metropolitan area dropped very steeply, but then it recovered very, very fast and now it's actually totally back to normal and, and, and, and, and it's much faster than the, the recovery of commuting trips. And you can see the same thing with Chicago is pretty much exactly the same and San Francisco, okay, this is a weird place, alright? I don't know why people just don't go out. And so, but you can sort of still see the same kind of pattern, right? This big drop in urban foot traffic, urban amenity, foot traffic, and a big recovery since then. And then, you know, some people say okay, maybe this is all about long-term vacationers the tourist. So one thing that we can do is to, to split the sample into weekday visits and weekend visits, right? So here you can see that weekend visit is really leading the way, you know, into the, for the recovery of urban amenity foot traffic and especially in places like New York, right? So you can see that visit weekend visits to amenity in center city New York is actually far outpacing the pre pre pandemic level, right? And same thing in Chicago and also you might be wondering, okay, maybe this is your visiting data that's like kind of weird, right? So here we can use the spending data like the safe graph spending data. This is credit card transaction. And you can see, so here I'm just plotting the, the fraction of all of the retail spending as a, the fraction of all the retail spending that occur in establishment within the urban centers, right? Out of all the, all of them. And you can see that the fraction dropped temporarily during the peak of the pandemic and it has since recovered and, and some more, right? So this is another piece of evidence and then here is the, the visitor coming from further away evidence, right? So here we're showing you the trajectory of the mean distance of the visitors, you know, visiting the amenities in the urban center versus the suburbs. So basically the visitors to all of the amenities are coming from further away. So people seem to have a lot of time visiting like far away amenities, but particularly for amenities that are located in urban centers. Alright? Okay. So all of those very interesting I would say, but they're coming from a dichotomy, sort of a very simple urban versus suburban binary structure. So one thing that we can do is that we can zoom in finer, right? So look at, maybe look at the like midtown Manhattan versus you know, west Village, right? So if the idea if, if this recovery that we see, some people are, are concerned that this recovery that we see is really driven by commuter behavior, right? So this, this really maybe, maybe used to be that commuters don't like to go out for lunch because of COVID and now they can, right? So what we can do is just to look at neighborhoods that just offer amenities, right? People don't probably don't commute into these neighborhoods. So, so the first motivation I want to give you is this. So here we rank the entire country's neighborhood based on amenity like popularity. So this is like foot traffic density before the pandemic. And then by percentile we just plot the trajectory of the foot traffic. And you can see that the really the highest, like the most popular neighborhoods saw the biggest drop in foot traffic, but then it has since recovered spectacularly. And then what's more is that the, the foot traffic now is actually they're attracting proportionally more foot traffic than all of the other amenity, all of the other neighborhoods. I think I'm running outta time so I'm just gonna hurry up and I have a lot of stuff. So I will just show you this regression and then I'll conclude. So the idea, one thing that we can do is to separate neighborhoods into places like CBD places where there's a lot of white collar worker where we expect a lot of loss of commuters and then we can look at other neighborhoods like West Village, right? There's a lot of like popularity there. And then we, we summarize these measures into rs. This is like remote shock, you know, measure. And then so this measures how, how much more likely this is a central business district, right? And then am is like the amenity popularity and then we can run a time varying, sorry, run a regression where we allow the coefficient to be time variant. So here's the result, and this is like really interesting, right? Because here the black line is the time effect by the remote shock, right? So this basically tells you that the places where you expect a lot of, you know, commuters to disappear the foot traffic around them to amenities, the foot traffic to amenities around these neighborhoods saw a drop and then sort of just stayed there. But then the places where there's a lot of amenities to begin with, it has nothing to do with really commuting. They saw a big drop and then it kind of recovered disproportionately, right? And then we did the same thing with consumer transaction data and we found basically the same thing. So this kind of really tells you that there is this two things going on in the city, right? One is the remote workshop that people are not coming in the other is like they're really attractive amenities and people are wanting to coming back, right? So I have a lot of other stuff, but I'm really running outta time. So you kind of already know what I'm, what I've already said. So basically we're trying to counter the idea that remote work is going to kill cities. It seems like cities' role as an amenity could really is like the key to saving the city and we show quite a number of empirical findings to, to back up the story. Although there are some more I wish I could show you, but, but that's, that's all

- Very interesting paper. I'm quite sympathetic to the overall theme about cities becoming centers for consumer amenities. One thing I'd be curious to see maybe in the next paper is has there been a divergence across the cities that were major sources of consumer amenities before the pandemic? But they've had very different recoveries. You already showed a little bit of that in the contrast between San Francisco and New York. My my casual impression is that San Francisco hasn't recovered because they've had bad city governance. This is back to my point about transforming tax revenues into valued consumer amenities. And I I suspect you'll see broader patterns of that. So you showed us mostly mean conditional mean outcomes, but I suspect there's a lot of heterogeneity in the recovery of cities because they've adjusted to the shocks you talked about with different degrees of effectiveness.

- What have, okay,

- Thank you. Sorry, this is such an interesting paper. I was just curious, I, I assume you have like time of day information so, and maybe that's one of the graphs you wanted to show, but I'm curious like especially in the weekday, if people's time is more flexible, if you see, if you think the future cities will also change in the timing of the amenities they offer, be it later in the middle of the day.

- Yeah, that's a, yeah, that's a, that's a really good question. So I don't have that graph so I, I'm actually, we were actually working on that using micro data so you can actually see exactly when and where people are visiting. So that we are definitely gonna look because you know, one thing that we saw was that when people work remotely, it's the day before that they're starting to go to have fun, right? Like so I'm suspecting it has something to do with like staying up late. So, you know, so there's some, a lot of those we can explore. Yeah, I think nick that or

- Super interesting. I was gonna say just what for firms I guess one, so what is, for cities like New York, their office sector seems to be doing fine in part 'cause it's next to amenities. So you kind of want your office next to someone nice. It also means with these disused offices it's easy to repurpose the first floor. 'cause you just put a shop in is more problematic for like floor 20 where no one, so yeah. Great. Really interesting.

- Yeah, great. Thanks. I think there's some over there. Oh

- Okay, great. Yeah, thanks. Quick question and then a bigger question. So the one of the last tables that you showed with the amenity value, the time, very amenity value, what was the outcome variable in that regression? Was that like a Oh, sorry,

- Yeah, it was foot traffic. Yeah,

- Foot oh foot traffic. Okay. And so then the right hand side was amenity and then one was remote and so you showed the amenity was one. Yeah, yeah. Okay. Super. So this goes back to a good point that Steve brought up about differential public governance. So do you know the emergent MSRB muni underwriting data? Okay, well this would be a, I mean, so we'll talk more afterwards, but basically seeing every underwriting record that counties have issued. And so the yields on that will tell you something about whether or not remote work is being capitalized into kind of city infrastructure. And so you can look at like general obligation bonds, revenue, bonds, et cetera. So I think that might be a critical mechanism for uncovering how remote is being priced and how cities are adopting.

- Yeah, yeah, yeah. Well yeah, that's a, that's a good idea. Yeah, yeah. One thing we're worried, we worried about like, oh, you know, remote, yeah, all these fun, fun people are coming in, you know, but how much like tax revenue are they gonna bring? So that, yeah, absolutely. I think,

- So I had a question very related to Steve's point, which is do you have any evidence for why the dynamic now is what it is with, with cities being centers of consumption? Whereas in the mid 20th century when cities really shrank and, and, and activity moved to the suburbs, there was more of a doom loop. Is there, I mean I, I think we can tell stories about, about how the characteristics are different now and then, but, but I don't know if you have any evidence or any ideas of where we might get evidence.

- Yeah, first of all, crime was like, you know, I, I wrote a paper on gentrification like way way before this. And so that was like the, the thing, right? Like it was really dangerous. And so yeah, there was a lot of, so the other thing is that I, I I think I briefly mentioned that it's actually a really important mechanism is that remote work actually makes people want to venture out, right? And I think, I feel that other people told me that, and then we can see it in the data, right? So now people are kind like, oh, I can work remotely, let's work in that blue bottle coffee shop. And so, and then, you know, when I'm done I could just immediately jump to these fun places, right? You know, that, that a lot of that is going on. So yeah, I think that back then we didn't have that.

- Yeah. Maybe the follow up question, which you don't have to answer is, is why wasn't there more of a doom loop in other parts of the world? So say in in Latin America or in Europe during the mid 20th century, as as activity could move far away from the city center, whereas in the US there was, that's

- Yeah, yeah. Complement my previous question. I don't really have a good answer though, but that's a good question. Yeah.

- So pre COVID essentially urban centers were really good places to work and now they've essentially transitioned into really good places to live. Would that be a, a, a fair perception?

- Okay, so, okay, that's related to gentrification. So when, when I was, you know, a PhD student here, I was working on the gentrification paper and so the, the the gentrification paper, so basically gentrification was like people, like all these high, high skilled workers are moving into the city, right? And so it's sort of like, but this, we, we actually saw kind of the opposite aside from New York. New York is like the, except the, the really notable exception, right? Like every other city we saw like high skilled workers like moving out and so, but their foot traffic seems to be moving in. So it's like there's some weird thing going on with like, yeah. So

- Living by meaning amenities are there, people can go in and commute and engage with those amenities and then go back to their urban or their suburban areas when they want to. My my question is, if you look at municipalities and you had like 20 seconds, top three suggestions or top suggestion, what would you say to municipalities that are looking at these problems and trying to solve it?

- Yeah, I, I, I think, you know, so, so, okay, I'll answer it this way. So, so in the, in the theoretical model when we're trying to work out the effect of the remote work, I, I think so, so remote work. So remote work actually has a first, first order impact on where people wanna live, right? So as long as remote work keeps, keeps going, this donut effect, it, it just seems like it's, it's not gonna go away, right? And so the only re way for these urban center to like, maybe not the only way, but important ways to get these people to visit 'cause like, you know, 'cause with remote work there, there's kind of little reason for them to just live there. 'cause like commuting is like a big thing that, that determines where people live. I think I need to, you know, maybe move on to the next.

- Yeah, thanks so much. That was great. And Stephanie great. And next up is Stephanie talking on shopping.

- Okay. Hello, it's a pleasure to be here. I'm Stephanie Johnson and this paper is called Shopping From Home. It's joint work with Nick and with our co-authors, Scott and Yana. Okay, so in this paper we asked the question, how does remote remote work change shopping? And to answer this, we ran three surveys on Nielsen IQ panelists asking them about how often they work from home, both currently, so this is in 2022 and 2023. But also we asked them to recall their work modality, both during the early pandemic and before the pandemic. We also asked them questions about their shopping behavior. So whether they primarily shop for groceries in person or using an app for curbside pickup or delivery and some other questions about the use of grocery apps. We can match this survey data to very detailed data on those households purchases. And we'll use this to answer several questions. So we show that remote work does change shopping behavior in several different dimensions. So firstly, there is a strong effect of working from home on shopping from home. So people who work from home are much more likely to do their grocery shopping online. We also see that remote workers are more likely to do shopping on weekdays rather than on weekends. We also look at how this changes our product composition. So we see that households shift towards purchasing more food and away from, you know, some other products like health and beauty products. We see that overall households spend more on the products that are covered by the Nielsen data. So for those of you who have used this data, this you can think about things that you would, you know, buy in a grocery store or in a drug store, those types of products. So we see spending on those things increasing and we also see households purchasing greater quantities of those things. We also see that households buy a broader variety of products than they did previously. So in this data set, what we see is the UPC. So you can think of this sort of definition of the product as being very specific. It is like a specific product and a specific pack size of that product. So we see households purchasing substantially more distinct UPCs when they start working from home. And then we also see some fundamental shift in terms of households like price sensitivity. So when people start working from home, they actually become somewhat less price sensitive than they used to be. And we also, we see this in several dimensions. So households obtain fewer deals, they are more likely to buy somewhat more expensive products within a particular product category. And we also see that they become like less price sensitive as measured by their price elasticity. Okay, so we ran three sur, three, three waves of the survey. The first wave starts in mid 2022 and then we had two follow up waves, one later in 2022 and one in 2023. At, at each wave we get about a third of households responding, which is reasonably high. And across the three waves we actually cover around half of the households with at at least one wave households in the Nielsen panel, like stay there for quite a long time on average six to eight years. So even though we, we can only get survey responses for the panelists who were there in 2022, we actually have a lot of information on those households back prior to the pandemic as well. So this is a couple of screenshots from our survey. So the first question we ask households for each day of the week, you know, whether they were working from home or at the employer or client side or whether they were not working six or more hours. So we ask them about every single day of the week and we do this for three different time periods. So the time period you see here is the pre pandemic period. So we tell them February, 2020. We also ask them about early pandemic and then about their current work modality, which would be in 2022 or 2023 depending on the wave. We ask 'em about their shopping behavior as well. So this is their primary method of grocery shopping. So you can see the options for that are primarily in person, primarily online and use delivery and primarily online and use curbside pickup. So the survey captures pretty well the overall trend in work from home. So we see the big increase in work from home during the pandemic and also that this comes down a little bit and remains like at a, at a pretty high level since then in the paper we also do several other tests to validate the survey and we can see that, you know, so for instance we look at the a CS and look at occupation by state cells and see how our recall survey responses line up with those. And actually it lines up very well across all three periods. The pre pandemic recall, early pandemic recall and current. So overall the survey as far as work from home is concerned looks, looks pretty respectable. Okay, so then we merge the survey data with the scanner data. There are two datasets, a consumer dataset and a retailer dataset. So in the consumer dataset, basically the household is scanning their grocery purchases so you can see everything that they buy. You can see exactly what it is. You have this UPC code, how much they purchased the date of the purchase, the retailer, they purchased it from the price that they paid for it. And you can also see whether they used a coupon or whether they perceived the item to be a deal. So we also have this perceived deal indicator, but on top of this we have a retailer data set as well. And so this gives us data on posted prices for store by week by UPC and we can combine these data sets to give us a sense for, you know, not just the price of that household paid, but in principle what were the prices that were available to that household at times when they didn't buy the product and in other stores where they didn't buy the product. So our main work from home measure is constructed using up to two household heads. So Nielsen defines a male household head and a female household head. In our survey, we actually allow potentially all individuals in the household to respond. Not all of them do, but we use information on the household heads that we have data from the survey for. So our work from home measure is the sum of work from home days for the up to two household heads divided by the total days worked. And for most of the analysis we use a long run change in work from home. So this is the work from home post pandemic. So at the time of the survey minus their work from home before the pandemic, which was this, the record work from home. And because we have three waves, we actually have three data points for each of these time periods potentially if households respond to all waves. So we average the responses that we have for each household within each of those time periods. Okay, so our main empirical approach is we use this, you know, post versus pre pandemic change in work from home and we look at how, you know, various outcomes evolve over time for households who, who moved more towards work from home than others. So we have household fixed effects, time fixed effects, and we're gonna plot this estimate of the B two T. So we are tracking the dynamic response and the idea with this is that the pandemic basically gives us like the, the timing of the shift to work from home for these people. But by defining this work from home measure as being post pandemic minus pre, we're focusing on, you know, people where that work from home shift was persistent. And when I applaud the coefficients, you expect to see firstly that the, if there is an effect that the beta will change around the time of the pandemic. But another thing that we're looking for is that this effect is persistent and in pretty much all cases we see that the effects we document persist all the way through to 2023. And so this shows that what we're, what we're documenting is like a fundamental effect of work from home, not just some kind of transitory effect that is due to the pandemic. Okay, so first I'll talk about shopping modality. We see that work from home increases online shopping. There are a couple of ways that we can measure online shopping given the data we have. This chart shows effects on the share of spending at what Nielsen classifies as an online store. So the retailers in Nielsen are anonymous, but they do provide information about the type of store it is. So for example, is it a grocery store, is it a drug store, is it an online store? And so this is nice because it comes from the scanner data, which means that we, you know, we have timestamps on it and so we can show the, the pre-train and everything. The downside of this data is that it only captures spending at retailers that are kind of truly online. So thinking about Instacart and things like that. Whereas if what you you're doing is say you're using the Kroger app for curbside pickup or delivery, Nielsen is still gonna classify Kroger as grocery and from the scanner data alone, it's actually not possible to determine whether this was an online purchase or an in-store purchase. This is actually one of the advantages of our survey that it allows us to pick up that type of online purchase as well. You can see here that there is an increase in this online shopping share that does coincide with the pandemic and persist to 2023. And this, this effect is actually very large. So the baseline for this online channel is 3.4% of spending in 2019. And so what we show here is that if you shift from this work from home share of zero to one, that increases this online spending share by about one and a half percentage points. So that is a lot compared with the 3.4% initially. And so we also asked households in the survey about whether they're using the grocery apps and this just plots, you know, broadly for pre pandemic, early pandemic and currently what that looks like depending on people's work from home. So if we were to use the, like the main specification, the estimate we get here is that this form of online shopping increases by about five percentage points. So if you look at the overall pre pandemic level, about 10% for the somewhat work from home groups, about 10% of people were shopping were primarily grocery shopping online and this was about 5% for full-time in person. So this five percentage point increase is also very substantial and kind of in line with the other online shopping measure as well. And we see households also increase weekday shopping. So weekday shopping increases by about three to four percentage points. And typically people do do about two, two thirds of the shopping happens on weekdays. Okay? So looking at spending composition, we see that households spend about 10% more on the products that are covered by Nielsen. And we also see that quantity increases substantially as well Later, I'll show you that as well as the increase in quantity, they actually also pay higher prices for things too. If we look at this by product category, we can see the most pronounced effect is for food and there's really not much discernible effect for things like health and beauty products. We see some shift towards general merchandise as well, which includes things like, you know, printing and office supplies and that type of thing. And then we see that people actually buy about a 10% broader set of products. So it looks like they changed their product composition fairly substantially, but they don't actually change the number of retailers they shop at. So we didn't find any, any effect on number of retailers. And then finally we find that they become less price sensitive. So this is the Nielsen deal measure. So working from home implies about a one percentage point reduction in deals and the baseline for this is about 20%. So it's about a 5% reduction in deals. We also find that people buy more expensive products in the same category. So let me just explain exactly what this measure is. So we compute what their kind of counterfactual spending would be if they bought exactly the same set of UPCs at the annual average price for those products. That's what's in the numerator. And then we compute if they paid for each of those UPCs actually the average price of the product group in which the UPC lies and then add that up. And that is the denominator. So for example, if you are thinking about cheese, which is one of the product groups, it's things like cheese, you know, cereal, crackers, that kind of level of aggregation. If you are kind of buying the average type of cheese in terms of its price point, this measure is gonna be zero. If you are buying more expensive cheese, this is gonna be positive. And if you're buying cheaper nasty cheese, it's, it's gonna be negative. And so what this shows you is that they're basically, you know, paying 2% more, not for an identical product but for a product that is in the same like product category. Okay? And then this is a measure we construct to think about whether they're shopping at more or less expensive retailers. So the numerator is again like what their counteract spending would be if they bought the same set of UPCs at the an annual average price of the retailer that sold them that UPC and then in the denominator, this is what the counterfactual spending would be if they sold, if they bought that UPC at the kind of national retailer average price. And so it looks like they are not shopping at more expensive retailers, but they are buying more expensive products within product categories. And this table shows a, you know, a few different outcomes together. This is a somewhat different specification in which we actually control for the online shopping survey measure and where we use all of the variation in work from home. But we see very similar results here. So in particular, so column two that tells you that they're paying 0.3% more for exactly the same product. So this is like diet Coke in exactly the same pack size column three says they're paying about 1% more for a very, very specific product category. So this is something like american cheese slices and then in column four, which is also what I showed you with the other specification, this is that they're paying 1.5% more for a very similar product like cheese. Okay. And then finally we showed that work from home does reduce price sensitivity. So we compute price elasticities and to do this we need to do some aggregation because obviously we're gonna have a like a lot of zeros here. So we aggregate to work from home group by week by product. So the work from home groups here are now we say that they are switches if they increase work from home by at least 50% of days worked from pre pandemic to current. And we see that there's a pretty substantial proportional reduction in price sensitivity for those who work from home and during the post period. Something to keep in mind here is that actually a lot of the variation in like grocery prices here is the prices are typically set at the national level. So it is pretty well documented for large retailers. They kind of tend to set like basically the same price everywhere. So think about that in terms of like interpreting the price responsiveness. Okay. So we find that work from home does have substantial effects on household shopping behavior and on consumption it increases online and weekday shopping. We find they buy broader sets of products that they spend more on these grocery items and overall that they seem to be less price sensitive and may be prioritizing convenience in terms of things like their online shopping and shopping around for deals and so on. And this in turn has potential implications for retailer pricing strategies. Thank you.

- This is a wonderful paper. Thank you so much. I do have two quick questions. How should we think about sort of the welfare of having this broader variety in UPC codes, right? Is this something that like, oh I did wanna try that other type of cheese or is it, I was trying to get the same Ziploc bags but I accidentally got, you know, Ziploc extra large accident?

- Hmm, that's a good question. It's hard for us to know that given what we've done. But thanks for bringing that up because it is, you know, if one of the channels is maybe household spending less time on shopping, it could be that part of this is could be like accidental but it's not something that we've actually looked at. I

- Think the other thing that I'd love to hear about is how we can think about the less price sensitivity, right? Because you imagine that if I'm shopping online it's, it's much easier to have a tab open for Walmart and a tab open for target rather than if I'm shopping in person and I actually need to go between the stores in order to find my deal. So it seems as though we just have greater availability of information, which should make it much easier to find the cheaper price. But that seems not what's

- Around it. Yeah, that's, that's an interesting thought. So the first thing I wanna say here is that actually the price sensitivity and deal stuff is not specific to people who shop online. We also see it for people who report primarily shopping in person throughout the entire period. But like another interesting thing we ask them about the way that they use the app and one of the things we found is that when people tend to add from the saved cart, they actually end up paying more. This is something that I personally do a lot of, like I spend a lot less time shopping now and I always use the same app and I just add the same stuff from the previous orders. And when you do that you are much like, you don't look at oh what other stuff is on sale or available or you're just adding what you bought last time. So it could go either way. Yes,

- Just a point. That's a great point. So the, the surprising thing is it's about work from home. We also at the interaction with online

- That's right.

- So it it it's work from home rather than online that drives it, which is maybe surprising.

- Thank you. I just wanted to ask you whether it is the price of income effects following up on the previous question, meaning people, marginal people who are tipped into working more, I have more income and so they look for high quality varieties of goods or like the previous question I have more time to shop so I, even if I'm buying a higher quality item, I've been looking for the lowest price such item and so I still take the advantage of having, I mean lower cost at comparing alternatives.

- This is an interesting point. So I think something we would like to do is to try to look at how much time people are spending because it's also not obvious to me which way it would go because if for example, you know you can now shop on weekdays but you have also have work commitments on weekdays. Even though you're working from home you may actually spend less time shopping and you may be more conscious of, you know, I have this flexibility and I have you know, fun activities i I plan to do and things like that. So it is not clear which way it would go in principle as far as income and so on is concerned. So I suppose you might be worried that these work from home switches also have different kind of like time trends in terms of you know, income and so on. Actually like the results are all robust to including controls, you know, interacted with the time dummies as well. We actually know quite a bit about the household, you know, their occupation, education, age, some you know, information about their income in buckets and so on. And actually like it's, the results are very similar both kind of with and without those controls.

- So I was interested in sort of how much you think this could partly be. People aren't eating out quite as much. Like particularly if they're working from home, they're not eating lunch out and so now it's just like they're, they might have some sort of mental accounting where it's like I used to be spending X on food now I'm still spending way less than X on food 'cause I'm not eating out. I don't know if there's anything in your data that could get at it. Like if you've ever asked people that or whether you could just see whether it's concentrated in people who are just buying a lot more calories and so you know that they must probably have been eating out more and now they're just like eating at home more or something like that.

- Thanks, that's also a great point. I have a feeling we did ask a question about frequency of eating out but we haven't actually used it in terms of being very precise about this though we really can't see how much money they spend on the products that are not covered by Nielsen. There may be other ways we can try to get at this, but that's certainly like a possibility.

- Could you also look at the retailer data to see if they're changing price Strat pricing strategies but also seeing if they're introdu introducing new UPC codes? UPCs,

- That's a good question. It's not something that we have looked at. I think we sort of thought about it and you would have to use some kind of like, because with the retailer side, you know, it's not like we have a work from home, you know, measure for the retailer. So you would have to use something like, you know, zip three work from home shares, which does get a bit coarser.

- Hello, interesting presentation. I had a question about inflation. I guess I'm wondering whether the shift to work from home is inflationary because you obviously had inflation in 2021 onwards and in particular you had inflation in 20 in, in grocery, in grocery prices. So I'm wondering whether you are able to kind of gross up your estimates of price elasticity to kind of decompose the rise in grocery inflation in the in the US into a non work from home effect and a work from home effect.

- Thanks, that's also a good point. So you know, there's various senses in which we show households are like paying more, a little bit more 0.3% more for exactly the same thing, up to maybe 1.5 to 2% more for very similar things. So yes, like in principle this could affect, you know, retailer pricing and inflation.

- I also have the question about inflation, but since that was us, I also wonder why you use a share work from home per household. Because I think a more relevant matter would be whether there's a primary shopper in the household with remote work or not.

- We do, we did also ask questions about who the primary shopper is, but fundamentally, you know, all of the, the outcome, the shopping outcomes are really mostly at the household level. So this is why we defined work from home at the household level. Great. Okay. Can we.

- Yeah, thank you for having me here today. Thank you for organizing the conference. So I will be brief, I hope I will also be brilliant. I'm a fifth year PhD student at University of Bond and this is actually my job market paper for next year. So in 12 months. So your feedback is really appreciated for today. And so let's start by what we already know. So we all know that workers don't just care about wages, they also care about all sorts of non wage amenities. So work from home on the job training, all these different dimensions. And what we, we implicitly assuming is typically that if workers care about all different dimensions, wages, work from home, all these dimensions that firms also know what workers' preferences are and that they set some optimal wage and demand bundle. But to provide this optimal bundle, they also need to have unbiased beliefs about workers want. And this might be very challenging. So what we try to do in this paper is we want to test this assumption and we try to understand whether firms have correct beliefs about workers' preferences. So we have three things we're doing in the paper. I can only talk about the first one today. So the first one is we asking whether firms hold correct beliefs about workers, valuations of non wage amenities and how we're going to do this. We're going to run a large scale survey with firms and workers in Germany where we, we will compare workers' preferences and firms beliefs about workers' preferences. And second, yeah, I can only briefly talk about the status to trying to understand why firms might get it wrong and finally what consequences in equilibrium are now whether they are this actually matters. Alright, so the setting is we ran a large scale survey with firms and workers in Germany. So this is from all different types of occupations and industries and what we have, I think this is quite, quite nice. It's mostly actually owners and high level managers within the firm. Yeah, so the respondents on the firm side are actually people who make the decisions within their firm. It's often small to medium sized firms. So most firms in generally actually very small and owners are often like everything at once. So they do the wages, they do the amenities and everything else and we will also be able to link the responses to firm level admin data later on this year. Alright, so what we're going to do, we will have two surveys, one firm survey, one worker survey. In the firm survey we will measure firm's beliefs about workers' valuations. And in the work worker survey we have to benchmark their real valuations and the amenities we look at is work from home, which will be the most interesting for everyone here today. But we're also looking at particular scheduling on the job training, feedback meetings, and the four day work week, which means same number of hours but only on four days instead of five. Alright, so in design, how we measure beliefs and valuations, we'll use the same paradigm for both workers and firms and we will always have discreet choice experiments where workers and firms can choose between two different types of jobs. Yeah, it's the same task content. The only difference between the two jobs is that one job pays a bit more but doesn't have the amenity and the other job pays a less but has the amenity. Alright? And we can do this iteratively through to recover valuations and beliefs on the individual level for each amenity. Alright, so this is how the decision screen looks like. So we always have this job, A work from home is possible. Job B, work from is not possible job A pays a little less, job B pays a little more and then we ask to work as prefer job A or job B. And when you click on job B, you get a different salary differential if you click on job A. So depending on what you say, yeah, we change the salary differential to infer precise willingness to pay for these amenities obviously. Alright, so for the workers we use the exact same paradigm. So also discrete choice experiment, hypothetical choices between the two jobs. Again, they only define in these two dimensions, but now we don't ask about beliefs, but we ask, would you choose job or job B yourself? Alright. All right. Same decision screen again. Okay, so let's drive state straight into the the main result. So we're interested whether workers valuations and firm beliefs align. So this would be the 45 degree line in this figure. If firms overestimate how much workers care about amenities will be above the 45 degree line and if they underestimate it, you would be below the 45 degree line. So before I show the results, I would like to see your predictions first. Maybe someone with a microphone, maybe Emma, you have a microphone right now. What do you think from skater? They sometimes overestimate, sometimes underestimate.

- I'm sorry, I've seen the paper. So

- You've seen the paper. Okay, you can, okay. Perhaps next maybe Nick,

- I I I was on the committee so I know the, okay, everyone else

- I'm gonna shame. Maybe they, they underestimate.

- Okay. For all amenities or only for some

- For working from home? I'm thinking

- Working from home. Okay. Alright. Anyone else?

- Maybe there's gonna be a curve. Some level they're gonna be underestimating and then there'll be overestimating maybe.

- Okay, I think we have two predictions now going in opposite directions, which is always great. So what we find is for all amenities we look at firms underestimate how much workers value these amenities. And this is quite sizable for, for example, for predictive scheduling, valuations are 37% higher than beliefs and for work from home also quite, quite large the gap. So the main result is firms underestimate workers' valuations. So for all amenities, so for all amenities it's significant. And the immediate next question is whether these beliefs might matter for amenity adoption. What I'm going to show you next is the same figure, but now we will split the firm sample into firms who adopt the amenities and firms who don't adopt these amenities. And what we find is that the amenity adopters actually quite unbiased. So for work from home they get it completely right And amenity non-ad adopters are like the, the gray dots, they get it very wrong. Yeah. So it seems like these beliefs have something to do with what, what firms are actually doing. Alright, let me give you like a very short outlook because I don't have the time to talk about like the entire 126 pages draft we have done now. So what we find is what I already showed to you, like firms underestimate how much they care. Then second we are asking why they get it wrong. Yeah. So this might just be random chance or our methodology might be wrong. So what we hypothesize and also find is that managers themselves have very different preferences than workers. Yeah. So managers that get older, they're more likely to be male and might also just work harder in general and perhaps might care more about money. And when we ask them about their own valuations, we find that they value these amenities much less and that their own beliefs are very strongly linked to their preferences. And also managers who don't care about amenities also think that nobody else cares about this. So it seems like there's this strong difference in, in preferences between managers and workers that rise the result. And finally we are looking at like consequences in equilibrium. So this might not be relevant at all. So we set up a simple model to look at Yeah, how these misperceptions might have consequences and what the model predicts is quite simple. So firms should adopt an amenity when the amenity is cheaper or the costs are lower than work workers valuations. Yeah. So if work from home costs 50, $50 per month to to for a worker, but a worker would be able to willing to, to have a wage cut of 20% then would be optimal to offer work from home. Yeah. And what we see is first in the data that will be optimal for basically everyone who doesn't adopt when they can offer it. And that the, the model predicts that this leads to excess labor costs. Yeah. So you pay too, too high wages, too low amenities and then you're less competitive than your competitors in the labor market. And this is also what we find. So firms who underestimate are doing less well in the labor market. Alright, so why is this interesting? So why perhaps for this crowd? So what do we see in the past couple of years? First after COVID, the spike in work from home and then some firms started to have these return to office policies and I think our paper might be one explanation for why firms are yeah, stopping these work from home policies even though they seem not to decrease productivity and are quite cheap and workers really valued. One explanation might be that these older and more male managers who have different preferences than workers underestimate how much the, the work has valued. Alright, so quick outlook for what we're doing next. So right now we are in the field to have an info experiment with the managers to see whether shifting their beliefs might change adoption. We will then be able to link the firm responses to German admin data and we will test some more predictions in aggregate admin data. And of course the next thing we'll do is implement your suggestions. So thank you. I hope we'll have some nice discussion. Thank you.

- So why don't you direct questions actually going forward we'll have the speak direct you

- Just point to who you wanna Oh yeah, sure. Over there.

- Thanks. I had two quick questions on the methodology. So should, should we think of this as like marginal or average valuation? So in the light of sort of the work by an Italian Emma for instance, like the selection to these things, maybe the manager is like, no one wants to, wants to amend you when we offered it because everyone already had it or something like, like the marginal and average might really matter here.

- This is a great question. So

- And then is this pretax or post-tax when they think

- So we, we say that it's pretax and both workers and firms know that it's pretax. Okay. And what we also do is we randomize whether we ask the manage about the the mean worker, the median worker or the marginal worker and it doesn't make a difference in our setting. So they always underestimate it. And finally there are too many workers in the market who like and work from home for few firms to few firms who offer it. And it doesn't seem like we can just split the market into two groups.

- Yeah, yeah, yeah. Great. It seems that to to me that what you may want to know is the beliefs about the relevant worker pool. No, not necessarily about the average worker. And given that you find this effect of those who offered the amenity actually are unbiased, I was wondering whether that could be driven by the fact that they actually have the correct perceptions about the worker.

- Yeah, exactly. They might have learned over time. I think it's also possible that we might like underestimate it too much right now. So what we can do later with a, with a firm admin data as we can kind of like create your own workforce and like we know your workers are have demographics, X or Y and then we can kind of like create some yeah. Artificial firm for you. Yeah, please.

- It would be awesome to see, you might be underpowered to do this, but heterogeneity and the type of worker to see whether it's just, oh I, I am the residual claimant, I'm the owner of this company and so I have different preferences or it's actually about being old and male, right? If I have workers who are old and male or workers who are in managerial positions or you know, more experienced or something, do they have more similar preferences to the firm owner or firm

- Deciders? That's a very good point, thank you. Yeah.

- Related to that, what predicts bias and beliefs? Which managerial characteristics, which firm characteristics?

- So you mean like from size or manager characteristics? So

- More managers characteristics probably more interesting but size as well. If you just, if you, I I, I'm guessing you've looked at a predictive model.

- Yeah.

- Try to relate the bias and beliefs to observables of the firms and managers.

- So observables, it's mostly being older and male, but your own valuations are much more predictive. So like your own valuations explain like 33% of the, of the variance, but demographics don't do too much. Okay. And it's also like there's huge differences between industries so it people get it right and more like manufacturing people tend to get it wrong over there. Yeah.

- How good are workers at assessing the value of the amenities to make an accurate trade off to wage?

- Yeah, I think it's a great point. One assumption that we have to make when we write this paper is that firms, their workers kind of like know what they would choose if they had the the options, even if they don't know. So if they would just click randomly this couldn't explain our results, so we have some exercise then like if they don't know what they're doing, you would like assume that all valuations for these amenities are very similar but we find a huge variance. So yeah, it seems like they kind of know if they like work from home more than other things,

- More in terms of them making an actual trade off on the, on the

- Work decision based on whether

- Yeah, E centering.

- Yeah. Yeah, it's a great point. I mean we have to assume it in the end. I think one advantage we have is that we have the same paradigm for workers and firms. So it's at least like they have the same questions that they see and it seems like workers and firms seem to have different responses to it. Yeah. All right. Anyone else? Or perfect then. Thank you. I was very brief. I think maybe too brief, so thank you.

- Alright. Hello everybody. My name's Kyle Sherman. I'm a PhD student at Harvard Business School in the strategy unit. I am on the market, but this is not my job market paper. So that explains the tie. This is joint work with Raj Tory, one of the co-organizers, Miguel Espinoza, my advisor Tarana and Christus Ti, who's actually in the audience here and presenting two days from now. But I'm excited to share with you the results of a field experiment we did in 2020 when COVID, we didn't know what the future was gonna look like. And so for a while things looked a little more optimistic and we had the opportunity to run this experiment. But the research question I'm going to ask is how does the intensity of work from home, and when I say that, I mean how many days are you spending as an employer or a manager in the office versus at home or a third place during the work week? How does that affect intra organizational communication along hierarchical lines? So that would be managers communicating with their teams or managers, communicating with other managers, workers and workers, et cetera. And obviously I think in every presentation we have this kind of latent undercurrent of RTO policies becoming really prominent and that can be motivated by many reasons, but two potential mechanisms that managers might have in mind is that there's a coordination tax and that if workers and managers don't have that opportunity for face-to-face coordination, it might require them to be more verbose, put more effort into coordinating tasks because they lack that kind of tacit knowledge. There's also this idea of management through monitoring, right? So managers might find that it's more efficient to kind of manage by walking around and they might feel like they're more confident in assessing employee performance if they're able to actually see them in the office. And so we asked this question about compared to a situation which would be kind of the status quo for this particular organization of all employees in the office, does email and asynchronous communication serve as a compliment or substitute for tasks along the coordination and the monitoring domains? And so to give a little bit of background on this experiment, so we worked with Brack, one of the world's largest non-governmental organizations headquartered in DACA Bangladesh. They have over 35,000 employees globally and over a billion dollars in earned revenue. So it's a, it's a, it's a, it's a modern contemporary organization that has a lot of the same corporate functions you'd see anywhere in the world. But essentially just thinking back to early 2020, COVID was in code, it was starting to hit the world. And in Bangladesh, it was in March of that year that the government released a mandate saying that basically all these employees that had been working almost entirely from the office suddenly were working a hundred percent from home or third places, right? And so for a while we know that in the future things looked a little different, but summer 2020 things looked a little more optimistic. It looked like, you know, this might pass very quickly and the government allowed this organization to bring back a certain percentage of employees back to the office. So for kind of health and safety reasons, they would want to cap that at around 30 to 40%. It varied each week, but they were, they were allowed to bring back a certain density of people within the office. So that gave us the opportunity to put in this experiment where for a nine week period, of which about seven work weeks, 34, 35 days of actual work time was included in this experimental period. We randomized each week on an employee and day level whether an employee would be present in the office or at home. So this wasn't blocked, randomized in any way among the 130 employees and managers who took part of this experiment. It was a lottery conducted every day. And so what that looked like at the end of this period is that you had a certain proportion of workers that were randomized into kind of working fewer than nine work days from the office. In that 35 day period, you had workers that were randomized into an intermediate condition, which translates to about two to three days per work week in the office. And workers randomized into this low work from home condition. And crucially we observed all of the emails between these, between these employees and managers in that sample. So although 130 employees are part of the experiment, only 106 gave us consent to use their emails. And so we, we observe the message text, we observe all attachments and we observe subject lines of a very rich set of email communications within this organization. And we did restrict that set before even doing exploratory analysis to emails where both the sender and recipient consented. So this is a closed tic set where, where every employee in the sample consented to having their email data shared with us. So thinking about, to answer to our answer, our research question about what is the role of monitoring and coordination as a complement or substitute for in-person collaboration, we classified all of these emails into four different categories. And so you, you can there, there's also a kind of fifth positive unclassified category, but that's a very small set of our sample. And so importantly for this presentation are kind of the coordination emails. Those would be emails where you're kind of asking somebody to help out on a task or you're trying to assemble a team to put something together and monitoring emails. So that would be, for example, in this case a manager kind of chiding an employ for, you know, why haven't you done this? Kind of get back to me about that. We also had two different categories for just personal emails that would be kind of like greetings, how are you, et cetera. And also sending and receiving help and knowledge. And so I'll present because, because it's a brief and brilliant presentation, I hope I'll present just a few of our results. But I'd like to start first with just a basic understanding of the email volume and length. We found that compared to workers being co-located or not co-located, there's really not that much of an effect on overall message volume between employees except between work when you're looking at workers between different teams where you see kind of an increase and so that that could be explained, it's consistent with an idea of kind of workers from different teams not having kind of a shared tacit knowledge base. So that might explain why they need kind of more email volume. And that also explains kind of the intensity of that email communication. So when workers are, or when a sender and recipient generally are not co-located, they're either, you know, one's in the office, one's at home, or they're both working from home, the emails do tend to be more verbose and longer, and that's particularly the case when managers are emailing workers within their teams, quite surprisingly a very significant increase in the number of words per message. And that is present in, you know, among the types of dyads that we observe that in it's present, regardless of whether, you know, it's a mix of person at home, one in the office, et cetera. So I, I'll explain some of these coefficients. So like WW team would be a worker emailing another worker within the same team versus you look at that I guess six column, you would see a manager emailing a worker within their same team. And so, so that effect is not, it's not driven by a work modality of one or the other. It's kind of consistent, varying levels of intensity but same direction, same significance in most cases. But the more interesting results that I that speak to that very first question I was asking are about what does non co-location mean? And what we see is that increase it when employees and managers or employees, employees are not co-located, like they're, they're in different work modalities that increases the share of emails that are related to coordination. So that's quite consistent with the idea of, you know, when you're, when you're in the office, you can just kind of turn your chair around and do some, you know, like just follow up on something or put together some quick collaboration and that's obviously not possible when you're working from home. So we do observe this consistent effect of kind of an increase in the share of the emails related to coordination and that's particularly the case for worker worker diets in the same team. But surprisingly, and this is actually an intriguing result here because that that first effect is seen kind of regardless of whether you're talking about managers employees horizontal vertical hierarchy. But when we look at monitoring emails, my manager emails actually decrease the intensity of monitoring when those local, when the localities differ. So it's negative but not significant for many of those diodes. But in the same team workers engage in slightly less monitoring of each other and workers engage in less cross team monitoring. But I think what's actually surprising here is that you see I'll, I'll show the coefficients themselves. You see, if you look at that bottom table of monitoring emails, you see that when when workers for example are at home or they're not matching their actual manager's modality managers are engaging in less monitoring of those employees but the employees maybe because they perceive that their manager isn't able to monitor them appropriately, actually increase the volume of monitoring. So that's why you see between the fourth and fifth columns here on the bottom table, a difference in sign. So that suggests that we see, you know, less monitoring employees kind of thinking that they need to make the case to their managers suspecting that managers aren't able to do that, but managers actually feeling like they can monitor without that. So I would love to skip the limitations but briefly I'll just say obviously this is a very quick sample period. We know that there is a negative assorting that goes on if organizations adopt these sorts of policies. Of course we only observe email communication. We did run a follow-up survey with a similar population of workers, not with these workers but a similar population that showed that kind of email communication volumes rose in tandem with other forms of communication following COVID. So we're slightly less concerned about that but of course it is a concern that we're missing some of those other forms of communication. We also have some limited results that we're working on right now kind of trying to control for prior communication within a dyad. So not necessarily obviously, but if I've been working with somebody in the office for a long time, I've probably been able to establish a shared tacit knowledge base with them that might require less coordination if we're in different modalities. And in fact that is what we see when we start to control for that pre that pre period exposure. And of course work from home schedules in our experiment, you get some very clean identification when it's randomized on the day by day level. But that is not obviously how organizations tend to implement these policies. So that's something to watch out for. But takeaways briefly, co-location seems to be a substitute for horizontal coordination related communication, whether that's between, you know, workers in the hierarchy or managers, but it seems to be a compliment for vertical monitoring related communication in terms of the manager's monitoring their employees. Workers within the same team are allocating more time to communication or coordinated coordination related communication in terms of email length, et cetera. And I guess the managerial takeaway here would be face-to-face interaction is important for that monitoring, but it also helps employees build relationships. So this could be kind of one study that an employer or HR manager might take into account when they're designing a deliberate policy, but it's not the, it's not the final say. And with that I think there's four minutes for questions.

- I, I didn't understand how you got to the conclusion that face-to-face interaction is important for efficient managerial monitoring. I, I I I, maybe I missed it. How did that follow from the evidence?

- So that would follow from in this table we see if, if we look at those coefficient of managers to workers within the team, it's not a significant coefficient, right? Versus workers actually have a different perception within their team. They actually increase their monitoring I guess from the evidence this is what would be consistent with an understanding of it. But we're not quite getting at the mechanisms here. So this is, I I suppose like a bullet point that would be consistent with this explanation. But we, we haven't dug right into that.

- Thank you. This is super interesting. I wonder some companies, they have used the like AI tools to monitor their employees. Like say you had to check in every 10 minutes. I wonder whether you control for that kind of effects.

- Yeah, so we're not controlling for that. My understanding from doing some interviews with people in this firm is they didn't have those sorts of monitoring technologies in place. They didn't have kind of the Microsoft Office 365 monitoring technologies there. I think one other thing that we didn't really need to consider but we would have to consider now is, you know, email volumes. Email lengths are being driven up by the use of gen ai. This experiment was done in 2020 where access to those tools was not widespread. So we don't have to kind of consider that.

- I was wondering whether you can say anything about the real implications for the firm and also for workers themselves. One idea would be to get the HR data on performance evaluations to see whether certain types of communication basically leads to better performance or promotion and so on.

- Yeah, so actually not, not in this particular paper but in a previous manuscript we had, we did actually collect some data. One, not subjective data in the sense that managers were employ or assessing their employees on a scale, but it wasn't it, it's what the manager was assessing. So that that is a real impact in organizations. We also went back and collected kind of three years after the experiment, any promotion data and any promotion and salary increased data that we had among the employees. Unfortunately there's high turnover in this particular department. So by the the time we got to the third year there were, there were about 50% of the original employees really remaining. So we didn't find significant impacts of kind of this randomization into that. But that could be because of the limited length of the experiment. It could also be, it could also just be driven by the lack of power because of the turnover we saw. So yeah, we didn't, we didn't observe from the data we have any particular impacts on kind of measured increases in salary et cetera. But we did see workers in this kind of intermediate condition did have higher managerial ratings and they tended to report self-report better satisfaction with their work-life balance versus workers in the other two categories. Thank you very much.

- Hi, I'm Ein and I'm going to present you my paper. So dear applicant, do you mean working from home or shirking from home, which I wrote together with my promoters of my PhD, which I defended last month. So especially since the COVID-19 pandemic, a lot of applicants want jobs with the possibility to telework. But should applicants be explicit about the switch and the reason for it in the application process? Well the literature is unclear on its consequences on hiring chances and therefore we are gonna investigate, we investigated whether mentioning a desire for teleworking one or two days a week in the early, in an early phase of the recruitment process, whether it impacts the likelihood of being invited to a job interview. And first we are gonna look if it's motivated by productivity reasons and then second if it's motivated by work-life balance. And we also investigated what mentioning this desire for telework signaled to recruiters. Again first when it's motivated by productivity reasons and second when it's motivated by work-life balance. And these are the investigated signals. So we investigated whether there was an impact on anticipated achievements, driving anticipated commitment, anticipated availability and competence. So how did we investigate this? We set up a factorial survey experiment in which real recruiters evaluated fictitious applicants. They each evaluated five fictitious applicants and these applicants deferred on eight variables which varied randomly over two or three categories. And therefore because of this randomization causal interpretation was possible and we employed eight different jobs to increase the external validity. So the most important vignette factor was the preference for one or two days telework. So the last one it's had three levels or no preference was mentioned or a telework preference was mentioned and it was motivated by productivity or it was mentioned and it was motivated by a work-life balance. So we did this in flounders, which is a northern Dutch speaking part of Belgium and we did it in the, we collected our data in the spring of 2022 and we sent our email with the link to the factorial survey experiments to recruiters, which their email address was on the Belgian's largest website, which is the Public Employment Agency of Flanders. And it, we collected these email addresses from jobs that were similar to the ones that we used in the experiment 266 real real recruiters participated and they each evaluated five fictitious applicants. So in total we had 1,330 observations. Well first before I go to the results, I want to and say what we expected to find well for a telework preference, preference that was motivated by worklife balance, we expected like a negative consequence on hiring. Why? Because it's for most recruiters it de deviates from like the ideal worker. The ideal worker is the most desirable, desirable worker who is always committed and always available for work. But especially the productivity motivation. It's interesting because on the one hand it also deviates from the ideal worker image, but on the other hand you actively emphasize values related to the ideal worker. So if you actually say that you want to be productive, it is expected to please recruiters and so it has the potential to counterbalance the negative effects of saying that you would like to work from home. Okay, so the results first for the first research question. Well we, we did a multivariate to regression analysis with the interview scale as the outcome variable and with the job and recruiter characteristics as controls. And what we defined is that compared to a preference compared to applicants who did not mention a preference for telework, recruiters are 2.1 percentage points less inclined, inclined to invite applicants who say that they want to work from home for a productivity reason. And recruiters are 5.1 percentage points less inclined to invite applicants who say that they want to work from home because of their work BA work life balance. We did several robustness analysis including one robustness analysis in which we only used the recruiters that had real life experience with the job they recruited for and we find that they even punish more severely. So that's the first indication that it depends on recruiter characteristics because in our moderation analysis we also see that other recruiter characteristics play a role for example whether the recruiter is apparent or not. But also job characteristics play a role, the level of physical activity required in the job for example. And also applicant characteristics play a role for example the level of relevant working experience. So in some asking to work from home motivated by a work life balance reason is always punished by recruiters. And if you are motivated by productivity, well it really depends on the context, depends on recruiter characteristics, job characteristics and applicant characteristics. Okay then I go to the second research question, the signals and these are the signals more more in detail. So for anticipated achievements driving we have items like applicants with such a profile are usually eager to advance. For anticipated com commitments we have statements like willing to make sacrifices for the job or make work a top priority for availability. We have items like he will be, he or she will be usually able to work a substantial amount. And then for the competence factory we have three specific competencies. Social skills is the first one and then we have self-reliance and productive use of time as they are linked to work ethic. That's why we took these two signals. And what did we find? Well that expressing telework preference motivated by work life balance has clear negative effects on anticipated achievements, striving on commitment and availability but not on competence. And when it's motivated by productivity we do not see these clear negative effects on the signals, on the investigated signals. So then I come to the conclusion already. So a telework preference motivated by worklife balance is clearly punished. A work, a telework preference motivated by productivity depends on the context. The punishment for worklife balance is less is is bigger than the punishment. Punishment for productivity reasons and tailor preference for work life balance has a clear negative effect on anticipated achievements, driving commitment and availability. So what would we advise to applicants? We advise to carefully consider whether they want to express this wish for teleworking in the recruitment process for them there's a trade off between being honest to see whether there's a good fit with the organization and to potentially reducing their chances on an interview. And for companies we advise that they clearly indicate in advance whether they are open to telework and if they are open to telework, why? Because this will increase the likelihood of attracting pot good, good fitted candidates. And also if they proactively communicate this then the, the applicant isn't faced with this dilemma that I explained. Okay, so thank you. Yes,

- Super interesting. It ties up a lot. I'm sure Simon is thinking that yeah it ties up very much with your paper 'cause in equilibrium it's, I'm a manager, no one ever says that they wanna work from home. I'm a, you know, and I just think that they don't and I believe what they tell me, of course they're not saying it 'cause they know they're not gonna get the job. So completely consistent and it's actually hard for firms. I mean they learn by actions but they don't learn by listening.

- Yeah, yeah, true. Okay

- So do you know, have you had a chance to ask these recruiters how often they were actually seeing in their experience that this being request a request? Like do applicants indeed ask for working from home or is this something that maybe strategically they do not,

- You mean in real life whether this occurs? Yeah, well no we don't really know but we do see that telework, teleworking becomes more and more important for especially maybe for younger employees and and from real life context we know that a lot of people for example, accept jobs that are located far away only if on the condition that they can telework. So it's not implausible that they already like pronounce it in the first stages of recruitment process. And we also frame it like to the recruiters like you are the the, and you have to help in the second round of the review process, a former a colleague has had had a a former conversation by phone and so they don't not do not know whether the applicant themself like brought it up or the colleague like asked questions about it or so someone else,

- Can I make another? There's a I also, the other thing, if you think that recruiter's interpretation reflects like their valuation on someone that likes work from home, it's very consistent with the M and Natalia paper because there's another like treatment and select 'cause often wonder why firms don't offer it enough. And once there would be allowing a given person to work from home say may increase their productivity, but the types of people that like it are, you know, are low productivity types, which just like your paper that the selection effect is bad. And so that would be also very consistent with this one equilibrium even because you know when I've done RCTs that's a treatment effect and the treatment effect could still be positive if the selection effect is negative. And I think in you guys, one of the versions at least you found they offset each other. And so it's also, as you know, it's another kind of thing why it would explain why there's not enough in equilibrium is it generates this negative selection effect.

- One of the findings I found very interesting was that you don't seem to have any impact at all in expressing a preference for temporal flexibility.

- Yeah, yeah that's true. So that looks like it's more yeah common or more accepted like yeah indeed. That's an interesting finding.

- Is that more common in Flanders?

- Yeah, it is more common. Yeah, yeah indeed. But we would've expected some negative consequences but it doesn't seem to have an effect at all. So that's really like common Yeah like I dunno the words for it like accepted. Yeah, thank you.

- Yes, Thank you. Two questions. I think one of the regressions you, my understanding was you're explaining the 10 point survey response with an all S regression. And if so have you thought about using techniques that take into account the orality of the outcome variable? And second, so the whole survey is obviously non incentivized and one concern might be that, you know, there could be some sort of social desire be social desirability bias among the hiring people because they know, oh the public is on our asses for being more lenient on work from home and allowing that. So of course you see that they punish if somebody expresses wanting to work from home. But I wonder if you have thoughts about could the actual effect be even worse than that because they have sort of pretended to be a bit nicer in this experiment whereas if this was an actual applicant they would be even harsher.

- Yeah, well the first question I can be rather short. Well, because it's a, a scale from zero to 10 till 10 we, we saw it as an yeah continuous variable which is we, we didn't do an ordinal regression, which I can still, which is a good idea. And then for the second point, well we did include a social desirability scale and we saw that the recruiters that are sensitive to us answering social in a socially desirable way when we ex, how was it again? So the the they are, they punish more severely so is that correct what I'm saying? So if we, if we take them out and we only look at the people who are, are not as sensitive to social desirable answering, then we see that the, the results are more severe. Yeah,

- I think it's super interesting. I wonder whether you found any interaction effects by gender as we know the literature for like parenthood penalty, right? That's a motherhood penalty but there's a wage increase for fatherhood I wonder.

- Yeah, so we expected maybe like a flex, a flexibility stigma for people that are parents, for applicants that are parents. But it was the other way around actually there were recruiters that are parents that were more negatively, especially for the productivity reason because maybe they found it less credible. If someone says I want to be, I want to work from home because I'm more productive and if I as as a recruiter I am a parent, then I don't believe you that much. So that was the thing we saw. Alright, okay thanks. Thank you.

- Alright. Hello everyone, my name is Jacob Sano, I'm a second year PhD student at was City University and today I'm going to be talking about this paper wage penalties for place amenity evidence from remote work in the United States. The research question here is very straightforward, it's just how much income do workers sacrifice for the ability to work from home? And we'll also look at whether this amount differs across groups. So I'm not gonna go too much into the background literature here, but there are two main contributions to this paper. The first is an empirical estimation of the wage penalty for work from home with a sample that's generalizable to workers in the United States. Studies in this area so far have mostly either used self-reporting evaluations. So essentially asking people directly in a survey how much they would be willing to pay for the amenity or experiments that usually use theoretical job offers to back out that valuation. Examples of this include OL in 2021, massive play in 2017, OL in 2024 and also the paper assignment just presented three presentations ago. So thank you for that. But what we're gonna do here is have an empirical estimation that allows us to see what happens when these theoretical evaluations meet actual market forces to create a wage penalty. And then the next contribution is that for this kind of estimation, the estimation of a wage penalty for work from home panel data is usually required. But in this case we're gonna overcome that with the available cross-sectional data. The reason that panel data is usually required here is that in almost all cases the people who are making the most money are also the people who have the highest levels of workplace amenity. So when you do the cross-sectional analysis on that, it actually looks like there's a wage premium for these amenities, which doesn't really make sense in theory. And this is something that was touched on in masses at all in their paper in 2023. The closest existing work to this paper is DEFRA AL in 2022. They also do an empirical analysis on this topic, but they look at the UK instead of the United States and they also take a very different approach to the estimation. So for data, I don't think I have to explain this too much, I'm using the survey of working arrangements and attitudes as to the aa. So thank you to everyone who's worked hard to collect that data and also make it available for public use. Just one quick note on the data here is that the SWAA does have data available from early 2020, but for this paper we're only gonna be using data from January, 2022 onwards, which is the point at which the employment rate in the United States went back to its pre pandemic level. And this is to as much as possible get rid of the compounding effects of the COVID to 19 pandemic. So the data period for this paper is going to be January, 2022 to December, 2024 here. So to get to the actual estimation, the equation that we're gonna estimate is as straightforward as possible. So on the left hand side we have long income and on the right hand side we have our vector of controls, which includes pretty much what you'd expect. And then importantly we also have a measure of worker efficiency on our work from home and including this is gonna make sure that what we're estimating is not just whether people get more or less productive under work from home. And then our variable of interest is going to be this work from home level variable here. And we're actually gonna define this as a categorical variable. So if someone is working 0% of their days on a work from home, they're gonna be the non category. Zero to 50% of work days done on a work from home is low, 50 to 97% is high and then 97 to a hundred percent is full. Using this as a categorical variable I think helps to make the results a little bit more interpretable and it also helps to deal with any potential nonlinearities in the effects as well. But so if we were to just estimate this equation as is, we still have the problem that we saw earlier where people who are working from home the most are also the people who tend to be making the most money. So here we have a chart that was created again directly outta the SWA data, but on the horizontal axis here we have work from home frequency and then on the vertical axis we have income and we can see a very clear positive correlation between these two. So if we were to estimate this equation, our estimates for beta would be biased upwards. So we need to take care of that. And to do so, we're going to add in this income 2019 variable. And I believe this was probably available from the screen questions from the SWAA, but this is going to allow us to control for the workers' ability or ability to gar our high wage. And critically since 2019 was before the mainstream legitimization of work from home as an amenity, it's going to allow us to do that for most people as a fixed value before they had any sort of decision on how long they were gonna work from home. So this is gonna give us a very clean way to sort of pin down their ability prior to work from home setting in. So then we'll go ahead and estimate both equations on the left column here we have the equation without the control for income 2019. And we can see here very clearly, or I guess you can see here looking at the full and high categories here for work from home. On the left hand side there's not really anything going on, but when we add in on the right column the control for income 2019, the coefficients for the full and high levels of work from home becomes significantly negative. And the base level here is going to be the low level of work from home. So what this is telling us is that for workers to go from the low level of work from home to the high level of work from home is associated with approximately a 1.8% reduction in income. And then going from the low category to the full level of work from home is associated with approximately a 2.3% decrease in income. And these are relatively modest decreases and they're certainly less than what previous studies have estimated for sort of the theoretical valuation of work from home. So this is potentially telling us that people are currently not paying their full valuation for the amenity. We can also look at the bottom here, the non category and we see that we have this very counterintuitive negative coefficient for the non category here. What's actually going on is that in this non group there are people who could potentially work from home but for some reason or the other they're not working from home. And there's also people that just because of their job description could not possibly work from home. And people in this later group generally have very low relative wage levels. So that's where we're seeing the negative coefficient there. Unfortunately in the data it's very difficult to separate out, separate out these two groups for, so for our purposes we're only going to be comparing the low level with the high in full levels here. So we can also do some heterogeneity analysis here. One possible dimension of interest can be industry. So in different industries there's very different cultures, likely very different ways of looking at work from home. So it's, it's possible that the wage difference or the wage penalties can be different across industries. So to check this, we can run the analysis for each industry separately and then we can order the industries by share of jobs in each industry that can be done remotely. And for this we're going to use the index from Ding and Neiman in 2020. So we do that estimation and we get the following results. So on the vertical here we have the industries ordered ascending by the percentage of jobs that can be done remotely. And then on the horizontal you have the estimated wage penalty. So the further left you get the steeper that wage penalty is gonna be. And then as you move toward that zero line there, the penalty starts to disappear. And what we can see is that when we're at the top and we're in industries where there aren't a lot of jobs that can be done remotely, the wage penalty is very severe. But once you get to the bottom an industry such as information finance and insurance and education where there are a lot of positions that can be done remotely, that wage penalty is essentially non-existent. All right. Also, one of the really attractive things about work from home is I can help groups of people who historically have been for whom it's been very historically hard to get into the office itself to increase the labor supply and also to become more integrated into a workplace. One of these groups is working mothers, so we check the wage penalty for them in this interaction here. So we add in interaction for being female and having children. And what we see particularly in this high category here is that when we add in the female and has children interaction there, we get a significantly negative coefficient. And this is telling us that working mothers are currently facing a steeper penalty for the amenity to work from home than our other groups, which is an unfortunate result here. Alright. And then to recap, we find that workers are paying a significant but moderate penalty for the men to work from home and that the penalty is heterogeneous across groups and that it's smaller in industries where more jobs can be done in work from home and larger for women who are raising children. One point at the bottom here that I wasn't able to get to in this presentation is that there is a, a slight issue with the measurement timing of some of the key variables here and that they're measured in slightly different periods. So we make sure that our estimates are numerically valid by pulling in data from the survey of income and program participation to impute some of these key values, IMP inputting those values and then running the analysis again, we find that the estimates are in fact robust to using imputed values for the key variables. So if anyone's interested in the details on that or how we do the imputation, please feel free to ask at any points and I will wrap up there. Thank you for your time.

- Yes. Yeah, interesting paper. I just, I would encourage you to approach the data with a slightly broader conceptual framework.

- Okay. - Because in know basic theory says there's, if there's an amenity value to work from home, work will get some of it, but it also increases the geographic extent of the labor market.

- Okay. - For those workers who can work from home, some are all of the time, which would lead you to expect that they'll be able to get better matching efficiency.

- Sure. - There's a productivity effect that comes from that that won't be, as I understand it, captured by your income 2019 variable.

- That's correct.

- Yeah.

- And also work from home affects the extent and nature of labor market competition, which could alter how match surplus gets shared

- Right

- Between worker and firm. That could go in either direction. So it's useful to get these net effects that you have here, but I don't think you can immediately jump from your net result numbers and include and conclude that's the consequence of how the amenity value of work from home works out in equilibrium because there are these other forces also.

- Right,

- Right. Changing from the pre pandemic to the post pandemic situation.

- Understood. Thank you very much. Yep.

- Excuse me. Very interesting. So you mentioned the maus at all paper.

- Yes. Finding

- The positive relationship between wages and amenities. Did they also try controlling for income or is the key innovation here being able to control for a lag of income or a lag of income prior to the proliferation of remote work with COVID?

- That's a good question. I think, yeah, the, the masses and all strategy I think was have to kind of go back and look at the details on that, but their strategy I think was, was quite different from what we're doing here. Yeah. The income 2019 is doing is it's, it's kind of one piece of longitudinal information within a repeated cross-section data set. So yeah, it's pinning down essentially, you know, where someone was at right before the explosion of work from home. So yeah, the, the idea here is that we do have this sort of like one pinpoint piece of information that's going to allow us to do this estimation within an otherwise entirely cross-sectional.

- Really quick, but just on your last side, did I see right that you do show, as was mentioned earlier, like a fatherhood premium 'cause you have full with has children being positive and high in has children being positive, so, so is that pointing to the fatherhood premium from work from home compared to mothers?

- Very, yeah, really good question. That certainly could be it, but this, this, especially on the full level here, when you had that positive coefficient coming out, this is very strange. I don't think there's really a reason we should see that, that has children being positive without a lot more, you know, detailed analysis. I can't really see anything firm on it, but it's possible that there's a little bit of like reverse causality going on here as well. So potentially people who know that they can work from home fully without facing too much of a penalty in their job, it could be easier for them to actually have children. So yeah, without more in depth analysis, I can't really say anything firm about that, but that could be something that's going on there.

- Yeah, I mean the study of compensated differentials is such an interesting one. There's a, yeah, I'm trying to think about how to say this in 10 seconds. Yeah, so there's all these workplace amenities. I guess the main thing is just that controlling for last year's income is just, it's a, it's sort of a crude proxy and so thinking about what other things you may be have in the SWA that you can maybe saturate the model a little bit more. Sure. We found, and we have a, I have a paper coming out in management science where we measure workplace practices very in a very multidimensional way. We show that controlling for these workplace practices sort of functions like that sort of unobserved heterogeneity, but Anyways, I think I'm been past the one minute, so we'll chat more when

- Down. All right. Thank you very much. Thanks very much.

- Thanks for having me. This is a super interesting conference. I have done a little bit of work on remote work, but nowhere near the extent of the breadth and depth that we're seeing here. The title of this paper is that remote work makes organizations top heavy and you will know everything that you need about the paper from the title because the title is the result. I also promised that I was gonna have a draft to circulate and there's this law that everything takes longer than you anticipate, even if you know the law or the rule. And that has certainly come up here. And so I am looking forward to your feedback genuinely, because I actually haven't written the paper yet. Here's the context. We know from the work of a lot of people in this room that remote work skyrocketed during the pandemic and remains at roughly 28% of days at home. Today. I don't actually think the 2025 citation is the right citation because this is from a paper that I think you're quoting from that was published in 2024, but you've continued to update the data, which is a great service to all of us. But remote work was essentially below 10% of all workdays pre pandemic, I think, across data sources and is now a substantial phenomenon. And there is mixed evidence about the effects of remote work from many people in this room. If you look at call centers or the US PTO individual productivity goes up. If you look at my work with Barick, which I think I'm the seventh author on this, small businesses that were induced to switch to remote work during COVID initially had a dip in productivity, but then tech investments and process improvements meant that most small business owners reported positive changes sometime later. And then on the negative side, if you look at knowledge workers at Microsoft, their collaboration patterns become more siloed. And then at unnamed tech firms by Gibbs at all and Emmanuel et all, you see that collaboration and helping patterns fall. More importantly, managers of large firms report continued difficulties with work from home and the Wall Street Journal has a series of articles on workplace monitoring technology that has been implemented in response. My favorite is a picture of a guy in a cutoff shirt with no sleeves who has a mouse that he's tied to a fan and the fan oscillates. And so it moves the mouse to trick the monitoring software. And so it, it looks like there has been a very different response at the firm level than at the individual level, both based on how managers perceive the implementation of remote work. This paper is simply going to show you that firms responded to the remote work transition by adding managers. And then the economic question is what is that in terms of an overhead cost? And thinking about whether the managers are kind of right because might slow down response times. It might add bureaucracy, but it also is going to add sort of the direct cost of management. So there are three reasons why I think that this is important to look at in terms of looking at the share of managers at the firm level or at the economy level. The first is that the span of managerial control is sort of a key parameter in most models of hierarchy of earnings, inequality of the firm size distribution. Many of you probably reflect on Lucas with some mixture of admiration and pain in thinking about what this will look like in terms of aggregate efficiency in the economy. But is, and much of my work is consistent with this second sub-bullet. There is little evidence on the micro foundations of what determines the managerial span of control. I have done a lot of my gray hair is due to trying to benchmark managerial productivity, but as far as I know, there is nothing that kind of looks at the micro foundations of how spans of control vary across different people or different places other than sort of idiosyncratic variation. You can correct me if I'm wrong. The second is that managerial quality shapes human capital development and workers careers. And there's emerging evidence that suggests that workers who are young who are exposed to remote work may not have the managerial impact that they need or the managerial or mentoring support that they need to foster long-term careers. But you can debate that I think Steve's point earlier about, you know, matching across geographic places and forming better matches is something that's going to push back there to where maybe the initial match is better for workers, especially if they were coming from labor markets that were not sort of frontier labor markets or high innovation cities and forecast. Finally, forecasts of the future of remote work are mixed and the organizational changes that have occurred in response to remote work I think are helpful for contextualizing the unit economics around how this is going to look from the perspective of the firm's income statement compared to the benefits that accrue to workers. Okay, there are three things that I do in this paper. The first is I build with the assistance of chat GPT, an assignment model that captures spans of control, wages and managerial skill that vary with what I term a coordination cost of remote work. And this is going to mean that this is a general equilibrium type of model where you will see marginal managers either pushed in marginal workers, pushed into management or away from management depending on some parameters of that coordination cost, cost function. And then I use two data sources. The first is using industry year variation from the OES, which is the occupation and employment statistics. That is a survey produced by the Bureau of Labor Statistics, and I merged this with the ding and nyman occupation level telework ability data. That was simply a forecast of what occupations and occupational tasks from O net could be done remotely. And then I look at managerial shares after 2020 as a function of suitability for telework. I'm still at 30 minutes of time, so I'll continue to keep going, but if you want to gimme the right time, I'm happy to also stay within the constraints that you've imposed. And then the second source of data that I use is light cast data or burning glass data on job ads for managerial positions. And this allows me to look at firm level differences in hiring or sorry, in advertising for managers compared to production workers. Okay, here are the main findings and then you can fall asleep. The raw data shows a dramatic increase in both managerial jobs and job ads. Pre COVID 5.3% of the labor force in the OES data is in managerial jobs. That gives you a ratio of about 18 workers per manager post COVID. That number is 6.7% or about 14 workers per manager. Now obviously I'm not gonna explain all of the increase with remote work. There are a lot of other things that are going on. And so you have to ask what the factor loadings are gonna be in terms of remote work that explains this variation, and I'll get to that in a second. But the second data source, which is like cas, shows that there's a one percentage point increase in managerial job ads moving from 11.3% to 12.3% between 2019 and 2022. Now, using the two sources of variation that I have, remote work accounts for about a 0.4 percentage point increase or a 7.5% increase from the baseline or a 28.5% increase or 28 per point 28 ish percent of the total increase in the managerial share over time using a cross industry variation and telework ability. If you take that estimate seriously and you multiply by some measure of managerial wages, you get an increase in managerial costs per worker of about 550 to about $1,500 depending on whether you allow wage adjustment as an endogenous part of the model using the light cast data. Remote work accounts for an increase in job ads of about 0.2 percentage points or about 2% over time based on that one percentage point increase in managerial job ads and light cast. I think the second number for those of you who have used the remote work data out of light cast is probably subject to a lot of measurement error that I have not corrected because light cast thinks that about 6% of jobs are remote jobs on average, and I haven't done the corrections that some of you in the room have worked with. But as a result, the light cast number is probably a lower bound on what the the true number is. Okay. There are a couple of auxiliary things that you should also be aware of in in the model, in order to get this increase in managerial shares, managers relative wages have to increase in order to induce more people to become managers compared to production workers. We sort of know that wages increased across the board and most of the economy as a result of the post COVID economy. But managers wages actually increased more in the OES data and managers in relative wages increased most in tele workable sectors after COVID. Now this project is a little bit weird because the way that I started it was that I wanted to understand whether there were different demands on managers in the remote era or in remote work. And we essentially detected zero despite a ton of effort changes in advertised skill demands for managers in remote positions. I still think we could be wrong about this, but as of now, what I was hoping to find was some increase in the demand for communication skills, social skills, or as my colleague David Dimming would tell you, some secular increase in social skills over time I thought was going to be accelerated with remote work. We don't find that. And then the identification here is going to be subject to the police and the police are not gonna be happy with anything that we do or anything that I do. But it's hard to tell a story about some other factor that is driving these results. If you look at things that would be confounded with the COVID shock, like episodes of new technology adoption, if you compare the COVID shock and presume new technology adoption to past episodes of new tech adoption, the AI or ML boom or cloud computing and job ads wouldn't be large enough to drive these effects. Demand shocks or from from growth don't appear to drive this. The composition of hires actually kind of boosts this because workers and tele workable sectors actually had larger spans of managerial control. And so if you're adding tele workable workers to accommodate remote work, this would go the other way. It also turns out that Elon Musk is not responsible for this. If you exclude X, Microsoft, Netflix, whatever, or other tech firms in general, you don't overturn this result. Okay, let me do five things very, very quickly. First is I'm probably gonna spend too much time on theory, but, and then I'll get into the industry and firm level estimates and then I'll talk a little bit about manager wages, skill demands, and some alternative explanations. And the eight minutes that I have left, okay, here's the theory and it's predicated on a key benefit of being in office, being face-to-face communication where coordination costs of collaborating or of reading body language or things like that are going to be lower. And in the model, this is gonna show up as follows, we have a single good economy with a unit measure of agents and the good price is normalized to one. There's gonna be a separate sort of numer error, which we'll call y. That is just the output level of a worker who is working in autarky with no coordination costs, but you're gonna need coordination to actually produce anything. And then agents are going to each have a coordination skill, which I'll label S like a social skill drawn from a distribution, capital G of S. And then they face an occupation choice where production workers are just going to earn a flat wage W, and then managers are going to coordinate production. And to make things simple on me, they're the residual claimants on their effort where they get why the production worker minus the coordination costs, minus the wages that they pay the the production worker and coordination costs are simply just fee, which take a value of N and R, where N is the number of workers that the manager oversees. R is whether the manager is leading a remote or an in-person team. And then S is that manager social skill or their coordination skill, which just drops the level of fee for any given level of S. The two important assumptions here are that coordination costs are going to be increasing and in and increasing in remote work. And then a type S manager is going to solve a problem where they just choose the number of workers in to satisfy the maximization where they take Y minus their coordination costs, which is again increasing in the number of workers in that they're hiring less the wages that they pay per each for each worker. And so a higher skill manager s is going to be able to support a larger team. So what does the equilibrium of this model look like? It's three things. You need a team size function in star of S, which is going to point wise vary with S. So higher skill managers are gonna hire more people. You need a skill cutoff for the marginal manager, which I'll call S star and a production worker wage W star, so that the marginal manager is going to be indifferent between becoming a manager and a worker that you get appropriate sorting or less skilled agents are going to become production workers and those who are more skilled become managers and you just need markets to clear. And then market clearing is just going to imply that I can characterize the aggregate team size or the average team size in over bar as the number of agents who become workers over the number of agents, one minus G, who become managers at this at this point a star. So this means that there are is a lot that I can do here to explore what that fee is going to mean about team sizes and who becomes a manager. There's one empirical restriction that I'll show you in just a second, which is going to help me pin down what the form of that fee function is. Before I get there, lemme show you a couple of comparative statics that require some regularity conditions that I'm going to skip over. The first is that if you move from in-office to remote work, you can see that under remote work the identity of the marginal manager falls in terms of skill S, which is on the x axis. But team sizes, which is on the Y axis also fall. That proposition that I just showed you said that if I draw more people into managerial work, average team sizes are going to have to fall because we have a unit measure of agents. But this second thing here that is noticeable is that team sizes fall more under this parameterization for the highest skilled managers compared to lower skilled managers. And this is going to be the empirical bite that I'm going to have in a little while. Second, if more agents are going to become managers, we actually need relative wages to move in a way that supports this. So here you can see that there's a cutoff where there is a kink in these curves that determine where production workers W earnings are going to be compared to managerial earnings, which are the upward slowing parts of each of these kinked peace wise functions. But under remote work, both earnings fall, but earnings fall faster for production workers. Now this is where the numar rare thing comes in because we sort of think that COVID brought inflation and so you need to interpret wages as sort of relative wages rather than wages in a, in real terms or in absolute levels. And then finally, I can show you a comparative static that larger teams are gonna shrink more when the marginal harm of remote work is going to outweigh the marginal harm from congestion of having more team members. And formally the chat GPT proof is that bigger teams shrink more if the left hand side of the inequality is bigger than the right hand side of the inequality. And so this is going to at least gimme some empirical bite about the the harm of remote work in terms of coordination costs based on the third derivative compared to the harm of adding another production worker. Okay, that's it. Let me show you a little bit of data. I use OES data in annual surveys from 2017 to 2023 covering 254 industries by 938 occupations. These are industry by occupation cells. The thing that you have to be careful about using the OES data is that these are three year rolling surveys. And so the 2020 and 2021 samples are gonna reflect some pre pandemic averaging. And then I merge in the dingle and nyman measures of telework ability at the three or four digit occupation code level and collapse the industry. To give you a measure of industry level telework ability, the key dependent variable is the share of managers in an industry which I take as the major occupation group 11 codes. This does not include frontline supervisors, which are a separate occupation code. And it turns out that most of the action is happening at higher level management levels rather than frontline supervisors, which is not exactly what I expected going in. But also might tell you something about the coordination cost function, that if you're a frontline supervisor, it might be relatively easy to monitor what people are doing if they're in telework, but there are losses and coordinating between other people. The top four occupations by headcount here are general and operations managers, financial managers, et cetera. And then gambling managers are some of the bottom four headcounts here. But there are a whole host of detailed occupations based on what you can see. And then light cast data, I think most people are familiar with. We have 2.2 million unique firms out of the light cast data that does named entity recognition. But we condition on having some data pre and post pandemic for most of the, for most of the sample. But I'll also show you how new firms compare for firms that are launched in 2021. Okay, here are just some summary statistics. I already kind of told you this, I'm running out of time. What what is so also notable is if you look at the dingling nyman telework ability measure on the x axis and the share of remote job ads at the industry level on the Y axis, these two things are highly, positively correlated. But the light cast data here does not adequately picture the state of remote work and the economy on the Y axis where you can kind of see that the median is around 6%. We obviously know in other work that remote work is happening in a much greater clip and so there's gonna be a lot of measurement error. And on the Y side, one could sort of think about instrumenting with this telework ability measure as a reduced form, which is the approach that we take. But here are the estimates in the OES data and the light cast data and the light cast data with a variety of controls, you end up getting point estimates that look sizable and positive accounting for between 0.4 percentage points to 0.2 percentage points of the increase in the managerial share between the OES and light cast. You can see this visually if you look at the OES telework ability measure the light cast, measure the light cast firms using the telework ability measure at the industry level or the light cast remote job ads measure where data are compressed. And then if you ask yourself what firms are actually responding, you can sort of see that it's the firms that had the smallest managerial share on the right hand side of this figure in the post COVID period in the light cast data that are increasing their share of managerial job ads. More so most of the masses coming from very efficient firms with a high span of control who are, who are adding managers relative to firms that were already kind of top heavy in bureaucratic before. This is the the nimble firms that then become relatively top heavy. And so most of this is happening at sort of the most efficient firms or the firms that tended to look like they had many more workers per manager. Okay. The next thing that I'm just gonna show you is a triple differences regression out of the OES data showing you that managers wages and tele workable industries increased most in the post period, which is consistent with the model. And then I'll kind of skip ahead to say that we can decompose these costs of managerial overhead into a direct cost of adding managers per worker of about $554 per worker. No, this is a pretty big increase because the support cost of managers per worker is about $6,500 per worker in the pre period. And then there's an increase in manager relative wages of about $965. What we haven't done is attempted to quantify output changes or whether firms move to a new ISIC quant, which is gonna depend on the functional form of fee or any benefits from office expense savings, commute costs, et cetera. I'm way over, I'm gonna skip managerial skills, new firms, et cetera, et cetera, to just kind of take,

- We, we have a whole paper using like cast data to, and it sounds like you know this paper,

- This is with Steve Hansen and Han Hanson,

- But our, our work from home rate, which using a supervised machine learning model heavily validated, gets about twice as much work from home as you're getting from the same underlying data. So you might wanna think about using our algorithm. I think you, it would, you can just feed your data through our, I think we've got it on site right Nick, you can just, you can just feed it through our model and, and then use it inside your analysis. I presume you're doing this, so yeah, something you might want to think about. We get, I'll stop there 'cause see if other people want to say things.

- Oh, I'm supposed to call on you. Yeah.

- Oh no, super interesting. So just to make sure I understand it and the question, the idea is it's harder to manage people when they're remote 'cause of these coordination costs and therefore it takes more time to manage any given person, therefore for a given size of the worker force, you need more managers.

- Yeah, I, I guess this kind of depends on what you think the managerial role is, whether it's sort of direct supervision solving problems. In the garo style version of this paper, you would maybe expect it's frontline supervisors who have a higher helping cost to H and that means that their teams shrink. I actually think that the thing that is happening here is that there's more to coordinate. And so as a result you add managers where the direct effect on frontline supervisors actually isn't present in the data. In fact, that's relatively flat, but higher level managers are being added relative to managers who sort of oversee production work directly.

- I mean one story that would work, I, from talking just to tons and tons of firms, the story that comes up over and over again, someone mentioned this earlier, you, when someone's in person, you can see what they're doing. So like if Natalia is managing me, yeah, am I at my desk typing away, it's five out of 10 management, you're basically managing inputs. But she doesn't need to have a good performance evaluation system. Whereas if I'm remote, she needs to evaluate what I'm doing. And for companies that's quite time intensive. If you've gotta do a 360 review, like in McKinsey, the saying was it would typically partners were spending 20% of their time just on doing evaluations. So 'cause she can't see me, she now needs to have a proper evaluation system sit down with me once a call to say, here's your performance, here's what you've done. Well, here's what you've done badly, et cetera. It's actually better management, but it takes more time. So that'd be completely consistent with everything I've heard.

- I think that's true. And I, I mean I think one of the limitations here is that we're not gonna take a stand on how this changes production at the firm level, but there could imagine an unconstrained world firms chose not to do this. And then remote work happens through the, the COVID induced shift

- There. There's another subtle issue here, which is the cost per unit of labor efficiency is spatially differentiated in the economy. So by moving to a model that allows for more remote work may allow more managers, it may more managers as your model emphasizes at the same time the cost per labor efficiency unit supplied for the production workers goes down because you're tapping into these other markets.

- Yep.

- And I don't know whether you want to try to quantify that and feed it into your overall assessment. So another way to say it, it the, the per worker metric is not obviously the right metric here. It's something harder to observe, which is the per unit of labor services. And at least qualitatively that effect will, will cut the opposite direction, I believe from what you are. That's true. Focusing on

- You, you can write down a version of this model and you can show it in a special case that if the wages fall enough team sizes rise because there's this opposite pressure where high type managers would then say, I'm getting a deal on the labor input. And so all of this depends on that functional form of the coordination function. But you're exactly right.

- Can I ask any question? So just two questions. One was about the importance of this hazard, right? The geo overwhelmingness gaw uniform versus our distribution. Small, small question. And the other one is, what about the role of complementarities between workers and managers and workers? You have cast a problem in terms of coordination, but there's that that, that is another aspect. I was curious about your thoughts.

- Yeah, so my answer to Steve earlier was in fact incomplete. That this also kind of depends on the functional form of the distribution of talent. And in this case, if you become a production worker, all production workers are equal. The talent only matters in the managerial sector. But you could easily sort of imagine that a model that has two different types of talent where you would get sorting between management and production workers is going to then feature some endogenous match that is going to determine output. I just kind of shut it down because this turned out to be a lot more complicated than I thought it was gonna be going on.

- Yeah. But as you know, moving in that direction makes this model, this kind of assignment model much more complex.

- Yes,

- It can, it can easily become very hard to figure out the equilibrium at all. 'cause you don't have this kind of single crossing threshold.

- Right? And this model is surprisingly complicated as it is and it's a model with very few moving pieces.

- So in, it seems to me that a lot of comp listed companies, if you look at for example, their earnings calls over the past few months there's been a lot of talk about cutting middle managers in particular and big tech. And so I wondered, I guess this pertains somewhat to Nick's question C. Can you distinguish between, you know, managers who manage managers and then kind of proper managers, middle managers? Can you, can you get any granularity there? Like who's, who's actually increasing in numbers the most?

- It is possible. I, I've thought about doing this in the light cast data. The thing that gives me concern or pause with the light cast data is you don't actually see head count. You see job ads which reflect things like the beverage curve or labor market tightness. And so to the extent that these layoffs have happened, they're going to distort the interpretation of the light cast data because the labor market tightness has changed for a particular preoccupation. That's why we sort of made the choice to exclude the, the, the Twitters and exes and big tech of the world because we did know that there was the perception that firms had become bloated at the middle, middle manager level and they cut staff. So what I can tell you is it's not that that's driving this, the results are almost identical if you exclude big tech. But I think if we had a good measurement strategy to get around some of these concerns in the light cast data, that's exactly where we would want to go. But we've been a little bit hesitant to do it because it's imperfect. Oh yeah.

- Do you see anything in technology affecting the coordination cost? 'cause that's a big thing between whether you're together or not. Huge investment since nine 11, or excuse me, since COVID has that changed that waiting?

- I I mean other people in the room can kind of tell you this based on their own work. But in the survey work that we did with small businesses, small business owners pointed to their investments in technology is the thing that caused their perceptions of remote work to change. My suspicion based on this work is that our focus on small businesses and my prior survey work is kind of missing most of the action because small businesses are gonna be less affected by this span of control thing because they're less complex and they're run by managers who are on average less talented. Because if they were more talented, they would've grown. And so I I, yeah, okay, that wasn't supposed to be a joke Natalia, but so, you know, most of the action here is happening with larger firms where that coordination cost is gonna be more pronounced. But I, I do think that there are obvious tech investments that have allowed some adaptation to remote work, especially among smaller firms.

- This project is about the idea that perhaps the hot two technology trends of the last five years remote work in generat AI might have something to do with one another. And it's motivated by this idea that there is lots of generative AI adoption around, it's all over the news we're all hearing about all the time, but it really varies a lot across different places. So some firms struggle with it. There's lots of news about sort of proof of concept projects not really going anywhere. And so it's raised the question, what makes it easier to adopt generative ai? Why are some firms lagging and some firms moving ahead? And I build in this idea from the management literature that organizational adaptation plays an important role in terms of how well firms can use certain technologies. A key channel is that you need to make complimentary investments in it and change your firm organization to really take advantage of the new technology that comes along. And that then raises the idea that perhaps it has past dependence, right? So if you adapt your firm to one technology that comes along, you invest in related skills, perhaps that makes it easier to then also a adopt the next technology that comes along. And I'm gonna use remote work in this context as a natural experiment that sort of shock some firms into having to build out a digital infrastructure backend. And I'm gonna build on this idea that, that people of other race in other contexts, I used to actually have a quote by Nick on the slide, I took it off recently about, about this idea that working remotely interacts with AI because it either might make you more or less replaceable, but also it it forces firms to adopt certain technologies. And so to the degree that remote work induced some exogenous variation in terms of what firms had to invest more in technology, we can use this as a shock to then see what that does to those firms attitude towards gene fi. And so I'm gonna look at whether or not there is something that I would call an organizational technology ladder where adopting one technology then causes a firm to find it easier to adopt the next technology. And so I'm gonna show you first evidence that remote work causes greater generative AI adoption. And I'll flip this around and see what happens to remote work after gen AI is adopted. I'll give you a little bit of intuition for why this might be happening and I'll focus in on two mechanisms that I think can explain this. The first one is around organizational adaptation. Firms change their work processes and the skills they hire for when they adopt a new technology, and that then makes it easier for them to invest in the next technology. And also there's a more general mechanism related to management quality that some firms are just better at dealing with new technologies, including remote work. Okay. The, the data that I use is the, the same light cast data that Chris is talking about. So this is, I measure a skill demand for gene AI in job postings. So I see whether or not a job posting mentions the words gene AI or a large language model. And so I'm just gonna define generat AI adoption, the degree of generat AI adoption as the average share of job postings by that firm that mention generat AI remote work is defined as the, the share of jobs that are remote. I'll include hybrid here. It doesn't really make a difference, but for now I'll, I'll, I'll lump remote and hybrid together and then I'll use a couple of job characteristics for later analysis based on sort of existing papers around what particular job characteristics might matter for firms such as social skills, inflexibility of jobs, or decision making intensity. Just to give you an example of what I'm talking about here about measuring gene AI and job postings, this is an example of a job posting. As you can see here in the, in the last line, the, the tasks for this job, this is a product specialist, associate this sort of an entry level position at a real estate fund. It says that this, this worker has to utilize generat ai, specifically J-L-L-G-P-T to support and optimize specific tasks and initiatives within the organization. So they've clearly build some sort of internal chat bot and they want this worker to work to work with it. And so we can tell from this job posting that this company clearly uses generative AI internally in particular wants this particular role to have some interaction with generative ai. And in the empirics in the paper, I'm mostly gonna try to basically convince you that there is a causal mechanism underlying the graph that you see here, which just plots the change in remote work for occupations from 2019 to 2022. So sort of during the main period of the pandemic and then the change in the generat AI share. So the adopters generat AI after chatt BT is released in the, in the subsequent period. And there's a strong positive correlation where occupations that went more remote are also then much more likely to adopt generative ai. Most of, you're probably gonna be skeptical that just the correlation itself is, is sufficient here because there's a big omitted variable bias. Particular firms in particular occupations that are more likely to adopt particular technologies might also be more likely to adopt other technologies just because they're more innovative or they have sort of more capabilities in this regard. And, and so we wanna deal with the fact that some omitted variable might be driving this relationship between remote work and generative ai. On the one hand, I'm gonna try to control for a lot of potential confounding factors. I have industry fixed effects, I have some within firm regressions where I can have occupation and firm fixed effects. I control for general firm remote work, suitability the firm's generative AI potential and exposure to generative AI education requirements and previous hiring for technology skills in the firm. I'm also gonna come up with an instrumental variable strategy that builds on this idea that remote work, and this is actually a great conference for, for this remote work, is perceived as a benefit by workers, right? And so if other firms around you are offering remote work that should lead to labor market pressure for you to offer this kind of remote work as well. And there's, there are other papers that have shown related results. And so my first instrument is gonna be the degree to which other local employers are remote friendly in the labor markets that you normally hire for before the pandemic. And so I'm gonna, I'm gonna interact that with your own firm te telework ability as sort of a measure of could you actually go remote if there is pressure from the outset. And so the, the instrument is just how tele workable are the other firms in the markets you hire in interacted with. How telework workable are your, is your own hiring all all measured before the pandemic ever happens? I'll have an alternative instrument in case you don't, you don't find, find this one convincing that uses commute time instead of telework ability in your markets as sort of a driver of, of remote work benefits. If you have to drive really far, you're gonna put more pressure on your employer to let you work remotely during the pandemic. And I, I basically find very similar results using, using this instrument. I also do some with infirm regressions where I'll additionally interact all of these with a sort of leave one out measure of how much hybrid work is normally offered in this occupation in other locations in the US as in order to be able to identify a variation within firms in terms of which occupations might have greater labor market pressure to let those workers go remote. Okay, what do I find? I find that that the effect of remote work on generative AI is strong, significant and positive with a 10 percentage point increase in remote work in a particular firm associated with about a 0.4 percentage point increase in generative AI skill demand across firms and within firms, the roles that are, that are more likely to be remote also have a greater generative AI adoption with about a 10 percentage point remote increase associated with a 1.1% increase in, in generative AI skill demand. And so this is evidence for, for what I mentioned earlier, this sort of idea of a technology ladder that you adopted, you adopted remote work and that now causes you to also find it easier to adopt generative ai. This this effect is sort of positive across the board. It's much stronger in the tech sector. And I'll comment more on sort of the heterogeneity later, but it's generally, but it sort of exists outside of the tech sector as well, right? So there, there's a, there's a positive effect in in other areas of the economy. It's, it's robust. I i it's robust using different definitions of what the word remote means. If you look at hybrid or fully remote, whether or not I use this alternative IV and whether or not you put in MSA fixed effects to capture some sort of notion of a tech hub here, you always find this, this positive technology ladder effect. Okay, now let me turn this around. So, so you, once you adopt gender DAI, what happens to the remote workers now? So do you keep them around? Do you use more of them, fewer of them? And so in order to study that, I'll do an event study where I look at what happens to firms that are more exposed to gender. FAI so have sort of preexisting exposure to the technology, what happens to them after chatt BT is released. And so this is sort of a, a standard diff and diff design and I'm gonna look at what happens to their remote work share. However, there's a big issue here and the issue is that for diff and diff we require parallel trends. We want these firms to have similar remote work trends before chat GBT is released. But that is simply not plausible in this setting. You would not, because of all the emitted variable concerns that I that I mentioned to you earlier, you would not expect firms that that have higher generative AI exposure to have the same remote work trends as firms that have lower generative AI exposure because these are different types of firms, right? And so a standard diff and diff is not really gonna work here. I can show you this in a picture if you sort firms by Gen V exposure firms with hydrogen V exposure, that's the, that's the top line, have much steeper increases in remote work during the pandemic, right? And so parallel trends are sort of by definition almost almost violated here. You can interestingly, you can also kind of see that with the chat BT release. You see this, this big kink in the, in remote work trends. And so what I do instead is I use synthetic diff diff, which is a relatively recently developed method where you basically construct under the same intuition as in synthetic control methods. You basically construct a a control group that exactly matches your treatment group's pretre in terms of the variable of interest. And so it deliberately finds firms that have a similar pretre in terms of remote work dynamics before TPT is released and then looks at whether or not this sort of match control group and the, and the treatment group where the only difference between them is, is higher exposure to generative AI then diverge after the release of chat GPT where, so oh, and where, where treatment here is defined as being the top decile of generative AI exposure for affirm. And this easiest to see in, in just a dynamic effect plot. As you can see this, this matching procedure procedure works pretty well. It basically exactly matches these, the control group and the, and the treatment group in terms of their pre-chat bt release remote work behavior. And then after chat BT is released, you see a big decline in in remote work usage for the firms that had more exposure to generative ai. And so this suggests that there is a sort of technology substitution effect where once you start using generative ai, you then use less of, of these remote workers. This is the same in in in numbers. So it's about a 2.7% decline in, in the remote share. And so this should, this should say remote right here in in the remote share. And that's driven by an overall decline in hiring and an even more than proportional decline in, in remote hiring. And so that's about a 20% reduction in, in overall remote worker usage by, by the, the most exposed firms. Importantly, this is not identical across all types of firms. There's a, there's a much bigger effect here for tech firms and I'll show you a little bit more about later about why I think that's happening. But for now, so the average effect is, is negative. There's a, there's a, a decline in remote work for firms that have greater gender AI exposure. What's going on? So I'll, I'm gonna give you a very simple framework for how I think about this just to sort of guide our, our thinking about this. That is, we'll use one equation and two figures. I I promise it's not, it's it's easy to digest. And the idea is that, that we can think of this in a model of production where you're producing this sort of output, you have some automated production, okay? And so there's some automation frontier where you can, you can make more of it automated and you have some remaining tasks that a human worker is doing. So there's some, and there's some non-decision tasks that they're doing. And then there is something that I'll call decision making and then some remote time savings. The decision making piece is the sort of interesting piece here where I'm gonna assume that the worker has not just sort of road execution tasks, but rather they, they produce tasks that have a certain quality, they have a certain decision quality that is produced using some managerial input where your manager supports you in terms of making good decisions over your, over your output. And then I'm also gonna assume that you can get some bonus from working remotely to capture this idea that remote workers have more time available and can be more productive. Okay? And so remote work in the setting is just this idea that that two things happen. On the one hand you get this remote bonus, right? You become more productive in terms of time, you just have more hours in the day 'cause you're not commuting. On the other hand, it makes manager input for worker decisions worse. So it lowers the ability to sort of coordinate with your manager and make good decisions because you're not in the office and talking to them in person. The second, the second piece of this is in the same framework we can think about generative ai where gene AI in this framework is just the frontier of automation expanding and in particular expanding among the most automatable tasks. So let's assume tasks here are sorted. You have some sort of continuum of tasks that we've sorted from lo from high to low automatability. And then there's, there's an automation frontier that can sort of proceed and, and eat up these, these lower end tasks. And I'm gonna assume that different sectors can vary in the degree to which the tasks that are eaten. First by generative ai, the tasks that are most automatable are low decision intensity or high decision intensity. I'm gonna call this a judgment skew of, of generative AI automation. And so you may have different sectors, I labeled this here as tech and non-tech anticipating some of the later results already. But this is just sort of a stylized example where in a tech sector you may have what I I'll call low judgment skew where automation mostly eats the tasks that are relatively low in terms of decision intensity, whereas you might have some other sector. My go-to mental go-to example is sort of education where gender AI automation actually may eat some of the higher decision intensity tasks such as syllabus creation and, and not really changing the decision intensity of the remaining human tasks. Okay? So if you then have a change in, in the, in, in the technology. So if you adopt an EI, you have some automation frontier here that moves across these tasks and then either, either increases the, the decision intensity of the remaining human tasks or leaves it the same depending on what sector you're in. Okay? And then adoption of new technology in this model just requires your firm to invest in some level of IT capital, which importantly can serve multiple technologies, right? So you can reuse some of these IT workers or the, the, the data backend digital infrastructure, whatever it is that you've invested in for one purpose. Okay? And so then you get, as a direct implication of this, you get this technology ladder effect where if the, sorry, where if the, if the remote work time savings are not that large and, and you get, have a lot of this IT complementarity so you can reuse a lot of the IT infrastructure, then if you first invest in, in remote work, it then makes it easier for you to also invest in generative ai. 'cause it's just cheaper. It, it, it also tells us what we should look for when we're further digging into the data. We should look for firms investing in these technology skills when they adopt remote work and then those technology skills also being associated with more generative AI adoption. And that's exactly what I find. So I look at, and I'll show you a couple of graphs that all have the similar structure where I'm gonna look at different categories of skills within a firm. So I'm gonna look first at a category of upskilling that sort of the complexity of tasks, due demand, advanced degrees, high experience, are the tasks repetitive? There's a a a block here that's communication intensity. How social are these tasks that you're, that you're hiring for? There's decision intensity blocks. So that will directly map onto this idea that different firms have different decision making intensity in terms of their tasks. This, their tech skills. So these are, does your hiring mention a lot of technology related skills? And there's also managerial skills. Do you hire a lot of managers and do you hire for a lot of leadership related skills? And so the first result here is that when firms go remote, they hire a lot more for workers with high experience. They hire a lot more for workers with a lot of decision making intensity and they hire a lot more for technology skills and for managers. And those skills are then actually directly associated with greater adoption of generative ai. If you have generative AI exposure. So if you run a regression of generative AI adoption on do you have pre-existing benefits on the technology and you look at what happens if you have more or or fewer of these skills, I find that companies that have more advanced degrees, more experienced workers, more decision making tends more tech skills are, have greater sort of conversion of their preexisting benefits into actual adoption of gendered FEI. And so it looks like remote work skews hiring towards exactly the kinds of skills that then make it easy for you to adopt. Gene fi I, now, let's go back to this, this sort of judgment skew idea. I already mentioned these sort of two examples you might have in tech, you might have, you might have GenFi I taking over a relatively low judgment, sorry, low decision intensity tasks. The, the road coding, leaving the software developer to mostly do planning tasks and think about which things need to be done first and do a lot more. So decision intensive tasks and teaching it might be the opposite way, way around. And so we would expect the technology substitution. So the degree to which you, you keep doing remote work after you adopt gender, DEI you, you might expect that to vary across firms that have different deci judgment skews in the setting. And the reason is that if your tasks become more decision intensive for the humans, that makes the penalty from being remote greater, right? Like it makes it more costly to you that those workers can't coordinate with with one another. And so in, in the sector here, here the tech sector where automation mostly eats the low decision intensity tasks, you'd expect those, those sectors to be more likely to then stop doing remote work after the adoption at FEI. Okay, and I'm just gonna show you the results on, on this. So what I find is with regard to skills companies that are, so the, the red ones are the, the high firms and the, the gray ones are the low firms in this. And this is the, the degree of the substitution effect. So the degree to which they, they end up substituting for remote work. And so the high decision intensity firms have greater substitution, a more negative impact on remote work. The high tech skill firms have have greater substitution and the firms that have good management skills or lots of workers that are good at communicating have less of an impact here. So this suggests that there's some level of, of organizational adaptation where if you, if you have the sort of organization that might make it easier to deal with remote work, it might be easier for you to keep doing remote work. Whereas if you have a lot of tech skills or you're very decision intensive, you're more likely to stop doing remote work. I'm almost out of time. One thing I want to show you is this substitution effect is very concentrated in the tech sector. It's very negative in the tech sector and mildly positive or sort of non-significant in some of the other sectors. And so we really want to explain why this would be so negative in tech and maybe even a little bit positive in something like education. And, and it align, it aligns exactly with this idea of judgment skew. And I can show you direct evidence of this if you look at what happens to your skill composition after you adopt, after the research chatt bt in these different sectors. It turns out that in tech hiring shifts a lot towards decision making skills. So tech firms suddenly start hiring a lot more for decision making skills. And that's exactly this idea that I introduced to you, that that suggests that there's sort of more of a skew where generative AI eats the low decision making parts of your job and leaves the the workers that you still wanna hire to, to need to have greater decision making skills. Okay. Okay. I'll spend 10 seconds on this one. One great piece of evidence of firms that that don't like remote work are firms that have returned to office mandates. And so I'm gonna use those as a proxy for firms that probably have really badly a adapted to remote work. Jamie Diamond always has great quotes that, that sort of support the story from this paper where he says that remote work doesn't work for creativity, it slows down decision making and therefore every, everyone has to go back to the office. And he, in contrast, he loves AI and he thinks it will take over a lot of the drudgery of, of, of jobs. That's sort of this idea of AI eating low decision intensity jobs first. So I, I scraped return to office information from Flex Index and then just, and then just look at how this effect of remote work on generat AI varies for return to office firms. And I find that it's much larger for firms that have told their employees to go back to the office as sort of evidence that they don't really, they're not really coping with gen, with remote work very well. It's, and it's an across firm effect. It's not within firm. Okay. I'll skip the implications 'cause I'm out of time. Yeah. Looking forward to questions.

- Yes. Three things I'm struggling with is I listened, you covered a lot, but great. First, historically, the industry's sectors that were exhibited the most automation were agriculture and manufacturing, which have nothing to do with remote work to first order approximation. And I just like to get your reaction to that. Second, I'm a little concerned about how, and I'm not sure I, I didn't catch how you're measuring employment because in some of the sectors that you mentioned where there's a lot of this AI take up, there's also an increase in outsourcing use of abroad and use of contract work. And I'm guessing that that's outside your measurement frame. Third thing I wanted to comment on you, you had rather wide standard errors on key coefficients given the number of observations in your regressions, which suggests there's an awful lot of stuff in the residual and then I'm makes me worried again about omitted variables, biases of various sorts.

- Can you go into more detail what you mean by agriculture manufacturing have traditionally been sort of adopters of technology, I guess.

- I mean, I think that's where we've had the biggest labor saving tech automation in history in, in agriculture. It goes back, you know, nearly a century, maybe a century in manufacturing. And it goes back at least since, since the 1930s. So, and yet, so I, maybe it's just not relevant, but, but that's where we've seen huge automation. Okay. Which is what you're thinking of. You're mostly talking about generative ai, I take it. Yeah. Okay.

- Yes. - Where there, the issue there is automating of kind of routine, more routine oriented tasks. Well that played out hugely in manufacturing and agriculture. It had nothing to do with remote work. So remote work was not the stepping stone in your technology ladder there. That's the point.

- Yes. I, I think, I mean the, the point I would like to make here is that there is such a thing as a technology ladder and remote work sort of gives us one potential piece of that in, in recent years. But of course this can come through other channels in other contexts, right? This doesn't, it's not, it's not a necessary component.

- And I guess that ties into the omitted variables issue because if there are many paths to which, many reasons to, for which AI might facilitate the automation of these routine jobs, those other things could easily be that are, and you know, there's a whole host of them, I suppose,

- Could - Easily be correlated with the remote work despite your best efforts remote work, takeup.

- Yeah.

- Despite your best

- Efforts. I think in this, in this context, I I, I think the, the standard errors are relatively wide because of this IV approach, right? Like that the, at the end of the day these a sort of instrumented regressions with lots of control variables, lots of fixed effects. And I can only sort of hope that that eliminates all the potential channels that, that, that I can think of in the setting. Yeah. Specific suggestions for what else I can include would be helpful. But yeah, your question regard regarding employment, so yeah, I mostly focus on hiring here. So like when I, when I, when I talk about sort of hiring more or less, like I talk about job po I'm talking about job postings basically. So it's sort of extensive,

- Do employee positions in the United States. That's the point.

- Yes.

- So coding jobs outsourced. That's right.

- Yes. - There's in the same sectors there's a lot of just take up of of of contract work. Yep. Sometimes through leasing firms, some, sometimes otherwise

- I don't, I don't have this in there. I actually think it's a really important question though. Like to what degree? Both remote work and gender API actually interact with the demand for outsourcing. If anyone knows good papers on this, I'd love to see it. Yeah.

- Great. This was awesome. Two questions. First is, when you're constructing the iv, it looked like you were focusing on the MSA level. I I was, maybe there's maybe a quick answer to this, but like why not construct it at the firm level and look like pre pandemic exposure at the firm level? That

- That is what it is. So it's, it's, oh, it's the, it had the word MSA in it, but it's the, it's the firm's exposure to different labor markets that have other firms. Gotcha.

- That would makes sense. Would offer sense. Second, and this goes to some of what Steve was just saying around other omitted variables and the role that organizational practices play in how organizations adopt. So there's a, I have this result on a panel of 20,000 respondents a quarter and what basically finds that individuals that have trust in their leadership not only are more likely to use ai, but there's also more psychological safety that they feel to experiment with ai. And so one, and obviously you don't, you it's with this data, it's hard to get measures of culture. It's may impossible, but I wonder if the va like the duration of the vacancy, the length that the posting is up could be a proxy for how attractive the company is and maybe, but people that work with job posting data, like Steve will have a quick answer

- To that. Yeah,

- Okay.

- Vacancy duration data,

- Maybe it's a no then Okay.

- But, but, but I like the idea of maybe trying to do more with, with culture and, and trust in that. Thanks.

- Well I was gonna say there's a lot of conversation here on behavior and management practices. So if you look at COVID as a shock that had to be reacted to immediately and you one might argue when chat GPT hit the market three years ago, same thing. Is it more sophistication in re in reacting to quick crazy market shifts that happen to include work from home and happen to include what's going on with gen ai?

- Yep. Yeah. So the general sense of which some firms are more flexible or more, more adaptive to general trucks. Yeah, let's go again.

- Yeah, thank thanks Greg. Remote work creates a corpus of text because people start using slack or recording teams meetings and I'm wondering if that corpus of text is the thing that gets picked up if you are then going to train an LLM or use rag or something like that in an LLM to make generative AI more effective because now you can pull in slack messages or teams messages or zoom messages or whatever that gets recorded in systems for the purposes of sort of purpose built ai. Your, your example with, you know, that general or that junior real estate person with a custom GPT is surely trained on the stuff that they collected in the text during the pandemic from remote work.

- Yep. Yeah, no, i, I, I fully agree. I think, I think that's definitely one of the channels and that's actually when I first started writing this paper, that was one of the things I had in mind. This idea of sort of like you're creating your own training, you're creating the training data for your replacement in some ways through remote work, right? 'cause everything becomes digitally legible in some ways. Yep. Other questions?

- Just very quickly, yeah, the discussion this, this lunchtime, which head Jones who was talking about the role of generative ai, we talk a lot about the extensive margin, how, how much more remote work is feasible. But what about the labor augmenting role of AI for like coders like tropic? I think most they, they always talk about the fact that a lot of their most innovative type of coding generative LLM comes from coders actually relying on them to generate and peak complete problems type of solutions. So do you have any thought about it? Because I find judgment a bit confusing. Like how do you at highly cognitive jobs, like the one of coding new, new processing, new software, where does it fall?

- I mean, so I fully agree that I, I think there are lots of dimensions of automation that are not fully captured by the sort of judgment decision making piece. The, at least for coding, I think, I think it works reasonably well in the sense that I think coders do sort of have the like rote writing tasks executed through LLMs and then think more about system design and sort of overall architecture. I think the, the augmentation piece that I think philanthropic talks about in its, its in its economic index is kind of an interesting piece here where they, they define augmentation as actually being something that is about like repeated interactions and sort of rather than sort of one shot getting the answer from something. And that might be an interesting follow up to try to see if I can, if I can map their augmentation measure onto some of this. I am always a bit skeptical that the idea of like if you interact repeatedly with a chat, that that somehow means they're augmenting you rather than replacing you rather than you just sort of bad at structuring your request in the first place. Or like I I, I'm not sure I trust they measure that much, but I think some form form of that measure might be interesting in this context of whether or not augmentation or how different uses of LM sort of matter in this context. Thanks. Yep. Thanks very much. Alright, thanks.

- Hello everybody. Thank you very much to the organizers for this great conference. I am Kara Bernardi, I'm a PhD candidate from Queen Mary University of London and I will be presenting this paper working from home and sorting of female and male workers. So we know very well that there has been a positive trend in work from home adoption at the firm level since before the pandemic with a peak at the pandemic. And today we see that these arrangements have become very popular across firms, but still there is variation, meaning that within the same local labor markets within the same industry some firms adopt, some firms do not adopt. And we cannot fully account for what explains this variation. But if we go to the firms and ask them why are you adopting working from home? Which is something that I can actually do in my, with my data, we would see that three in four firms basically see working from home as a recruitment tool. So they will cite as their main reason for having adopted reasons like improving work life balance for my employees or decreasing their commuting times. But in any case, in 75% of the cases they are seeing working from home has a tool that makes the workplace more attractive. So given this, what I ask in my paper is not only which are the firms that adopt, for example, are they the most productive firms? There is productive firms, what characterizes them but also which are the employees that in the end the, this firm through this recruitment tool is able to have, which are the remote workers that were matched to the firm? So as a natural follow up question, after having characterized the firm side and the employee side, I ask how do these two sides source? So is there a change in the way firms and worker match due to working from home? And to answer this question, what I do in this paper is I leverage a unique data set on working from home. I will talk a little bit more about this in a second, but essentially this is a match the employer employee data set representative of the German economy. Meaning that my analysis is also representative of the entire economy and it has a very good information on working from home at both the worker and the firm level. And I first provide a descriptive picture of working from home adoption with both firm and worker side. What I see, so for the sake of this presentation, I will mostly focus on productivity of both workers and firms. In my paper I also look at other characteristics, but I see that the firms, the remote firms are indeed more productive than the average. But this effect is mainly driven by early adopters. Meaning that as this policy spreads across the economy, the adopting firms become more similar to the average firms. And then at the worker level, I can really look not, I can not only add the characteristics of the worker, of the remote worker with respect to any workers, but I can look within a working from home firm so I can remove the self-selection of workers across firms and within of course occupational education. And I can see, I see that remote workers remain positively, strongly, positively selected even within working from home firms. And then to answer my last question, I look at, I exploit the dynamic dimension of my data, which is a panel data. And I look at what happens to firm's hirings after working from home adoption. And what I see is that after working from home adoption, the average productivity of the firm's hirings increases, but not of all the hirings, but specifically of those that are transitioning job to job. So these are particularly are an already particularly selected subgroup of hiring. So more productive than the than the average. And the firm really picks the most productive, so increases the productivity of the subgroup. This effect is only visible for female workers and not for male workers. I also look at other characteristics of these newly attracted, highly productive hires. I see that they are also retained for longer and that they come from more productive employers, meaning that they are downgrading the firm quality to get this amenity. And I also build a simple measure that helps me strengthen the interpretation, my interpretation of the previous result, which is that overall I see that working from home decreases the intensity intensity of positive assert matching between firms and workers. And this might maybe come as a little bit puzzling since I'm been telling you, oh okay. Both firms and workers are positively selected on working on productivity. So we should see an increase in positive assert matching or maybe this is what intuitively one might think. But to help us really think about this, I have done this very simplified sketch in which I basically express what I see in my data. So even though firms are selected are positively selected on productivity, they're not so different from the average, from the average firms. And instead the work that this firm seem to end up attracting are very positively selected. So if in the end working from home acts as a tool that improves the likelihood of matches between these two subsections of the distributions, this means that it'll overall decrease the intensity of positive ulcerative matching. So for the sake of time, let me skip on the dis on the contribution in a case we're all very familiar with the literature and let me talk very briefly about this data. I would like to say more things but, so this is dataset representative of the German economies called Lincoln personal panel. I use the last four ways of this dataset and it has this very high quality work from home measurement. So at the firm level I have the manager being interviewed and reporting whether the firm allows remote arrangements. And at the worker level I have intensive and extensive margin information I use both to create my measure of re of individual level remote workers, meaning that I can really capture workers that do systematic working from home. And I can match this data with information with a proxy on productivity on both workers and firms. This productivity proxy is the wage effects estimated through the A KM methodology. And what is key here is that this estimation of productivity, which I have for each workers and each firm in my dataset is pre, oh sorry, is pred with respect to the period of my analysis, which is particularly useful for my event study analysis or when I study the dynamic effect because I will be using a productivity measure that is unbiased by the direct effect of working from home on productivity. And I rescale this, the, the wage effect estimated with the A KM because as you may, as you may know, these are just numbers so that are difficult to interpret. So for interpretability, I rescale them. I also have very interesting qualitative information on firms and worker motives and preferences on working from home. I want to be able to discuss this into this presentation. So let's, let me just tell you briefly what I find relative at the cross-sectional level. So as you can see this, these are the results of cross-sectional regression at the firm level and with of course a rich set of controls and fixed effects. And I see that when I have all firms in my sample there, there is a positive and significant relationship between being more productive firm and having a work from home policy. But when I restrict my sample to only later adopters, this positive relation is no more, is no more there. And when I look at the worker side status, so meaning as I was mentioning before, that I am really looking within firm and also of course within occupation within the same occupation I see that the remote workers are different from their colleagues because they are more productive than them. So the last piece of descriptive evidence I would like to show you before going into my event study is this one. So here what I'm doing, I'm trying cross-sectionally to pin down the differential sorting intensity of workers to firms for remote and non remote for the remote and non remote labor markets. So what I have here is a comparison of working from home workers in working from home firms versus non remote workers and non-REM remote firms. So what what this beta two coefficient should be capturing is the sorting intensity in intensity, the between non-REM remote workers and non-REM remote firms. While this beta three will capture the differential sorting intensity for remote workers and firms. So my results suggest indeed that as one might expect the there is positive sort of matching between workers and firms, but when I look at their remote workers and firms, this seems to decrease the coefficient is negative and for female workers is is also highly significant, therefore suggesting this different and lower positive sort matching for remote workers firms. So I then run my event study. So this is the standard event study equation in my case the outcome variable, it'll be run of course at the firm level comparing control and treated firms before and after adoption. And I will have as outcome variables the average productivity of hires or the average productivity of outflows. So outcomes that represent the productivity of the new and exiting workers at the firm after the implementation of working from home. And my treatment can happen in 20 16, 20 18 or 2020. I can study the outcome in every year because I have information on every year, even though I have information on working from home only every two years. And as an estimator I use the local projection difference in difference which you might know is one of these newly developed estimators that tackles the problem of the two way fixed effect estimators. So my result, this is my main result, I, when I study changes in productivity of the inflow of inside the working from home firm, I see that the effect is concentrated among job to job hires. So tho this subgroup of hires that are already positively selected, and this is what I call in my paper of pick, pick of the crop effect. So among the hires that are already more likely to be productive, the this working from pharmacies to become more able to really attract the top talent. But as you can see, this effect is only visible for female workers. Still it's very large and positive and significant. I do not see effect on other hires or workers coming for example from unemployment nor male nor female. And also look at other characteristics of this newly attracted highly productive workers. So I see that their retention increases, so their probability of still being with the firm after one year of being employed increases. And I see also that their average work from home propensity increases. What does it mean that the firm is hiring at each year after working from adoption workers that are very remote? So this is just to strengthen my interpretation that these are workers that are entering the firms because of working from home because they will be remote workers. So this propensity is compute on on the basis of education, occupation and neuro of observation and is compute on a different sample with respect to the one analyzed here, I also do a series of other checks that help me strengthen the, my interpretation of these results. I see that these female workers are coming from a broader local labor market, which again supports the idea that these are remote workers. They are not only more productive when I rank productivity on the entire population of workers, but even when I rank with respect to the firm, they're going to join. So they are really, they really represent an improvement of productivity also when changing in the reference. And interestingly importantly, I see that their origin firm, the employer they're coming for from is more productive than the employer they are joining. So in other words, they seem to be downgrading the quality of their firm in order to get diss amenity. I do not see significant, I do not see any significant effect when I look at outflows. And I also do not see changes in the size of job to job in flows nor outflows. Meaning that the effect that I see is compositional and not driven by change changes in the size. And then the last thing that I do is I want to really say something more conclusive on the intensity of positive assertive matching. Because so far I've been telling you, okay, so these workers that are joining are more productive than they would've been otherwise than they're counterfactual. And these more productive workers are also coming from more productive employers. So this is, I would say suggestive of the possibility that working from home is having a negative effect on positive assertive matching. But it's not conclusive because in the end it depends on what was the level of mismatch before this change. And so this is not direct proof. So what I do in order to have some simple measure that helps me make this evidence conclusive is that I build an hypothetical benchmark. So I take following the basic idea of positive assertive matching as the SC out and culture, I take all of my firms, all of my workers, and I assign them on the basis of productivity, the most productive workers to the most productive firms until all workers are assigned to a firm. And then I look at how this allocation compares to the actual one. So what I end up with is basically a distance, a distance from this perfect positive sorted scenario. And then with my event study I look at how this distance changes and what I see is that for treated firms with respect to their match, with this newly highly productive female job to job workers, this distance increases meaning that we are getting more far away from the perfect sorted scenario. So just to conclude, as I've been showing to you across the German economy, I see that working from home adopting firms are more selected than average. This is mostly driven by early adopting firms. And as you might imagine, the more the policy spreads, the less this selection is evident. And then I have seen that within work home friendly firms, it's really the highly productive workers that are remote even when controlling for many characteristics including the firm you're employed in. And after working from home adoption, I do see that firms attract more productive workers, female job to job transitioners that come from more productive employment employers. And overall this effect seem to widen the gap with the perfect sort matching. Thank you.

- Can you go back to the last result on the reduction in Pam?

- Yeah.

- Okay. Just tell us what's the units, I can't read the units but on the vertical scale, but I also just wanna know how to think about them. 'cause it, it's hard to tell whether this is a large or a small deviation from the initial level of perfect Pam.

- Yeah, so it's a very large effect. The initial distance is around 20 points and it, it increases of around 30 points. So it's,

- That's huge. Well so let me understand you, you have some model and or some note you've ranked on the two sides of the market and there's a hypothetical sorting in which it's perfect.

- Yeah,

- Perfect sorting.

- Yeah. - And you were 20 points away from that initially.

- Yeah.

- And then you're 50 points away from that

- After the Yeah, so it with respect to the subgroup of workers to

- This women. So

- It's not the general of the firm just with respect to this subgroup of work. So female job to job transitioners where you can have more variability of course being a restricted group, but yes.

- Okay. So and okay, so female job to job transitioners and what percentage of all females who are working over this time period is that, are we talking about 1%, 10%?

- What percentage?

- These are just the, these are just the women who actually moved,

- Yeah. From another, from

- Employer to another. And I'm not sure over what time period

- Over this years or so between 2016 and 2020.

- Okay. So let's say it's 20% I'm guessing.

- Okay. Yeah, I think that's

- Okay. So, okay, so among those who did move in that period, there was a very abrupt there, there was a very dramatic reduction in the extent of their sorting.

- Yeah. For in the remote scenario or I mean here what I'm comparing is the difference between the remote and the remote. So

- Okay, it, it'd be good to just draw off the magnitudes of this, that that, that does sound like a big effect.

- Yeah,

- It'd be helpful to help us think about the magnitudes. 'cause it, it's, it's statistically significant but from your description it's actually a huge effect for, for the group that moved.

- Yeah. Yeah. I think it's really interesting and intriguing evidence. I really like your sketch at the beginning as well to sort of illustrate the, the decrease in positive assert M meeting. But I'm just wondering, conceptually seems quite hard to think about what could be the reason why the high productive firms may not want to use this as a recruitment tool to retain these productive workers. Do you have a sense of,

- So there are two things. One is that the most highly product part of the most highly productive firms might have already adopted before I implement my event study. So I might not be seeing them. And another thing is that seeing if firms, as it really looks from my data, see this as a recruitment tool. Maybe if you are already at the top of the productivity distribution, meaning you're already paying the highest wage premium, maybe you don't need such a help for your recruitment. And instead if you are still positively selected and probably this correlates with other things like having good internet connectivity or having being a bit advanced in technology adoption, you might benefit more from this recruitment tool as it indeed looks like.

- Can I, can I ask Yeah. Two, can I two follow up questions. One is your measure of productivity is the, is the a KM wage effect? Yeah. So correct me if I'm wrong, but if for example, the firms that offer work from home also offer lots of other nice things like flexi time and I dunno, you can think of other nice things, they would appear to be unproductive 'cause they don't pay very highly.

- So this is why I'm using the a KM estimation from before the introduction of working from home. So it should be less affected by

- That's true. So you, you're basically relying on Yes, I I, that was, that was a good thing to do. I still might worry a little bit that there's like five things I can do that are nice to employees, like job sharing part-time. I dunno. And I've done four of them and the fifth is work from home. So the, the other thought I had on Steve's I, and I don't quite know how to fix that, but I mean that's just an issue with using Yeah it'd be nice if you had some productivity data you could crosscheck versus

- Yeah.

- The other thing on Steve's thing is interesting thing here is imagine the productivity stuffs correct. So, you know, and I temperature, but the same point here is a bit surprising. But imagine the low, the high work from home firms, the low productives stop, these women leaving the labor force, it's not clear, they're perfectly sort of matching, it's not clear what the counterfactual is. So they may keep them working for a long time. So it's true the first best are in these productive firms, but that's not the alternative. The alternative, they just start working. And so the sec, this may be a second best. So I, I dunno what, it's kind of an interesting question 'cause of their attrition rate is you could estimate the attrition rate actually, you could see it I guess from,

- Yeah, so the attrition rate for this subgroup is relatively lower. This our job to job transition are full-time women. So they, they have relatively higher labor market attachment. But yes, definitely. This is something I, I can look into this but, but

- Can I say something I didn't understand? Usually the story, so the level playing field is after the adoption of work from home. So I'm providing an additional amenity. I see less of a movement towards a PAM outcome allocation. Despite the fact that firms relatively more productive firms are attracting relatively more productive workers. How do you make sense of it? Because usually the story for why you see stationary failures of Pam is that we are sorting on non peary attribute of jobs like working from home. Yeah. So I am Does it mean that it's not, like Nick was saying, it's not a very big dimension, not not a very big non ary dimension upon which workers are sorting to firms or how do you understand it overall? Thank you.

- Yeah. So here when I'm talking about positive alternative matching, I am talking in terms of productivity estimated through the camp. So of course this embeds the problem that Nico was suggesting. Oh, okay, what if the wage effect is not the best way because maybe these firms are offering automatic, and this is definitely something I can think more, but it is definitely relevant criticism. And on this dimension, the decrease in Pam is coming from, as I was saying, that these firms are positively selected but not really as much as the workers are. So there is a decrease in the positive sort dimension because you are increasing the probability that work firms that are at a given point in the productivity distribution match with workers that are way above in their productivity distribution. So perfect Pam would mean that they should sort when they are in this at the same level and instead it's increasing the probability that they match when they are at different, at a different level.

- Hi everyone. Thank you very much to the organizers because this workshop is super interesting. Very nice. Okay, so my name is Francesca Llano. I come from England and today I'm gonna present some results from a project with my colleagues at King's College London, rabbi Rebecca Riley and the Office for National Statistics, Lindsey Brown. So the aim of this project is basically trying to understand work in the post pandemic period and, and of course working from home and the shift towards working from home is a big component of the new settings in terms of work. So we do, so we try to understand how work has changed by using new and newly available time use data that have been collected by the Office for National Statistics. So before the, up until before the pandemic, the UK did not have a kind of systematic collection of time use data the way the US has got with DA two data and now the Office for National Statistics together with ESCO is trying to kind of fill this gap by ensuring like systematic collection of this type of data. Okay, so before talking about how work has changed and how working from home has affected the way individuals use their time, I think I should give a little bit of the ground about working from home in the uk. So here I plot the proportion of individuals who report working from home in the annual population survey, which is a national representative survey that is collected by UNS. And as you can see before the pandemic there was a pretty stable trend for the 10, 12 years before the pandemic, like men and women report men report around 16%, sorry, 16% of men report working from home mainly or partly and around 10, 10, 11% of women report working from home after the first two months of 2020 with introduction of the first strict lockdown and then the following two strict lockdowns, we see that this proportion steady increases up to reaching around 30 to 33%. Very interestingly, also the gap between men and women working from home kind of closes down. Okay, so given this premise, we know that the UK has been together with some other countries, one of the places that has experienced the most, kind of the biggest and most permanent shift towards working from home. I probably don't need to motivate to you as a crowd why it's important to study more working from home, but let's say the working from home has several implications for employees, employers and policy policy makers in terms of wellbeing, productivity, urban organization. And on top of that we saw like a, a vast body of research that shows that actually workers willing to forego part of their earnings or earnings growth in order to actually work from home, from home. Therefore, understanding a bit more why workers want to work from home is a pivotal issue I think, and time used data kind of can open a little bit of a window on why that's the case. Okay, so here we try to answer four main research. Question one probably is less interesting for you. The first that is we want to understand whether these new systematic collection of time use data can be used to track the trends in working from home. And the reason is because the UK doesn't have many repeated collection of data. So having multiple sources to try to track this behavior is important. Secondly, most importantly we want to understand whether working from home kind of affects the way individuals allocate their time across different activities and also whether there are implications for enjoyment. What I call IMP enjoyment is the, the instantaneous enjoyment the way individuals feel about an activity they are kind of implementing in the moment and also on self-perceived productivity and more general and wellbeing. And finally, whether there are gender differences, gender is always my main interest but is particularly interested in the context of working from home because we know that working from home from home might kind of drive or increase gaps that already exists. Okay, so what I do in this paper very simply is we use this new time use data to kind of produce a descriptive analysis of differences in time, use patterns between individuals who report on the day working from home versus individuals who report working from the office. We also this as I said, for work both for all the other main act, all the other main activities during the the 24 hours. We also try to understand whether there are differences in, as I said, enjoyment, self-perceived productivity and wellbeing. Okay, I skipped the contributions. Okay, data. Let me talk a little bit about this data because, so for part of this data we've been helping with the collection and we feel very proud about them. So as I mentioned before, the ONS started collecting new time data from April, 2020 because they understood that it was very important in the moment of the implementation of the first strict lockdown to understand how people were using their time. So since April, 2020 there have been nine waves of collection of time used data. The last was in March, 2025 and the the respondents are sampled from the nuts and panel. So far we have around 7,500 diaries of individuals who are employed for weekday, sorry, weekday diaries, individuals who are employed and reported list for our work. In each diary in this, for this particular results that I will present today, I will focus on the post pandemic period, which is from November, 2022 to March, 2025. And this includes around 4,400 diaries. So what is quite important about this collection is that respondents who agree about answering this survey are given two random days, one weekday and one weekend day and they have to report their time use during those days. They cannot change the days they can and if they report a different day or if they report less than two days, they're dropped from the sample. So they, their diaries are not included. So what is interesting about these time use data, of course time use data record the primary activities, that is the main activity you are doing. The respondent is doing over 24 hours, but also secondary activities where secondary activities are kind of things. Some sort of, you can think about it as multitasking, whether you're doing something while you are actually kind of also conducting your main activity. They, this diaries also collect is spontaneous enjoyment. That is how much individuals are enjoying on the spot of the activity for each episode reported in the, in the diary over 24 hours on a scale of one to seven. And then also we added a question in the last couple of ways about self-perceived productivity. I'll show you a little bit about how we did that in the survey. And so we kind of collect information about self-perceived productivity. This question is asked every time the individual reports one episode of paid work. So if there is an episode of pay works, individuals has receives a pop-up, they ask about enjoyment and productivity. I'll show you in a second. These surveys also collect a comprehensive set of socioeconomic characteristics. Several respondents respond to more than one diary in, sorry, more, sorry, more than one wave. However, at the moment I'm treating the waves as kind of sep different construction so I'm not exploiting the the kind of pan dimension. Finally, we con, we think that time diaries are kind of gold standard to kind of track behavioral changes, particularly because they minimize the recall bias and are less influenced by social disability. This is the way we see this type of data. Okay, so this is just to show you how we collected information about enjoyment and self-reported productivity. So until wave seven this popup window would open every time an individual was selecting one episode and was only about enjoyment. So for each individuals had report enjoyment between how much danger the activity between one and seven. In the last two ways we added the second part of the popup where we ask only for when individuals report paid work in this case is paid. You can see from the windows paid work from home whether they felt productive. And the way we define this productivity is with five values, whether they're 100% productive up to less than 70% productive. Okay, this is just very quickly to show you how we use time, use diaries to kind of add an additional measure of working from home and track working from home over time. Here I report the proportion of diaries, the report working from home either any time working from home at least for hours or the whole time working from home for the nine waves. And I compare, compare them to the most recent pre available pre pandemic time you survey that was collecting in the UK in 2015. As you can see, there is a big change between 2015 and the pandemic, the proportion of diaries reporting working from home increase and then kind of stabilize after November 22. So November 22 to March 25 is when we are focusing the results I'm showing you today. Okay? More importantly we want to understand how time across different activity differs between individuals who report on the day working from home versus individual report individuals who report working from away from home. We, we do so by running separate regressions for each type of activity. And what we do is basically running 10 regression, as I said, one for each type of activity where on the left hand side we have number of minutes for each diary spent in each activity. And on the right hand side we have an indicator for working from home gender, the interaction between the two and the a comprehensive set of individual characteristics. Here I report the, the estimated coefficients expressed as percentage differences relative to the mean number of minutes for each activity reported by non-working from home workers in order to give like a sort of mag magnitude to evaluate disco coefficient. So basically just to to be clear, if we look at sleeping, we see that individuals who work from home or individuals who report on that diary, they're working from home the day they sleep around 7%, 7.5% more than individuals who report working from the office. This is for men and around 4% more for women. What is more interesting of course is that the report around 76% and 71% less time in transport as you can imagine. And they seem to trade these saved time to spend more time in unpaid work. This is true for both men and women and unpaid care. Again, this is true for both men and women in terms of magnitude, both basically men and women spend around 20 minutes more per day working from, sorry, in unpaid work if they work from home. So this is, this 38% is around 25 minutes. Then we also saw the women spend more time watching television. Watching television like individuals report a lot of time watching television or Netflix or it's, it's quite, it's quite daunting actually to see it and I, it is crazy. I mean if you look at this, yeah. Anyway, and so women report spending a bit more time and in terms of leisure, they seem to, both men and women seem to spend more time in leisure related to wellbeing only For men, we find a very small negative coefficient for paid work. So they work around 20 minutes less per day if they work from home. But this is mostly driven by nongraduates working from home. If we look, if we separate the sample between graduates and nongraduates, graduates don't have this negative effect on this negative coefficient on working from home. We do the same for secondary activities, what we call multitasking. Basically we look at whether individuals report more time in other activities while at work and if they work from home. And basically we find that individuals work from home report more leisure while working and both men and women. And in the case of women we see like an increase in work in, sorry, unpaid work. It is not massive, it's just around like something like four or five minutes. But still it's quite big in terms of magnitude considering how much this is reported as secondary activity. Another interesting thing about time use is temporal flexibility. How individuals kind of arrange their working time over the day. I'm not gonna go into the details of these, these estimation, but basically if you look at the blue bandwidth, this is the 95 confidence interval of the estimated difference between the probability of being in work for every hour of the day, over 24 hours between individuals work from home, individuals work away from home. And what we gather from these figure is basically individuals work from home are less likely to work early in the morning, less likely to work a kind of lunch around lunchtime, but more likely to work in late afternoon in the case of man sometimes late over the day. So there is some temper of flexibility. They rearrange their day as you can imagine. Now so far we've seen, let's say the positive aspects which are about flexibility and kind of using the save time from commuting towards things that they need to be done or things that they want to do. What about enjoyment? Instantaneous enjoyment for working, for paid work in particular working from home versus working away from home. So what we do here is we run, we we try to rank preferences about activities we do so by kind of running regression where we have on the left hand side the enjoyment for all the activities reporting during 24 hours. So we have multiple observation for each diary and on the right hand side we have the different categories of activities where the baseline is leaping and personal care because these are kind of essential activities that cannot be excluded from everyday life. What we find this way is that individuals seems to consistently prefer, prefer working from home compared to working away, sorry, working away from home compared to working from home. We also find what is interesting about this is that we also find that traveling is quite disliked as we can imagine and as disliked as working from home. So in a sense, individuals who work from home don't seem to enjoy it to be individuals, work away from home, enjoy it more, but kind of dislike the traveling. So a kind of trade off we can see a kind of trade off here. Okay, what about self-perceived productivity? I mean I'll let you guess, but what we find is that working from home is associated with a negative self-perceived productivity and is particularly true for men. If we look at the, here I report the separative productivity as a continuous variable or for different cutoffs and we can see that for the highest cutoff, which is being productive 100%, both men and women are around five percentage points less likely to report being being productive at 100%. Okay, so I think Sava who just arrived, made a comment about this, sorry, made a comment about these results, mentioning the fact that of course there is some selection and is very linked to the comment that was made before about your paper. So it could be the individuals are less happy or enjoy less working from home and they feel less productive because of a difference in tasks. We can't test that with the time used data we have that I used so far to show you something. But actually we use a difference set of data collected by the NS in 2024. Among public sector workers in these time use data which are experimental and incredibly cool individual are asked to not only report their activities out, work in a kind of more aggregated way, just leisure, sleeping, personal care, but they're asked to actually say exactly how they spend their time during their work day in a sense like they have to, to say how they spend their time across different tasks. Now these, these is run in a way that each public sector worker has to report to two days here I focus on the subsample of individuals report one day working in the office and one day working from home and just descriptively we can say that there is a big difference in terms of non-occupational specific tasks and time spent in non-occupational specific task between working from home and working from the office. So clearly individual work from home, spend more time in admin and these, it's plausible to think that is my kind of drive, the type of sense of less enjoyment or less self perceived productivity. Okay, so just the very last thing about, okay so far, so we've seen more flexibility in terms of time, probably better use of time, but also less enjoyment and less perceived productivity. What's the total effect? Purely observationally on wellbeing. We have in the survey five measures. So wellbeing, we see that there's no, sorry, rather than effect association, what's the association between working from home and this measure wellbeing. We see no association like individuals who work from home don't report less life satisfaction, less happiness, more anxiousness or worthlessness. Okay, so very quick I will conclude our 20 seconds remote work. Persistent possibly is driven by time saving and flexibility, but not enjoy necessarily enjoyment and product productivity and probably wellbeing benefits kind of arise from outside of work, future research, definitely collecting more data because I do think the task and time span task and type of tasks is very important in understanding why individuals enjoy maybe less working from home. Okay, that's it. Thank you. Yes.

- On on the temporal flexibility, I'd love to see it broken down by people with chil young children and those without, because there's lots of other evidence that suggests that it's people with young children that put an especially high value on the temporal flexibility. So it'd be good to see that chart broken down that way. Yes. I really wanna see the question about self-perceived productivity. I mean I've used these questions as well, but, but exactly how you frame the question is potentially quite important here.

- Yes. So I have to say a little bit about this question. So the way it was framed, let me go, sorry, this is the, the way, can you see the bottom of the popup where it says, during this time roughly how productive were you? 100%, 90 to 99%, 80 to 89%, 70 to 79 and less than 70. So this question was originated by several focus studies that were done by the cabinet office in the uk, by the government

- During this time. Means during time you're working at home on a job where you do both,

- This question pops up for all types of work, work from home, work away from the home. So

- Oh, I see, I see.

- Every time an individual,

- Every time you're working paid

- Work. I see. Yes, every time.

- Okay, got it. Okay, great. And the other thing, the other, the other breakdown is married versus non-married on the enjoyment wellbeing stuff,

- Is it big? I don't, so I've done it because I, I have another paper on loneliness and time use and, and so while during CID that was crucial, this difference between married non-married parents, non-parents and quite interesting, interestingly, women enjoy more time with others than time with the family. Kind of understandable sometimes. But the, in this case, for the period post pandemic, I didn't see a much difference. But thanks for the suggesting because actually something I should add in the appendix to show this. Thank you. Yes.

- Great. I I had a que this section is on mental health, but there's an interesting question on physical health. So just an observation, it's this old literature that people here probably know where you look at countries which have exogenous retirement dates, like I think in Germany it was 60 and it went to 62. You see a big spike in mortality after retirement and it just stays higher. So oddly enough people think, oh, I'm healthier when I'm out of work 'cause I have all this time, but you don't, you know, there's a, there's a big increase in death rate for various factors. But if I look at your study here, I was looking thinking of physical health, the two things that people are spending more time on, one is sleep, which I think is increasing evidence. Actually most adults don't sleep enough. And the other that's much more important is exercise. So the the like health evidence and exercise is incredibly positive, like unbelievably positive. And so, I mean there's something just out, I saw this week from Stanford saying, you know, there's no, there's no drug or therapy for many things including multiple cancers that is better than exercise. This is incredible. So, and physical health, it would be interesting if you had any evidence on this. Like,

- Yeah, so that's the, the, the result I showed is this what you mean in terms of time spent? So

- No, I saw that if you had physical health outcomes, so like you have

- Mortality

- Rates or visits to hospital or, or if anyone else has it,

- We have some. I think it's a rough measure on health. Thank you. No, no, I actually use it as a control whether, whether individual report any health problems. But no, we don't have, also, since it's not a panel, we can't really track whether this change, but

- It's basically conceptually two by two work, not work. And if you are work, work on home, work in person and it looks like the work not work, you know, work is positive versus the least the evidence I've seen. But then conditional working this stuff is almost suggestive at least some days at home because of time use. But it would be good to see this.

- Okay. Okay. Thank you. Thank you. Yes, good point. Thank you very much. Yes.

- Yeah, actually maybe really related to this, I think this is really interesting because you have to sleep more, exercise more. I'm wondering about whether they eat healthier and I know I talked to Stephanie yesterday about, you know, whether you can see something about whether they spend time more cooking and doing things that might be actually good for them. But then on the other hand they see less people and I find it interesting to see this sort a contrast between the two papers of how, you know, the first three really should affect mental health as well in a positive manner. But then not seeing people is, is sort of playing a, a negative role. And I wonder if there is a way of putting all of this together.

- Okay, I'll think about it a very good point. Yes, yes. Sorry, sorry, I thought someone

- I very interesting. So I remember these couple of papers by bene region quarters on the uk. They were actually saying that at least during the pandemic in the UK perceived this productivity was going up for remote workers. So that was specific of the pandemic. So how do you reconcile,

- Sorry, which papers? By whom?

- Ben Ridge and

- CORs Ben. Well maybe also the question No, I, i dunno exactly the paper by Ben. I suppose he uses understanding society because this is what he usually uses. I, I, I dunno exactly. It's something i I need to see

- To comment on this one huge difference which you can see in your study is what you define work. Because you can be 10% less productive at work, but you save 20% of your total work plus commute time. So it's like this, let see is huge because if you look at, you're saving an enormous amount of time commuting and people dislike it as much as they dislike work. So if you are 5% less productive in the office, but then you don't have to commute, you're actually net more productive, including commuter. So actually these questions we've been looking as Jose a lot. It, it's not clear actually what, it's not clear what you want to define, but I think in your case you're defining actual minutes worked, but you're excluding work commuting, which is often, like in fact all these studies show commuting's typically hated even more than working. It's like people hate commuting.

- So I, I also had one thing about that is that in understanding society, the question is direct. Like, oh do you feel more productive working from home if that's the, the, the, the data set used. Whereas here, the thing I was saying about the bias and the, the social disability, the good thing about time is data is that people report things without really thinking about a direct question. The kind of has some implications so they understand whether there are some implications to the question or not and might respond in order to make the the person asking the question kind of happy. So I think that's the main difference and that's why we like to collect it. We did the same for ai. We added a popup question asking whether they use generative AI every time they report the timing work. So it's quite nice because actually you get slightly different percentages compared to other studies that ask direct question in a kind of questionnaire setting. Yes, can you bring

- It out by type of job and whether they're, can you break it out by type of job or occupation? And also whether they're a manager or a non-manager because some of the activities like meetings and supervising it might fit with some of the papers from yesterday about manager or monitoring and things like that because I, I thought God would be important.

- So can I can definitely do that for the other paper, the one using the public sector time use survey where we have very specific occupations as well as tasks. So is yes here we have occupations but we have a big, for the moment we have a big proportion of missing us. And the reason is because the ONS uses an algorithm to code strings into occupations, but we're trying to solve the problem and kind of gut 90% once we have that, that's the aim also to look at the bioc occupation, whether there are main differences and particular to understand whether self perceive productivity or something to do with the type of occupation or kind of is different. Yes. Sorry, sorry, I, I saw you before.

- It's super interesting. So we have some similar line of work in the United States. One thing we real we found is actually a secondary childcare increase when both men, when people work from home. And then if you look at what they're doing when we look after children's, basically watching tv. And I think it would be super interesting if you could collect data on children's term use. I think, sorry, on children's, in children's term use and wellbeing. I think that has been a gap in the data collection literature

- That is pretty difficult also because it's usually parents filling in the diaries. So yeah. But no, it would be amazing. And that opens all other kind of research. Yes. Finished. Okay, just one minute. Any other question? Okay, yes, sorry,

- I just want to add something to the discussion. I think maybe a nice idea of like reconciling the previous presentation and your presentation could be like there, there's a future discounting bias that in the moment people see the, the advantages of more sleep, more exercise, but that they underestimate a negative effect on the long term of like building social capital. So I think that's an interesting thought of yeah, like the difference between stated preferences and like the impact on mental health. And so for,

- For me immediate perception as well. Yes.

- Yeah. That there's a kind of future discount on

- This.

- Yeah, definitely. Yeah. Yeah. Okay. Thank you. Okay,

- Thank you very much.

- Okay, good morning all. It's nice to be here. Thank you very much to the organizers for having me here. I am Samir. I am a computational media psychologist. I work at the Harvard business with Mike Norton and Ashley Willens. I am a psychologist, not an economics, so excuse their smeta on the slides. We tend to take them very seriously. Okay, so let's just go ahead and get started. I am preaching to the choir here, but it is important I think to set some common ground for this presentation. So let me start with what I hope will be a joke and not an overused cliche, but I found it really funny when back in the day the CEO of Zoom essentially stipulated a return to office policy for the entire company. Essentially arguing that zoom's own revolutionary technology was not digitally rich enough to emulate in-person sociability. And this was ironic, but it was also emblematic of a significantly broader return to office mandates range across the big tech industry really. And when CEOs were pressed about why, you know, they were insisting on this, they said something along the lines of sociability creativity being a miss in hybrid models of work, particularly on remote days. And, and of course we have scholars here who have argued that, that there are financial incentives and the CEOs making those claims. But you know, it's also possible that the digital media technologies that we have at our disposal are just not sufficient for motivating or for emulating, sorry, the the in-person sociability that in-person interactions offer. So, so let me with that dive into some past research and kind of assess if this is really the case. Obviously we have a lot of research that suggests that hybrid work can actually improve retention. You know, we know that people prefer working in the office for three days or so and that nearly 21% of remote capable workers are, are willing to take a substantial pay cut just to be able to work remotely. So we know that that employees want this, but what we don't know is if the digital media landscape that currently exists is sufficient enough to offer the kind of interactions that constitute in-person interactions. And so that's going to be one of the primary focus of my talks today. So really we have three key foci in the current presentation. The biggest one is looking at day-to-day sociability. I think this has been really nicely set up by the scholars who came before me and, and kind of wanting to study employees over time. We take it one step further, we don't just want time used. Time used is great, it's very valuable, but we want to examine employees multiple times a day, ask them rich psychological questions about their experiences and understand if those constitute meaningful work experiences for them. And as a computational media psychologist, I was actually trained here with Gabby Harri at Stanford. I would be remiss if I didn't study digital media technologies in the richness that, that these technologies exist in the workplace. So that's a, a second key focus for the current research. And, and and really what we are interested in at the end of the day is asking this question if, as to if employees have become adept at managing their sociability and hybrid arrangements, right? So it's, it's fully possible that the technological landscape does exist, but the question as to whether employees are good users of these digital media technologies is an open question. And this really requires us from a researcher perspective to unpack the microdynamics of hybrid employees day-to-day sociability, both in person but also on digital platforms. Okay? So within this broad kind of umbrella concept of workplace sociability, we are particularly interested in what we call informal work interactions. And so Ashley has done a lot of formative research in this area where she finds that largely from qualitative evidence that management consultants in this case reported that informal interactions selectively went missing during hybrid work or during fully remote work. And these interactions are essentially characterized by their ad hoc nature. They can occur in a, a variety of contexts, you know, they don't have to happen in a meeting room. And in the absence of these interactions, workers kind of consistently alluded to missing out on a shared understanding our essential grounding work that they, that, that they did with their colleagues in order to make meaningful progress on their work goals. Okay. And so I would like to emphasize here that you know, you can, you can clearly see from these, these results that there is an emphasis on on a variety of contexts. So everything from Ubers to hallways. And so it seems that, that these interactions are context agnostic really. And, and that seems to be a crucial part of what makes them unique. So drawing on a rich management science literature, OB scholars have studied this behavior for many for many decades. We kind of define informal work interactions in the following manner. So we define them as interactions that happen in passing that can help give employees a new perspective on their work. We can go into more detail about this if necessary, but, but they are work related interactions. How that distinction is drawn should be questioned and interrogated further. But essentially these are informal interactions, unplanned, ad hoc, spontaneous and crucially they are work related. Okay? So that's kind of like our independent variable, what we care about in terms of workplace experiences and employees experiences at work. We, we focus specifically on psychological safety. Psychological safety is a well studied concept by Amy Edmondson. And essentially she finds that having this to take risks without the fear, without the fear of negative repercussions is a really crucial element of being productive and innovative at work. And for all the economists in the room, a lot of this past research has measured productivity in very objective terms, you know, and has has found these robust correlations between increased feelings of psych safety and increased objective measures of productivity and innovation. Okay, so with all of that background we have three research questions. So first, as as, as the scholars before me have set up diary studies, we study employees over the course of multiple work days and that kind of allows us to introduce exogenous variability in their work locations. And that kind of allows us to ask if daily work location predicts just the frequency of informal work interactions that employees are having. And then we ask if these informal work interactions are meaningfully predicting changes in how they experience work and and their work related outcomes. And so essentially the theoretical model is as follows in, in kind of one way of thinking, working from the office should be associated with an increase in the number of informal work interactions that employees have. I will go into a bit more detail as to why we suspect that might be the case. And an increase in informal work interactions should be associated with positive work experiences as well as increased self-reported work outcomes. Now the competing argument, which can also be fully plausible is that employees are master users of digital media technologies and regardless of whether they're working from home or not, they should be able to emulate the informal interactions that they have in the office on remote days using these digital media technologies. In which case it's kind of really anybody's guess as to what's going on and we might even expect to see a reversal of this conceptual model in terms of valence. Okay, so those are the two competing hypotheses. And now let me kind of dive into the sample that we collected data from. We collected data from hybrid employees. These hybrid employees provided nearly 5,000 observations multiple times a day over the course of seven consecutive work days. We took some, some measures to ensure that employees were generally representative of hybrid knowledge workers in remote capable industries in the United States. So a lot of employees came from IT, financial services, insurance and education industries. Importantly, these employees had to report working in a hybrid arrangement. So they had to spend at least one day in a regular work week working outside of their primary secondary office and our client customer office at a third location such as a cafe or their, and so I'm just, I'm sticking on spending some time on this because, so folks have a good idea of how we define hybrid work and it was kind of done very much in line with past gold standard research. Thanks to everyone in this room. We were definitely, and and very much focused on employees who are physically located in the United States. We would expect meaningful cultural variations here. So it's important to set very explicit cultural bounds. Obviously employees had to work regular hours and they had to be remote capable. So essentially we excluded employees who indicated being manual workers, laborers are are general staff such as receptionists or cashiers just because we wanted to have some kind of operationalization as contentious as it might be of knowledge workers, quote unquote, which the first three categories seemed to satisfy. Okay. So our service strategy was as follows. So building upon a lot of the past research, we surveyed employees multiple times a day, once at noon, once at three, once at six for seven workday employees during each ping completed either a five minute or a seven minute survey. The end of the day 6:00 PM survey was the seven minute survey which included a wider battery of questions that we didn't ask during the workday, you know, to prevent people from not working during our work service. And essentially these daily surveys measured self-reported psychological safety, primary team trust work engagement work, creativity, team learning behaviors. And crucially we assessed the employee's daily work location, which was the location they primarily worked from at the day of the survey, the extent to which employees were physically co-located with their teams. So this is really crucial because even though employees can have office days, if their team is not physically with them in the office, they're actually gonna resort to using a lot of digital media technologies and some of the benefits of in-person work might be reduced because of a lack of physical team co-location. So that was a crucial variable for us. And then employees also reported on the frequency of their informal work interactions. And then they also told us about features of those informal work interactions, namely their interaction partners, whether or not these interactions happened in person online, which channels they used to have these interactions, the extent to which the interactions were synchronous. So were they having rapid fire text message exchanges with their managers or was it something more inertial with, you know, a direct report, like an email, like a consecutive email exchange spanning tens of minutes as compared to a rapid exchange over text and if these interactions were recorded and stored on organizational premises. So there is a lot of research in media psychology about how the recordability of interactions kind of influences their utility, the extent to which people are candid on channels that they know are being monitored. So we decided to measure some of this variability as well. I should say that all of these measures in classical management science have typically been used in single one-time surveys. We conducted two qualitative, so there was one qualitative diary study and one quantitative diary study that was used to validate these instruments, validate their psychometrics. I can go into details about that. And then in the third study, when we were confident that our measures were actually capturing what we intended for them to capture, we deployed them in the full study. Okay, great. So for the results, what I'm going to present to you are person level analysis. In these person level analysis we're comparing a person and employees psychological experiences when they're at work as compared to when they are at home. So these are within person comparisons, they're conservative 'cause we are averaging out whatever variance might be attributed to to different people. But before that, sorry, my apologies. But before that, let me show you some descriptive statistics and, and then here the remarkable finding is that apart from the frequency of informal work interactions, we don't actually see meaningful differences in, in other features across remote and office days. So employees are more likely to have frequent and formal interactions when they're in the office, but, but we don't see statistically meaningful differences across how they're having these interactions. So people use different digital media channels roughly equally. And, and quite importantly, the, the frequency of formal work interactions does not change much across remote and office days. But, but let me kind of go back to the, the theoretical model that I was talking about. So looking at these within person mediations, we essentially aim to predict psychological safety just from employees work locations that day. And we find that there is basically a strong direct quote unquote effect. It's just a mediation terminology. It is not an effect, it is a positive association, you know, just to be precise. And so essentially working from the office is associated with higher feelings of psychological safety at the daily level. And this association is partially mediated by the total number of the, the total frequency of informal work interactions that employees have that day. So in other words, working from the office is robustly associated with having more in more in informal work interactions. And that in turn is robustly associated with an increased perception of psych safety. There is a lot to break down in this mediation. So essentially, let me just get into that in a minute, but I, I do want to emphasize that even though broadly there were no meaningful differences in the features of informal interactions across office days and remote days, there were some interesting findings with regards to using specific digital media channels and feeling different levels of psychological safety. So interactions that take place via video calls or instant messaging are, are usually associated with lower levels of psychological safety in comparison to in-person interactions. Okay, so, so that's pretty interesting. But, but let me break down the overall mediation into a more sequential set of analysis. So essentially there is a, a very strong relationship between being in the office and being physically co-located with one's team. So this is most likely because a lot of offices have, you know, they they require the team to be physically co-located on specific days. So that's good news. People generally are coordinated and when they're showing up to the office and being physically co-located with one's team is associated with, with having a greater number of informal interactions that day, which in turn is associated with increased psych safety and a range of positive work outcomes, all of which to be transparent are self-reported. Okay, great. Great. So with that I would just like to summarize some of the contributions here today. So essentially what we see is that first and most importantly constructs that management scientists have studied as, as relatively ossified constructs, as constructs that have been conceptualized as inertia over long periods of time actually display meaningful variability over the short term, particularly in the era of hybrid work where daily work location varies a lot right across days. And, and so this might be introducing more variability in, in, in workers experiences as well as their perceived outcomes. And we really do see that where employees choose to work from every day is, is crucial for understanding some of this variability. Physical proximity to teammates and frequency of interactions seem to be important predictors and in line with the, with a rich tradition of digital media research, we find that digital media is not adequately substituting in person interactions. Employees tend to have better informal work interactions in person which are associated with meaningful work experiences. And so essentially we need a combination of both. One could argue better digital media technologies, but also perhaps media literacy interventions that help hybrid employees use digital media technologies in a more meaningful manner in order to really get the best of both worlds that hybrid work promises to us. Okay, so I'm going to stop there because this ongoing research slide is slightly tangential and let me just go here and thank my collaborators and thank you all for listening to me and I'll take questions now. Yes,

- This is very interesting. Thank you. I had two quick questions. I didn't totally understand the setting of where you were finding your survey participants from, but then I think the other thing that I was interested in is do we see sort of people compensating? Maybe it's if I'm a hybrid worker, I, yes, I'm more, I'm having fewer of these interactions on the days I'm home, but then the days I'm in the office I'm doing very little work because I'm just doing the informal interactions. And so maybe over that whole time period I'm actually just sort of shifting when these informal things are happening, which I think maybe with variation in how frequently people are working from home, you might be able to actually capture and answer how much of this is sort of just shifting around versus total losses because I'm working from home.

- Yeah, so to answer your first question, these were participants that we recruited on prolific. They had to meet a very stringent criteria in order to constitute hybrid employees. And yeah, and we basically did the diary study with those participants, but they were sourced on prolific. And then yeah, we can get into details about how we verified their humanity versus their ai. The second question as to whether people are displacing some interactions across, so like, you know, within a work week, currently I'm running sequential analysis to see if, you know, if a particular day that's characterized by unusually high informal work interactions in the office does that day tend to, to lead to our, our predisposed subsequent days to not have those many informal interactions. Some pilot results suggests that there isn't like a very strong sequential effect. So people tend to be very clumpy in terms of having, like Mondays will be the office day and that's when I'll have my most informal work interactions and Fridays are always work from home. So then I'll, I'll have like no, you know, variation in my informal work interactions that day. And there are these like very, very strong temporal patterns that we see. And it's most likely a function of how the organization is like instructing employees to, to work from home. Yes, yes. Nick Nicholas. Yeah.

- Great. Really interesting. So two comments. One, oddly enough I asked Eric Uan why he canceled work from home at Zoom and he said a, he's moved to hybrids just to be clear so everyone's on the same page. So he said now the sales team has to come in Monday, Wednesday, the engineering team, Tuesday, Thursday. He said, we've grown so much, there's not enough space in the office, which is why they come in on different, he said they don't overlap that much. And he said, look, most of our clients are hybrid and he, he's literal phrase says, we've gotta eat our own lunch. So he said, look, our hi, our clients are hybrid. It was unclear whether I, you know, that that was his explanation for it, which is partly like, but it would be consistent with your thing, but it's partly this is how people are using it on the psych safety. My one interpretation, which you know, I was at McKinsey and even now is that you'd kind of pre-vet stuff. So certainly when I was at McKinsey, if you were gonna present something sensitive, what you would do is kind of walk the hallways a bit and check in with people kind of informally. You'd see them in the coffee room and say, Hey, I was gonna say this in the meeting, is that okay? And that would be very consistent with your findings and you don't need to pre-vet it everyone, you just, you know, you're gonna present to 20 people and you run it by three or four people in the room informally and they're okay with it. Whereas on Zoom you may not have that chance to do it. And I guess that's, that would at least be part of the explanation.

- Yeah, absolutely. And this was not, I mean Zoom is a great technology, it's just like with all video conferencing technologies, one has to plan the meeting in advance and that just takes away the opportunity to have informal work interactions. I know some companies have experimented with just having like office hours where when one leaves their zoom on and people run into others. But yeah, just thoughts about how these technologies can be better designed. Yes.

- So I was just you clarifying questions. Yes. So is this all comparing within person that it's like

- Yeah.

- Comparing their informal interactions. Yes. Okay, that's helpful to know. And then did you ask them in the beginning like just what their schedule was? I guess I just could be worried that like on a day that you don't feel great about your colleagues, you might decide to work from home. And so do you know like this is my normal days that I work from home versus work in the office?

- Yeah, yeah. We have data on what their plan schedule is for, for the week that we collect the data from.

- Okay. Yeah and that might just be useful to check.

- Yeah, that's a great point. Thank you. Yeah. Yes.

- Are you familiar with Jessica Met Hot's paper on Small Talk? It's an academy management journal and it talks about stall talk as a ritual and that it leads to positive emotions and then that helps people with OCB. But she has a measure and it seems that this idea of small talk and office chitchat is something that uplifts people emotionally while working and it seems very relevant to your paper.

- Yeah, definitely have heard about that paper. I think it's extensively cited in our working paper. I don't think that paper distinguishes between informal work interactions versus just like general small talk about life. Perhaps it does, but yeah, thank you for the,

- They define it as talk not related to task completion.

- Okay. Okay. Let's, let's, let's chat more about that. Yes sir. Yeah,

- I'm, I'm glad you're working on this, but it's such a complex terrain and I find the conclusions people draw to are very much related to their prior. So I wanna make a couple of observations first. It's not even clear apriori whether ad missing out on ad hoc interactions is a net plus or minus. There's an opportunity cost of time, the small talk chit chat and so on. And there's a greater opportunity to choose whom I interact with when I'm working remotely than when I don't. So I gotta trade off the water cooler conversation about the football game versus I can remote talk to Nick and you know, that that might be more productive if he's having a good day. Sec second point is there's just tremendous variability In that opportunity cost trade off across persons and organizations. Okay? If you are in an environment where the small talk is going to be highly productive for you, well that's great, but that depends very much on the environment you're in the stage of your career and so on. I cannot but think of, we just had a two day conference honoring Thomas Sowell. Okay. The guy, the man has published 45 books, 40 books between the age of 40 and 95. Much of that time he is writing eight newspaper columns a month. The guy is famous for never being seen at the office. Okay. So he's an extreme example, but I just wanna point out there is heterogeneity, you know, and even some people find the small talk stressful and annoying. So again, I think we have to understand there's just tremendous heterogeneity or across organizations and people and you know, so there's danger in just focusing on conditional means even though that's the obvious place to start. Other, other common is this, I I buy the basic idea that feelings of psychological safety contribute to positive communications in, in interaction. But there, but that is also something that managers need to think about cultivating In what is now a very different work environment that they grew than, than the one they grew up in, especially the senior managers. And it seems ev some senior managers just can't really conceive of how to replicate in a hybrid or mostly remote environment. Many of the good things that they associated with from traditional working arrangements. So the adaptation I like what I like about your stuff is you're, you're looking at the social adaptation process to these technological innovations and as you correctly point out, that's likely to take place over time, but the organizational adaptation and the managerial app adaptation may be even slower working.

- Yes. Yeah. The second point, I totally agree with the first one. The conditional means were used as so, so we did do a heterogene 80 analysis, so they were person specific slopes and we looked at the distribution, I should have had a graph of that. But essentially it's a, the, the the distributions mean is positive and, and yes the tails are a bit long. So there are folks for whom there's just either no effect or even a negative quote unquote association between having interactions and reporting lower psychological safety. But, but for the vast majority of the people in our sample, it was a positive association, but, but great points. Thank you so much. Yeah,

- Very interesting. So I was, I have two comments and maybe questions. The first one is can you do more on the mechanisms to what extent it is peer to peer interaction or manager to peer interaction? Because that knowing more about the mechanisms would really tell us as kind of coherent story. And then I think the second point I, I thought that things like psychological safety, trust culture, they are very slow moving variables. So it'll take a few maybe months, maybe years to basically observe. So related to this, to what extent you are worried about experimenting, demand effect or desirability bias because you see some fairly strong effects there. No,

- Yeah, we have, so for both, both the points that you raised, we have data to, to address the, there were no meaningful differences in psychological safety after interactions with different types of partners stratified on seniority. So people reported roughly the same level of psych safety with peers as compared when they had interactions with peers as competitive when they interacted with their managers. Number two, the entirety of one of the pilot studies, I shouldn't really call it pilot 'cause it was like a big like 400 person one week, like same design study that basically sought to establish, like sought to examine if there was variance in psychological safety over the short term. And there were a whole host of statistical tests and discrimin validity tests, construct validity tests that were performed and the only conclusion that we arrived, like we could, we could confidently say that this construct was varying meaningfully over short periods of time. Its normal logical network with other associated constructs was also varying meaningfully over that time period. So that was one of our propelling arguments in the paper is that we should reconceptualize this construct as more labile over the short term. But yeah, happy to chat more about it. Thank you.

- Favorite topic, return to the office.

- Thanks for that introduction. So this is joint work with Natalia Emanuel, who's also here in Mandy Palace, who's at Harvard. Okay. So this shouldn't be a surprise to anyone here. Americans are now more physically distant from their coworkers than they've ever been before, but we're also much more digitally connected and especially among remote workers. People are just spending a ton of time on Zoom meetings and other digital tools. And so our core question is in this increasingly digital world, does proximity to coworkers still matter, particularly when we think about training for tomorrow as well as productivity. Today we're gonna test this among software engineers in a Fortune 500 firm, which we're gonna think of as sort of the best case scenario for digital interaction. This setting's also gonna be nice for us because it gives us just a lot of data. We see information about program quantity and quality, and we also see the feedback that engineers receive on their code during code reviews. And so this is gonna give us a direct measure of a form of on the job training. This particular firm is also gonna have some nice variation in engineer's proximity to one another when the buildings were open. So because of limited desk availability, some engineering teams ended up split across the two buildings on the firm's main campus. There were about a 10 minute walk apart, whereas other teams could all be co-located in one building and even this short distance changed some team dynamics because even short meetings would then sometimes move online. So in some senses, those distributed teams were already acting a bit more like remote teams even when the offices were open. We're gonna look at how that difference in proximity manifests in terms of these outcomes of interest. When the offices are open, we're gonna see that basically when the offices close, all those differences are gonna disappear. And so our preferred estimates are gonna focus on that difference in differences strategy. We're also gonna be fortunate to be able to see this natural experiment kind of run and reverse and we're gonna be able to see what happens when the offices reopen. And we see a lot of differences reemerge at that point when there are again, differences in proximity across these different types of teams. So for those of you who have seen this paper before at this conference, I just wanna flag what's new to this draft. So we now have data on program quality, which we didn't have before, and we also have this information about what happens when the offices reopen. Okay, so before I dive in, I just wanna give you a little bit of context about how software engineers do their jobs. I spent one summer as a software engineer, which probably gives me just enough information to be a little dangerous here, but I'll try to be true to what the software engineers at the firm have told me about how they do their job. We think that this firm is relatively emblematic of online retailers in terms of the tasks that the engineers are doing. They're working on the website, they're working on the backend databases and internal tools for the company, and we're gonna see about 3000 total engineers over the course of our period. This gives you some sense of the software development process. So because the engineers of the firm are doing really important things for the firm's bottom line, they really don't want mistakes. So every time you work on a new program, you have to branch it off from the master code base, then you work on your draft, you still have to send it off to someone else who's gonna review it, then they're gonna give you a set of feedback, you'll have a back and forth. Then if you get the green light, then it gets merged back into the master code base and can ultimately affect production level code. We've tried to denote with the two different dogs that the one reviewing the code is often more senior than the one who wrote the program. We can do that more technically by looking at the histograms. So here we look at the histogram of the tenure at the firm among the program writers in black versus the commenters in green. And so you systematically see that the people doing these code reviews have a lot more experience in the firm. They also tend to be a little older, and so you might think they have more information about how the firm works and then also just more general information about how to program effectively. So a lot of this code review process is not oriented just at making sure there's no bugs in this particular code. We've been told by people at the firm that a big purpose of this interaction is to also make sure that people get these learning opportunities to write better code going forward. Okay, so with that background, I'm gonna dive into our first result. So here on the x axis is the time we've highlighted when the offices are closed. The y axis is the average number of comments that people are receiving per program as part of those code reviews. So what we see here is that engineers on these co-located teams who are all in the same building, they receive about 20% more feedback on their code than engineers on these distributed teams. While the offices are open here, we have controlled for things that might affect the amount of feedback you receive, like your age or tenure and what types of programs you tend to work on. If we just plotted the raw means though, it would look quite similar, but you might still be worried that there's just some unobservable between these teams that is different. And this doesn't reflect the causal effect of proximity. So that's why the office closures are helpful. If it really reflects proximity, we'd expect these differences to go away. If it was something else, we'd expect them to just persist as is. And so here's what we see when the offices close and we see that just feedback in general takes quite a bit of a tumble and it particularly declines for these engineers who had been co-located with their teammates and so are seeing a bigger loss in proximity. So our preferred estimate of this difference in differences suggests that losing proximity to your teammates translates into about 1.5 fewer comments on each program, which is substantial relative to the dependent mean it represents about an 18% loss in feedback, you might have imagined that this pattern could have gone in the opposite direction. You might think that if you're with your team face to face, you should actually have less of a need to interact online because you can interact in person. And so in many ways this is probably an underestimate of the total effect of proximity on feedback to the extent to which co-located teams are also interacting a bit more in person. We can look at the specifics of the comments exchanged to have a little bit more sense of what mechanisms are driving these somewhat surprising findings. So particularly what we see is that when engineers are co-located with their team, they receive feedback from more people. Oops, an error occurred. Oh, so they're receiving feedback from more people, a broader network and that echoes of findings in other settings. But then what is maybe more unique to our setting is we can also see what happens in this interaction about the code. And we see that engineers co-located with their teammates ask more follow up questions. So they clarify, oh, you gave me this feedback, why did you give me this feedback? And we think that may be helpful in the learning process. We also had external engineers rate how helpful comments were and then use a supervised machine learning algorithm to sort of label all our comments as helpful or not and along a variety of other dimensions. And we see from that analysis the proximity seems to be particularly important for the comments that people tend to view as most helpful and as doing things like explaining the underlying reasoning behind the changes. I won't go through much robustness in such a short talk, but I just wanna highlight one key set of results that we see that this pattern is really driven by feedback people are receiving from teammates and absent from feedback people are receiving from outside their team. And that's less consistent with a story that you might have where these patterns are driven just by one building team engineers sort of needing or wanting more feedback since you'd expect that to also show up in feedback they're receiving from outside their team. Okay, so hopefully I've convinced you that proximity is helpful in terms of generating more feedback. Now we can think about whether this really is on the job training. If it is, we'd expect it to if most matter for less tenured engineers who are building firm specific human capital and also younger engineers who are building just more general human capital about how to code. And so that's exactly what we find. So this first graph just splits by tenure at the firm where we've split by the average tenure at the time the office is closed. And so the left panel focuses just on these less tenured engineers and we see that the effects for them are much bigger. We see a very similar pattern when we split by engineer age. We have much larger effects for engineer age, engineer tenure and engineer age independently, predict how much feedback people receive and also mediate the impacts of proximity. And so one thing that I think is particularly worth pointing out here is that younger engineers really only receive more feedback than older engineers if they're proximate to their teammates. So if people are distant from one another, older and younger engineers receive exactly the same amount of feedback. Even though you might imagine younger engineers have a lot more to learn from their coworkers and in results I won't have time to talk about today, we also generally find that at any given age or tenure, the effects tend to be larger for women and we find suggestive evidence. That's because women seem to be particularly reluctant to ask for feedback when they have to do so remotely. Okay, so now I can dive into other outcomes. So you might think that with all this feedback we'd hope that people end up writing higher quality code if it's really building their human capital. And so unfortunately the the company doesn't have records on code quality that cover the period around the office closures, but they do have these measures of code quality that go back to December, 2020. And so we're gonna use this in two different ways. The first way is we're gonna think about long run differences in code quality across engineers that spent time on one versus multi-building teams before the office is closed. We're also gonna look at what happens to code quality around the office reopenings. Okay? So we're gonna measure code quality in two pretty standard ways among software engineers. The first is we're gonna think about the share of a of engineers new files that end up getting deleted in six months. This is often emblematic that your initial program was such a tangled mess of spaghetti code. Someone came back to it, they tried to edit it and they just gave up and decided to just start afresh. I'm sure we can all have experienced that with our own code. We just went back and we're like, what was I doing here? I'm just gonna try to start over. And so what we're seeing here is an engineers on one building teams about two percentage points less likely to add files that then pretty quickly get deleted than engineers who had been on these multi-building teams. And so got less feedback when the offices were open. One thing that we find particularly helpful in this analysis is that because these are measured quite a bit after the office closures, there's been a lot of team reshuffling in the interim. And so there's lots of teams that have people who are trained on one building teams and also have people trained who are trained on multi-building teams. And we find that even with team fixed effects, we're seeing this differential then on a given team engineers who have spent time on one building teams are writing higher quality code. We see very similar me outcomes when we're thinking about a more extreme measure of a problem in the code, which is introducing a bug which we measure as you make a set of changes and those set of changes get just immediately reverted, sort of like a control Z on the program. We just wanna take away our changes because they likely generated some sort of emergency in the code base. And so here we're seeing the engineers who've spent time on these one building teams and so receive more feedback about half as likely to introduce these bugs in the long run. Okay, so the other design we're gonna do is we're gonna think about the office's reopening. So for this we need just a little bit of background about our first stage of how much these reopenings are actually translating into changes in proximity. The first RTO didn't actually do much in terms of people spending more time face-to-face with their teammates. The second RTO that asked engineers to be back three days a week and was also a little bit more rigorously enforced, that led to a bigger increase in proximity to teammates and also a bigger difference in proximity between engineers who are on these co-located teams where everyone was assigned to the same office building versus these engineers who are on distributed teams where even though they may go back into the office, their teammates are gonna be going into a different office and so they won't be face-to-face with them anyway. Okay, so then we can see how this translates into code quality. And so here on the Y axis we're looking at differences in code quality between engineers on co-located versus distributed teams where being in the negative quadrant is good because you're having fewer that you add then get quickly deleted. So what we're seeing is that in the second RTO, we have some evidence that these engineers on co-located teams who are spending more time face to face with their colleagues are actually writing higher quality code. Now our story would be that this effect should really be driven by people who don't have a lot of experience because those are the ones where their feedback they receive is most sensitive to proximity. They also have the sort of the most learned from their coworkers. And so that's exactly what we find. We find that this effect is really being driven by these less experienced engineers where we either measure less experienced by either age or tenure. We see the effects are much more pronounced for them than they are among these experienced engineers. We see quite similar patterns when we look at this alternative more serious measure of programming problems, of introducing a bug where again, we're seeing that once people are co-located, they're less likely to introduce these bugs. And again, that seems to be more pronounced among these people who have less experience, particularly less experience at the company. Okay, so so far it looks like being proximate to your teammates is very helpful, but we can also now look at whether it has an opportunity cost, whether all this mentorship that happens in the office is costly for particularly senior people's time as they give these junior people all this feedback. So again, we're gonna look at that around the office reopenings and here we're seeing in this second RTO that engender this substantial increase in proximity to teammates. We're seeing that engineers on these co-located teams started writing about one fewer program per month off a base of about eight programs per month. So substantial decline in productivity. Our story would say that this should be driven not by the junior people but really by the senior people who are the ones who are giving mentorship. And so that's what we find here for inexperienced people, either by age or seniority of the company, we really don't see any impacts of proximity on their productivity as measured by the number of programs they write. They're writing higher quality programs and just as many of them when they're proximate to their teammates. By contrast, for the experienced people who are the ones giving a lot of feedback, we're seeing that proximity seems to reduce their output and has this opportunity cost. We see really similar patterns when we look around the office closures where again we see evidence that proximity to teammates is costly for senior engineers own output. Okay, so now in my last few minutes I'll just talk about more of the downstream implications of this, about the firm's decision making about whom to hire. So you might think that if proximity is really helpful for building people's human capital, that firms might just want to shy away from hiring people who need to build their human capital when proximity is difficult to facilitate. So our results are consistent with that theory. So this we're just focusing on the firm itself and who they decide to hire. This is a histogram of the age of new hires over these different periods. So what we can see in the late blue is before the closures, the firm hired a lot of people who were quite young and the typical new hire was in their twenties. During the closures they shifted towards hiring much older engineers. Typical engineer was in their thirties. Then they, I think more compellingly to us shifted back towards hiring much younger engineers when the, when they did the RTOs and again could facilitate proximity. You might be worried that this is all driven by macroeconomic trends. One thing that we can do to provide some reassurance on that is look at how these patterns differ across the different locations of the firm. We find that this pattern's really driven by people hired into the main campus. And by contrast, for those hired into satellite campuses who will always be distant from some of their coworkers, the firm just always hires older people. Okay, so this is just this firm Now we wanna try to see whether this result generalizes and whether it seems like a lot of firms are shifting towards trying to hire older people when proximity is difficult to facilitate. So to do that, we look in the current population survey and we focus on college graduates who are in promotable versus non-REM promotable jobs as classified by the dingle and nyman classification, which we've heard a lot about today. So here this is focused on people in promotable jobs and the Y axis is their unemployment rate. The x axis is their age, the dash line is pre pandemic 2017 to 2019. The solid line is 22 to 2024 sort of post pandemic. And what we can see here is that for young people in promotable jobs, we see this substantial uptick in unemployment. Whereas for older people we actually see some declines in unemployment for those in promotable jobs. So we can really see this pivoting in who is bearing the burden of unemployment in these types of positions. And so where the next graph is gonna focus on that age gap in unemployment, how much higher is unemployment for younger people than older people in promotable jobs? And we see that that age gap blipped up in 2020 and then stayed elevated ever since then. So younger people have consistently found it harder to find jobs when they have a history of these removable occupations. We don't see a a similar pattern for non-removable jobs. For them it looks like 2020 itself was tough for young people regardless of your occupation, but then it really returned to baseline, whereas that has not been true in these Mable positions. So the next analysis, we'll look at the triple difference comparing the change in the age gap in unemployment and re promotable to non-REM promotable work. And so when we look at that triple difference, we see that the unemployment rate of young people in promotable jobs rose relative to what you'd expect from other young people as well as others in promotable positions by about 0.6 percentage points. We find that's driven by involuntary and persistent for forms of unemployment, largely absent for people who have left their jobs or are on temporary layoff. And maybe more importantly we find that that result is robust to allowing for generative AI exposure across occupations to also matter using gregor's measure and allowing that to differentially affect younger and older people. And so our back of the envelope calculation suggests that this could help meaningfully explain a lot of the struggles of young college grads that we've seen in recent years in terms of their rising on employment rates. Since so many young college grads are in these promotable jobs, our back in the envelope calculations suggests that this phenomenon can explain almost two thirds of the total rise in their unemployment rates. So I see I'm out of time so I'll just leave this up here and move to questions, but I'm excited to get your feedback. Yeah, Chris,

- Very interesting. I I don't think I had seen this version of the paper with the return to office. So nice extension. One thing that is could be really interesting is to compare sort of the support costs of more senior people in mentoring juniors versus whether it would be in the interest of the firm to kind of cut out that support and maybe substitute with generative ai, which is gonna be the next wave of this conference I imagine to say, you know, how much of the wage cost of a senior engineer if you have wage data goes into the time use of supporting someone who is junior. And then if you sort of think about this as a model of hierarchies where that senior person is like a manager or a mentor, you can think about what the skill gap between the juniors and seniors would be in terms of how much time a senior is do supporting a junior versus doing their own tasks. And then think about what it would take for the firm to be on the same isic want as the technology changes as you move across the wage schedule.

- Yeah, I think that would be super interesting. I mean one descriptive thing is that it is the case that just as people spend more time at the firm, the amount of comments that they're giving other people just rises a ton. And so their own output is actually pretty flat even though you might imagine they're getting more effective, they're just reallocating towards like giving other people feedback. So yeah, I think thinking about how all these dynamics are gonna be shifted with generating that AI is a really interesting one. Yeah Steve,

- Very interesting work there. There's a contrast between your firm's specific results and the macro results the firm in for your particular firm. The firm apparently concluded after the pandemic that it was better better to bring people back to the main office at least okay that that was the more cost-effective solution. But in the aggregate data you don't see that as indicated by your unemployment rate data, which suggests that the ag aggregate of firms is making a different decision more in line with let's, let's just cut out these younger high maintenance workers and focus on the more experienced people possibly because we're substituting towards AI or outsourcing abroad at the same time. I don't know. But there is a, there is a contrast between your firm specific and your macro results.

- Yeah. And that point is very well taken. I mean something that I didn't have a chance to share is that like the people who are trained on one building teams are slightly more likely to quit for higher paying jobs. And so I think that that's some evidence like a lot of the at least general human capital part is not being well trapped within the firm. And so it might be rational for the firm to be like, well our senior people will be more effective if they just don't do this type of task and and remote work seems to be one way that it reduces maybe social pressure to engage in this type of activity. So that's a great point. Yeah.

- Yeah, it was great. It was super interesting to see the update. One question I to ask was in the earlier version you had a result I think that it was particularly female managers that got asked to give help most do you see in the return to the office, are they the group that loses the most and I if so, is it like there are more approach? Yeah, the question as to why,

- Yeah, so we do see that effect. I mean a tough thing for us in this setting is I think the results on on women are so interesting but we only have 20% of our population who's female. So always our standard errors get quite large. But yes, we do see that same pattern around the return to office that the effects on negative effects on output do look a little bit more pronounced for senior women than senior men. And I think that one likely reason is that they're just more a little bit more approachable. We don't, we could potentially try to do more in our data to suss out whether that's like the, the exact right mechanism

- And and related question is like this is does the, does the firm take into account the multitasking? I mean I think we talked about principles. So you know the old, I think it's Armstrong Milgram that if you have two activities you want someone to do like write code and mentor juniors and you only reward one write code, of course they under provide on mentoring. So do you know in the performance review, 'cause in the data, if you look at managers with the biggest drop in performance, are they also the ones that are, are they more or less likely to get promoted? Because I can see it going either way a superstar manager builds their team and therefore has a lot of performance drop 'cause that, I mean there's genders one dimension but another is how much they've been promoted in the past or they're performance reviews or something.

- Yeah, I think we could definitely do more than that. Something we got in the additional data collection is we also have like the text of the performance reviews. That would be great. So we're, yeah, I'm not sure how much we traction we'll be able to do on this it, I'll just have to look back in the data of like exactly the time period of that. But

- You have, you have data on something that comes up all the time and if you talk to professional service firms, so law law firms are classic on this 'cause partners get rewarded for fee income but partners are supposed to do two things, generate fee income and mentor new associates and they tend to only get rewarded for one. So totally under provide on the second and you know it's just a management failure but if you had some data on to what extent,

- Yeah I think we could parse sort of the feedback and see how much a manager when they're evaluating their underlings are talking about are they mentoring other people or are they just focused on like their own output. Yeah I think that that would be really fascinating. Yeah,

- This was so interesting. I was just curious if you were able to ask the firm that you're working for more about their intentions short term versus long term. 'cause I think it's interesting if you have the generative AI but then no firm is willing to train their engineers eventually you run outta supply. And so I don't know if that comes from the performance reviews of what they're valuing now in terms of initial idea like did they before value the training of younger engineers or not? Yeah. Where that that idea comes from.

- Yeah, no that's really interesting. So I think this firm was relatively late on the like generative AI bandwagon. So I was like embedded at the firm in the summer of 2024 and at that point they were just just starting to really like institutionalize it into their practices. So it was hard for us to get any direct data on that. I went back this past summer just to like have a chat with someone and their impression was that like now that there's generative ai, the offices are much quieter 'cause people are just like not talking to each other, they're just like writing into the chat bot they use. And so I think like my strong prior is if we got this sort of data we would see changes. We don't currently have the data so maybe that's a good reason for me to go back. But

- I think I've seen some research showing that mentoring actually you know, helps you think and become better as well. You know your a task that you might do regularly. And so I'm curious whether the folks who are mentoring, even though they might be spending less time, do you see an increase in their code output as a cause of like actually spending the time mentoring other people? Does that make sense?

- Yeah, it does. I mean we don't, in terms of code quantity, it looks like more maybe of just like a trade off that you just don't have as much time to program. I mean in terms of some of the code quality stuff, it looks a little bit more like when you go back to the office, you, the senior people also see some improvements on code quality, at least suggestively, which might be that mechanism of like the mentorship helps them too. But we don't have like direct evidence on that. That's an interesting thought. Thank you.

- Okay, so first of all I want to thank the conference organizer for having the paper on the program and after listening to the previous presentation that make, I don't know whether I should say make my life easier or harder, a lot of results are very much consistent. We study return to office using a different setting. So the title of the paper is the Impact of Return to Office Mandates on Equity Analyst. Equity Analyst. I, I don't know whether you have been to a earnings call, Tesla made earnings call yesterday and a few analysts asked a question. I'm studying those, those people, sales side equity analyst. Why sales side? Because the buy side, they also write report by Internalists, not public available. So you study sales side analysts. Some of those people are very well known on the Wall Street and not every one of them. If you want to talk about the total population of equity analyst, sales side equity analysts, there are about 3000, 4,000 of them in the US and half of them live are work in New York, New York City. Okay. I've seen we will send this, this graph like at least three or four times. I'm not going to repeat that. It it, it's, it's hardly debated topic even I think even today you see news articles on return to office mandates almost on a weekly basis. There's so many of them, the federal government, Washington Post Amazon, many of them, Jamie Dem is probably very well known. This particular regard, we actually manually collect, largely manually collect news articles on return to office mandate discussions. We got a few hundreds of them, a lot of repetition and and those type of things. But we did a very simple text analysis. Those are the topics they actually mentioned. Like when Wall Street Journal talk, talk about a particular return to office mandate, what type of words or topic they actually mention. There are not that many surprising things here. It's summarize basically most of the hypothesis we have been discussing organizational culture, employee engagement, employee autonomous work life balance. And there you also see in-person collaboration just covered the period roughly a few months after the outbreak of COVID. So at the very beginning, I think at least 20 20, 20 21 and small part of 2022, they were also talking about safety protocol because COVID was still going on. But 2023 and latter part of 2022, no one was still talking about safety anymore. Okay, the debate, the positive side and the negative side. I'm going to skip this since I mean I don't have to talk about this for this audience. I have a literature review and I just mentioned during the break with Steve, I think I need to update the literature review because the literature moves so fast and I didn't know I must paper had the RTO and I saw the 2023. Now we got a 2025 version. So it's got more and more related to what I'm doing. So sales analysts, there are high school workers who task a high stakes, very important. I can give you a rough number if analysts a typical analyst issue report upgrade, let's see from hold to overweight to buy and then the market reaction can increase by a percent. Okay? The stock price can increase by a percent that the typical analyst, okay, star analyst, the price impact will be even bigger. Those are very important. People on the Wall Street and and market perceive that they actually have a lot of skills in analyzing the firm. They actually cover, okay? It's creative, the coordination intensive, the very often they work in teams and the time consuming, we actually know when they actually publish their report. Very often it's 2:00 AM 3:00 AM So they work pretty hard and, and we have the analyst characteristics. We know their age and their not very accurate, but we roughly know how old they are. We know their gender, we know where they live. Not exact address but like kind of a which city. We also, we can also measure analyst performance since, so the the the, the most important job they do is to give you earnings per share focus for the next quarter, next two quarter, next year. And then we can compare the actual earnings per share. That's how we are going to measure their focus, their, I mean quality of their, their job. Okay? The methodologies we manually collect is certified return to office Monday done by brokers. They're, I I have a table showing which brokers they are like Goldman, Sach, Morgan Stanley, bank of America and they're all in the sample team. We do a stack difference in difference analysis, compare return to office analysts. Those are, those are treaty analysts and with other analysts who are not treaty, those are controls, but basically they are covering the same firm for the same period fiscal year. And then the issue the report roughly at the same time. Okay, very short time difference at the for for their issuance. Okay? That roughly can think they're doing exactly the same thing. One is basically working office, the others working from home, okay? The auto conveyor movie examines not just a focus error, the greens per share, focus error, but also focus timeliness. Majority, I think 50, 60% of the need to give you update immediately after revenue's call. I, I'm pretty sure today the analyst carbon on Tesla, they're super busy because they're, everyone is writing their report. Some of them already issued their report yesterday. Okay? It's so sensitive. You want to give that report as fast as possible because your investors are actually waiting for that. Okay? It's time consuming and timeliness is important. If they delay it for 30 days, then everyone else has issued their report. So what's additional value for you to give a late report? And we also look at the number of firms they cover and also their updating frequency. And typically they do that once or twice every quarter. Okay? And we also look at analyst turn over. So you can basically think this is very similar structure to the paper you just presented. Okay, I'm going to give you a a review of the results and in terms of focus, accuracy, focus, focus, error, return to office, decrease that focus error by roughly 14.7% of the pre RTO median. Okay? The pretty big impact and the effects largely concentrated on younger in experienced analyst. Okay? Analyst is older than 35 years old, you don't see any impact, okay? And female analysts the same as the earlier paper, the standard average are large but the direction goes to predicted direction. Female, they're productivity, their focus accuracy actually increased more after they returned to office. I think they probably got more distraction when they were from home. And another funny is analysts living in democratic states, the results are stronger. I, I'm from Florida, I know even during COVID we were required to come back to school to teach. So maybe analysts also kind of face the same pressure if they live in a a Republican state. Even before they were required to return to office, they were already voluntarily working office. So the marginal impact would be smaller for those analysts. And another thing is that the results are strong on the first forecast after each earnings call. And I said the first one is really 10 pressure you need to update within the next two days. The, if you can update yesterday even info Tesla, that'll be even better. So there's a 10 pressure I guess when you have 10 pressure working from office actually matters more than when you have a lot of time to actually decide whether to update or not. But we do have one test that's more direct directly linked to the mechanism mentor or coordination. So a equity analyst typically cover an industry is rare to see one analyst covering more than wine industry. Okay? But you may have colleague who work in related industries like think about the input off the table. You may have a colleague who's working in the upstream industry or the downstream industry so you can get information from them, right? And you can also think of there may be analysts work in different broker or different industries. They don't really have any other colleague working in related industries. So if you have colleague working in related industries, you can talk to them. So what's going on there so you can get the information, okay? And they in fact a stronger the return to office impact us stronger. If you have colleague you can actually potentially talk to, okay? You can learn from them then it's better to go to office and then talk to them. Okay? Another funding is, this is related to career concern and they cover multiple firms, sometimes 30 firms. So which ones should they actually prioritize when they work from home? They don't have enough pen, they got more distractions. And the prediction is very simple. If I'm covering Apple, okay, I'm also covering a stock bunch of smaller than Apple. I'm having 30 stocks of them. Some are very big, some are not that big. Which one should spend, I make sure that my investors happy Apple, the large ones. Regardless where I work, I make sure I'm gonna spend 10 hours on Apple. Well the smaller ones I can pretend I actually continue to work on because it doesn't really matter, it does not bring, I don't have many clients. It does not necessarily bring in that much revenue for my firm. Okay? The impact of returns office are stronger for small firms. They cover, okay because that interpretation that they kind of pre prioritized the larger firm when they are time constraint. Another finding is a return to office increase focus timeliness, not a gigantic amount from roughly 57 to 58.5%. Okay? This timeliness means that the update within roughly 24 hours after earnings call, okay, to become more timely. That's also an indicator of quality. Also indicator of quality, not as important as the accuracy but it's very important too. And then and turn over RDO increase annual, annual to over from 15.2 to 17.6% and we don't really have a very large sample. So the standard airway is pro is also pretty big. And an interesting observation for the turnover result is that's that funding concentrate the first two years after COVID 2023, largely the effect are gone. I think most firms already had return to office Monday. Even if you are not happy, then you really don't have any alternative, right? So you want to quit and to another firm that allow you to work from home. But there are not many firms like that anymore. So, so a partial equivalence type of thinking, general equilibrium thinking will give you a probably different policy implication. Okay? I've covered ma majority of the result. I'm talking about the, the data. There's a data called ABE that basically collect, it's actually analysts submit their focus to that, to that firm. And so they collect all the data and we got stock pricing information from CRISP and also thoro. The institutional holding data from from Thomson and the RTO news are, and many analysts ity, we collect them manually. Ibase got into trouble. There was a very well known paper some time ago seemed that they actually rewrite the history. They actually changed their, their focus exposed or something like that. And, and that results, it's not that easy for you to know which broker this nalytics work for. But since they all show up on earnings calls, it is a lot of matching and, and and link in search to get their characteristics. Okay, here's a list of a return to office mandate events we have in the paper. There are many more. If you look at, if you read all the news reports, those are the one that we use in the final sample. There are lot of others that we cannot use because they initiated return to office mandate and two weeks later they say, oh no, we are not going to do that. Then we cannot use that. Partially because in early period COVID become more important and then decided not to do that or, or an employee may actually the back there is a backslash and type of things. Or two months later after they said okay we do RD O2, two months later they say okay, we are going to increase the number of days you need to work in office. Then the post period is really, really limited and there are many other things going on. We make sure that in the six months period around the implementation of return to office mandate, there's no confounding events. Okay? In the end we got roughly 35 return to office events. Most of them don't have you how many days they need you need to spend in office. But there are a few of them said it is three days. And I, about the end of sample there are, I think the five or six of them said it, it's fab days in office. But we get a sense that the fab days in office requirements is never strictly enforced because several of them said fab day in office 2022 and they said the same thing 2023. So it's not strictly enforced. Okay? We actually split the sample if they see, okay, fab this in office. But the results are pretty much the same relative to overall sample. Okay, I got summary statistic, we got 1500 treaty analyst and the controls a lot more. That's roughly 22,000. There's those are not unique analysts in any if they show up once the condom ones, right? So there's a lot of repetition and they made a lot of forecast and roughly as I said, 1.5 forecast every quarter for every firm they cover. So on average they cover probably 20, 30 firms. Those are be the be the people anecdotal, I know a lot of them start to use chat GPT or a large longer model to write their reports. Maybe they'll get more and more efficient nowadays. Okay? Those are the baseline regressions. I just want to tell you that we control this standard definitive we control for firm 55 firm fi, the interactive fix effect. I can control pretty much everything you can I can think of and we are comparing the treaty analyst, control analyst. They're doing exactly the same thing almost at the same time. Almost at the same time. So that's the mean estimate focus accuracy, that's about 0.0 0.1. This is scaled by the price level and relative to the mean I think as I said is roughly 10 to 15% of the pre RT mean okay? The dynamic effect, if you look at that, you see the impact started immediately after the implementation. There's no pre trend, okay? You may wonder that why a firm decided to initiate RTO mandate. There's an endogeneity concern but we don't think that will have a significant impact on the result. Largely because of the timing of the result firm like Goldman announced RTO policy, typically a few months before they actually implement, even if they have any information, they are able to predict a change of their analyst performance in let's say three months. Okay? You, you need a story that's convoluted right? And mg I can predict my analyst will start to suddenly do better in three months. That's why I decided to put up the RTO now, right? And which we don't think that's plausible, okay? The impact is pretty, pretty large. It's unlikely to be some kind of reverse ality type of story. Okay? And this is, this is estimated by even you see that we got 30 fab, most of them we get consistent evidence. That's pretty strong evidence. We kind of think this is, we were also surprised by how robust the results are. Okay? And we did a, a placebo basically you can shift the re the actual return to office mandate date to a pre COVID year. Okay? And every of them is shipped because most of the broker have been there for a long time and can redo all the DID using those placebo days and got all the placebo coefficient. The, the P value is roughly 2%. Some of those, those, there are few data points that with a, with a pretty big coefficient, those all show up pre 1990. Okay? Where we really don't have that many of the visions. Very small sample. That's why the estimated error is pretty large. If we focus on after 1990, make the P value, that's exactly zero. Okay? Robustness, we get a lot of robustness. The controls have more analysts from small brokers because they are small brokers. Wall Street Journal, New York time may not be interested in covering them. They're m likely from small brokers. So we remove them from the control. The result, pretty much the same. An interesting finding is that the 2020 we'll have full year period, 2020 to 2023, it show up in the first two years. Also the next two years. So it's not a purely a COVID type of shock. 20 22, 20 23 I think safety concern were largely gone. Okay? This is a result of 10 years with to see, see that they become more timely and focus quantity. There is no change. They continue to cover the same number of firms. And updating frequency is also similar. And that's, that's why when we talk about productivity is also important. Think of the quality dimension because they can pretend there were by issuing a report it can also check be write a report without updating anything that's really tangible. That's not very useful for investors. And those are result of first forecast versus the other. The first focus results are stronger. Those are the experience result. You'll see a monotonic funding younger y experienced analysts, they're, they benefit more from return to office age and gender and the location. This is the result I mentioned about collaboration between c upstream and downstream. The pu pu table, if you have other colleague to talk to, then naturally you'll benefit more from return to office mandates. Okay? This effort allocation, so the, when they are working at home, they prioritize the larger firm in their portfolio. Okay, in their portfolio. This is relatively, I I can cover magnification seven, but the smallest magnification seven will still be considered less important to me even though they're still pretty big firms. This is consistent with the career concern and turn over result. That said there are about two percentage point increase in turnover if you require them to work in office. But the impact largely disappear in 2022. I guess it's by that time there are not many other different alternatives available anymore. Okay, let me conclude. A return to office mandates increase analyst productivity is quality but not quantity. This is largely expected if you are familiar with this particular setting, they're required to continue to cover those firms and update whenever there's importance in for them to update. But they can kind of lower their quality to some extent if they couldn't work at heart or got more distractions or couldn't get the mentoring from senior colleagues. The finding suggests that work from home associated with distraction or less effective collaboration or if you are younger, you receive less mentoring. Okay. You can think of a junior assistant professor who got their assistant professor job 2021. Everyone complains that no one shows up in office and they couldn't talk to anyone. And, and another funny that they actually strategically, strategically allocate their, their attention to some extent. Yeah, I, I look forward to questions.

- Yes, really great. I had one question I think you answered but just to check where often when firms make RTOs, they also make particularly recently kind of comments like if you don't wanna come in, we're gonna kick you out. So it's often associated with more aggressive exit policies. So you can in your study, separate out this the announcement from the actual date and just have both of them in there as treatment dates. Like when they announce it. 'cause when David Solomon announces it, he probably at the same time often announces we're gonna get tough and if you don't like it, it's like as the labor market soft. So can you just separately put in the date of the announcement from the date the actual RTO and check it's the RTO that's mattering. 'cause they, they tend to be temporarily correlated. So you said that your pretre suggest that it's not the announcement, but quite a few of them there's a relatively narrow gap and some of them is a wider gap.

- A typical case is about two months,

- But it'd be useful. So form just econometrically, I would separate out the announcement date.

- Agree. And you can just

- Put them both in.

- Right, right. I agree. That's a good point.

- Yeah. And they may both have effects. There's no reason the announcement

- Doesn't, I I think they both have impact because I just show you that the, even by even almost everyone, you get a, a negative coefficient

- Because the announcement could be I'm just stressed, I'm gonna get fired 'cause the firm's got tougher.

- Oh yeah, yeah, yeah.

- So it is not wrong, but you, you wanted to see a SA main, you know, the RTO on it side.

- Yeah, that's a good point. Yes.

- Thanks. Two questions. One, I had a little bit of trouble understanding the tables because there's a treatment row and a post row, but then you had analyst fixed effects and I would've thought that you couldn't estimate treatment status separately from an analyst fixed effect. Okay. Just something to think about. Oh, okay. Yeah, yeah, yeah. I guess the second more substantive point is that my colleague Boris Rosberg has done a bunch of work on the portability of human capital and analyst teams. And I was sort of wondering what the like team production component here might mean in terms of return to office, where a senior analyst might be reliant on a junior analyst's support or something like that. And was wondering if you might be able to capture something about analysts who tend to work alone or autonomously versus those who are more reliant on junior people's

- Support. We have a funnel in the table talk about that. And unfortunately we don't really have a good data that there was earlier paper on team analyst versus you can actually work just by yourself. The empirical measure is on the report. Whether you get one person signing the report, you get a multiple person signing in the report in the data is the lead analyst. They don't collect other analysts. But, but my understanding after checking the data more carefully is that this is kind of a style or internal role or particular broker. As a Goldman, you always have people, multiple people as long as you work you sign, but some others you get the team, but only the lead analyst sign. So you get 100% for this broker and almost 0% for other brokers so called team fraction. So empirically we don't really, we couldn't measure that. But what we know is a majority of them work in teams. Yeah. Yes.

- If you have the, I guess you have the text of the analyst reports, right? So I guess you could probably, you could probably collect those reports and then construct measures of the types of soft information cited in the reports, I guess. 'cause hard information could be from like a working harder, but then soft information you would expect to be harder to

- Collect. We, we are in the process of collecting analyst reports. Oh, sorry. And the data collection is not that easy. This is from Refinitiv, you know, every report they charge a hundred dollars and I have almost a half a million of them. Maybe Stanford has a deeper pocket than University of Florida and they can cover that. Okay. I got 100 reports for the whole university for free, but just 100. Okay. I know that's $15,000 already. So, and they charge by pages by the way. Yes.

- Thank you. Two comments. One is there's a neat paper by Laura Vel camp and two co-authors in RFS about the value of financial information. You might be able to actually quantify the value of this reduction in, in precision or the rise of precision despite kind of connecting it with asset prices. So there's some back of an envelope sort of thing that you could probably do there for end of the paper. The, the second question is on this kind of like state dependence is precision more, I mean, precision makes more, is more valuable during times of ambiguity and uncertainty. And I guess the que the question is, do you have enough variation in your time series with RTO that you could separately test for like periods of higher inflation, lower inflation, or just some sort of state dependence?

- That's a good suggestion. Yeah.

- Do one more. Hey, just had a brief question. So I'm wondering about whether part of this mechanism might be just a greater willingness to share non-public information in person. So if all of your analyst team is in the office, like there are legitimate uses for non-public information when you're doing this, but there's stigma and all of the communication technologies are recorded if you're working from home, et cetera. So, and it sounds like you're, you're going to look at the analyst reports and maybe you can pick up on a little bit of that, but I'm wondering if you could speculate about that role.

- That's always a possibility. But the current regulation is inside their CEOs or CFO not allowed to share private information with any individual analysts anymore. This being this case in the past 20 some years, even when they work from most majority of the return to office, they're hybrids like three days in office. A lot of them actually travel to company visiting headquarters and talk to people. There. It's, it is, it is always a possibility. And they may have a better technology to talk to people and then leave no trees. You can talk on Telegram, maybe that's one channel. So yeah. Yes. Okay. Thank you. Thank you.

- It is very nice to be here. My name is Francesco Verio, but everybody calls me Verio and I'm a PhD candidate at the European University Institute in Florence. So me, you might notice that I streamline the title a bit. I mean this is obviously due to the fact that I'm time constrained, but I also think that is more focused now it's more clear what I'm, what I want to talk about in this paper. Namely, I want to talk about how working from home might impact intergenerational mobility more broadly inequality. Okay, so let's try to set the stage and the framework that I want to analyze. So I think, I don't think I have to convince anybody here that working from home is going to stay. I think this is very clear from the presentation we have already seen. And I think there is also another important point that has been made in these two days of conference. Meaning that working from home is immediately distributed across workers. In particular, what I'm interested about is that working from home is mostly available for highly educated workers in high income jobs. Now if you think that these people are probably already advantaged in transmitting their skills towards the new offspring, what are the implication of working from home in the long term for inequality and intergenerational mobility? So this is more or less what this paper wants to talk about and obviously it's a model question 'cause it's a little bit ahead of the data. However, I'm going to try to convince you that there are two clear channels through which working from home might have intergenerational effects that can also already be visible in the data. And I'm going to show you some patterns in micro data through the American Chinese survey and the American community survey to inform my model in particular, the first mechanism that I have in mind is an education time reallocation mechanism. So we know from the education of economics literature that college people, educated people on average spend more time educating their children and anticipating the struggle of making it through the academic career. And if you think about the fact that working from people might save some time from commuting, they can spend this time productively educating their children more. So this is the first mechanism I have in mind. But I also have in mind the second mechanism which is related to the donut effect. The fact that with de couple jobs from the office place, which means that now people can relocate towards servers but might also relocate towards places with better schools. So here what I'm plotting are averages is just to give you an idea, these are just descriptive evidence of education time with children. So the education time is something that I code myself from the American Time Youth Survey. And it's basically a variable that embeds both direct academic supports of parents who children like doing homework together and cognitively enriching activities with children. Like for instance, going to the museum, reading together, this sort of stuff. So the only thing I wanted to grasp from these graphs is that first the magnitude. So on average, people spend one to one hour point, one hour educating their children and then there is a gradient that on of working from home, parents spending more time educating their children, which is, I mean clearly, sorry, I don't know the point but here. So it's clearly happening for working from home versus not working from home, but it's stable across different specifications. So in heterogeneous analysis on gender education and child youngest child age. Now obviously this is just descriptive. So what I do, I also run a regression, but I'm going to be clear about it. Like I don't want any causality, police to hate me. So it's not causal effect, but I see a pattern here. And what I do, I regress education, weekly education, time of parents with children on a bunch of controls, plus a working from home dummy. And what I found, it's a a significant effect that is robust to different specification of about a quarter of an hour additional education time that parents spend with children. Now let's go to the second mechanism. I'm trying to be fast, I'm sorry, but my time is limited today. So the second mechanism is something that we want to understand, right? So I'm working from home parents relocating towards neighborhood with better school. So in order to answer this question, I take the micro data from the American community survey post COVID of parents with children between five to 15. And I merge this data with the school quality data from a chat and cohort dataset. Then I estimate this profit model where I regress the probability of a parents moving towards the neighbor with a better school quality on a bunch of controls and then working from home dummy. Again, these are patterns, but what I find is that on average working from home, parents are 3.3 to 3.7 percentage point more likely to relocate towards neighborhood with better schools. Now with this true mechanisms in mind, I build a model to study the effect on inequality and intergenerational mobility. In this model I have a two period overlapping generation spatial model, which is some sort of, is also similar to the one that Michael presented yesterday. So in my model, agents have lived for two periods in the first period, there's just children, they are not taking any active decisions, they passively accumulates human capital. In the second period they become adult and they take a lot of decisions in terms of location, where to live in terms of labor supply, how much labor supply, housing consumption and time allocation, especially education time with children decision. So my in, in my model agents are heterogene in three key dimension. One is the ability to accumulate human capital. The second human capital serve that in my model it's somehow a productivity at work. And the four, the third DI dimension, sorry, is the working from home availability, like this dimension in my model is exogenous. So somebody's luck and is born with the, with the possibility to do working from home and somebody's not, obviously I can talk more how I estimate this probability and how I calibrate my model with this. But now just stay with me with this assumption. Then parents take decision where to leave, meaning that they can decide to leave in a central business district or in a suburb. They are different in terms of rents commuting time and crucially about the spillover effect of neighbor, which I'm going to talk about in the very next moment. In fact, there are two key technologies in my model. One is the working from home technology, which basically parents that are working from home, they can decide intensively how much hour to allocate towards office hours or home office hours through a CES. So I treat these two the choice variable as imperfect substitutes. And the second technology, this endogenous human capital combination of the children, which is a linear function of the ability of the children times the cop Douglas between E, which is the education time of the parent of the children and S of Z. Z is the location suburb of OR or, or CBDS is basically what I call the neighbor spillover, which is a proxy for school quality and is the expected human capital of children in that neighborhood. So this basically is more or less the two mechanism I was talking about. One is the education time of parents and the other one is the sorting effect in two different neighborhoods. So let me jump to the results. As I was mentioning in my model, it's not only similar to the one of Michael in the in in the, let's say in the the framework, but also in the calibration strategy. Meaning that I am assuming that in the pre COVID we are in a steady state. I assumed this to be 2013. And then we see a working from home shock and the compute, the post COVID steady states. And I used the pre COVID as a counterfactual so to speak. So what I found are two important effect. One is a direct within generation effect of working from home. Meaning that I find that people that are working from home are 3.8 percentage point more likely to end up in the top decile of the income distribution compared to non working from home people. This is within generation and it's, I mean I can ex I can talk more about it, but I think that the very interesting finding is this intergenerational effect. So across between generation effect, what I do find is that when I control, when I look at children that are so are people that are not working from home, when they had a parent that was previously working from home, they are 1.9% point more likely to end up in the top decile of the income distribution. So somehow this is through these two mechanisms I was talking about education and relocation and I I I'm sorry I'm running out of time. So I'm just want to plot here like this CDF of these different type of people over the income log income distribution and basically this gap that you see, sorry, this gap that you see the check curves are here in the panel A is direct within generation effect. And here the between generation effect then I decompose for all four types. And what I find in a counterfactual is that the main driver is the education time. So when I shut down the education channels, basically every the, the intergenerational channel almost disappear. So I'm jumping to the conclusion because I'm running really out of time. I think the only thing I want to mention is like the direct policy implication of this paper, meaning that maybe we have to ensure that the opportunity doesn't really become a perk of working from home. So thank you very much. I'm happy to take any comment. Yeah, please.

- Really, really interesting. So two great things, one, I'm gonna, Steve may say this, but the, the more educate, it's a tricky thing 'cause it increases inequality but many by improving it for some but not taking away for others. So that the, if you were to do a welfare analysis, it's not, it's not

- Bad, it's not clear.

- The other thought but more, more substantive is just your functional form is Cob Douglas between school and quality and parental time, which makes them both compliments,

- Right? You

- Could have tried like CES where they sub, it's not obvious. Like if the school system's terrible, it may be more valuable to have parents at home cs, I dunno, I I I'll stop there, but it's very

- Interest. No, this, this is absolutely a good point. So the thing is this, in order to defend, obviously depending on the function of formula you get, you might get different results. Obviously the way which I defend this is usually that most of the models that I am referring to in this literature of like intergenerational mobility, they use this specification like z and GR ham for instance, in which they have exactly school quality and education time. I was talking also in another conferences about the probability possibility to use a cs. So I'm thinking about it, but maybe it's an, an extension. 'cause by the way, this is my job market, so I, I need to have something ready in the few in the next month and I cannot like now start to run additional. Yeah, sorry, please. Yeah.

- So I look, I love this topic, but is is Nick anticipated putting the focus on the inequality is, is the wrong thing. First order effect is work. Work from home appears to be pro-human capital formation,

- Right?

- Okay. Right. Just because the rest of the profession is obsessed with inequality doesn't mean you have to be. So that,

- That's the headline. This is a good point.

- Second, second thing I maybe I missed it but, or I misunderstood, but it looks like your specification has the effect of moving to a another neighborhood working through spillover effects from other children,

- Right?

- Rather than just through a peer school quality effect. And I didn't really understand why you did it that way. It would've seemed to me more obvious to just start with some school schools are better than others and that's

- Right.

- Not because of spillovers from kids, but, so I just wanted to get your reaction to that and since you're already treading on controversial ground in this paper, why not go all the way and and tell us what happens to time spent with children by PE when, when one parent's not working,

- Not working,

- Not working. Yeah. Does that, does that, is that a pro-human capital thing? For some, for some,

- I mean to be honest, I haven't look at that. Maybe that's another paper. But that's, that's a good point actually because yeah,

- It may be that work from home is halfway to something that might be even more favorable for promoting human capital. And of course that's, that's not a politically correct thought these days

- You get a job, right?

- Yeah I know but yeah, I want 'em to get, that's why it's another paper you can do that after you get the job.

- No, but that's actually very good point. 'cause what I found in with this education time data is that there is a significant decrease in trend starting from the 2010. So I want to also analyze this because there might be implication. So there is a very famous paper in the education of economics literature, the race, which also was part of my previous title, that basically they saw an increase in these childcare activities from the nineties to the early two thousands. But now I, I, I see that this time is decreasing. I don't know if because parents pass more time on social media or Netflix with as we have seen or other top type of reasons. So I want to have an additional paper on this and this actually might be super interesting thing that I didn't, didn't think about. So thank you. Yeah. Yes, please. Sorry,

- Super interesting topic. I just want to make the model even more complicated. As someone who had done some work on sibling spirit over, we know developmental wise parents are very important when the children are very young, but then siblings and peers become increasingly important during like middle, middle school age. So, you know, model, if you could also model time spent with siblings, that would be very, very interesting.

- This is a very interesting point. So like, obviously I should speak also to the whole literature of Hackman and Kya, right about this differential productivity time with children in a different stage of their development. I haven't touched this in this paper to be honest. I, I try to get like some clear patterns that I can inform the model with. But that's a good point actually I can, I can extend this. Yeah, thank you. I think I am done with time, right? So thank you very much. It was very nice.

- Hi everyone, thanks so much for having me. I'm from Boston University, I'm a PhD student there and this is joint work with Shelby Buckman here at Stanford and Christina tele trio at the US Census Bureau. The usual caveat applies that these are our views and not necessarily that of the US Census Bureau. So to start this talk, I wanna focus on spouses. How does the opportunity to work from home for one spouse affect the outcomes of the other? So I want you all to think about a more traditional household. That's how we're gonna start. So one might think that if the husband has the ability to work from home, this might create more of an opportunity for the husband to help with non-market work. When we see that more traditionally that falls on the wife and one might think that this could in lead to increased labor force participation for the wife. On the other hand, let's say we have a world where the wife works from home. Well if she's already doing the majority or more of the work from home or the the work at home, then she might increase this. And when you might see an increased specialization whereby the wife spends even more time on non-market work in addition to the work from home. But also the husband has this increased market work potentially out of the home, maybe more of this greedy work and specialization that Claudia Golden touches upon. So the literature has touched on gender specific outcomes and a lot of the talks I think we'll see shortly and have seen talk about some spousal outcomes for work from home as well. So I wanna talk about what we focus on. So we have firm level exogenous shocks in our data using US census data, which I'll talk about more. And we use these firm level work from home adoption shocks to an individual to see how their spouse is affected in terms of their labor market outcomes. We can look at labor force participation, job switching, probability relative earnings hours worked. We can also look at the probability of the marriage dissolving as well. We also have some external data that we bring into the census, which I'll highlight more, but all to say we can look at hybrid outcomes. So we can look at how firm specific minimum day in office policies, minimum percent of time in office fully remote options affect these gender differences. And we'll also look at how these effects differ for parents versus non-parents. And we're really curious if work from home will change within gender and within couple gender pay gap. 'cause one might think in the two worlds that I talked about in one world maybe if the husband gains work from home we might see a closing of this gap. But if the wife gains work from home, we might see an expansion of this gap and we think we can think more about that here. So our data primarily uses the longitudinal employer household survey in the US census. So this has 95% of US private firms. We're using data 2010 to 2025. And so we have this firm information and individual information and we also use this to create couples based on their home address and some other criterion. We then merged this data with the annual business survey, which gives us a yes or no answer to whether or not a firm allows any sort of work from home. But we also and that that is annual. So we have that in 2019 through 2025. We also have the flex index, which is an external private data set that we've been generally allowed to use in years 20 23, 20 24, and 2025 with this much more detailed information on work from home policy. And we can also see if that changes across different firms. We also have access to the Americans community survey, which we're going to merge into our census data. And importantly this provides us with information on individual uptake of work from home. So we can use this for an IV structure where we know from the A BS or the flex index whether or not Affirm allows for uptake of work from home. But then we'll actually see does the spouse then take up this work from home? So we can really look at that channel. So our data set that we're going to collapse to, we'll have couples who are heterosexual ages 22 to 64 and stay together at least 2019 to 2022 for whom we have work from home information in at least 2019 and 2022 for at least one member of the couple. We'll also look at a sample that's more restrictive where we know the work from home for both members of the couple. But we look at both of this. So our samples will consist of first this L-E-H-D-A-B-S sample where we'll look at more of this binary shock to work from home in 2019. This firm does not offer work from home in 2022. The firm does offer work from home, then we can use our L-E-H-D-A-B-S flex sample to say what happens if the firm in 2019 didn't offer work from home. And now in 2022 it does offer a minimum policy of one to two days in office. So we can think more on the margin of how a certain level of work from home affects a couple. And then we also have this IV approach that I mentioned looking the using the A BS firm level data and the a CS to get some causal results. So our empirical strategy is an event study. We use this shock in work from home in 2022, but we can also test the shock for other years. And we're looking at again the firm not offering work from home in 2019 to offering it in 2022 and then seeing the effects on your spouse. We're controlling for industry so we're controlling for this idea of your industry's original propensity to work from home and really looking at that firm level switch shocking your spouse. And then you, we have two initial census results that I wanna highlight. These are sign and significance as it takes some time to get these results out of the census. But we find that when husbands have access to work from home, as we thought wives experience a positive effect on their labor force participation and job retention. So they're more likely to stay attached to the labor market, more likely to join the labor market. We have this suggestive evidence, I mean it is significant, but we find that when wives have access to work from home, their husbands are more likely to switch jobs. And what we wanna explore is are they switching jobs again with this greedy work idea so that they can work longer hours, they can work and earn higher wages. We still need to explore that. So we wanted to think about work from home through this model. This is based off of TI 2024, the basic ideas that you have a unitary household model where you have some market work and some non-market work. The main thing I wanna highlight here is our contribution. So if you just think of this part of the equation as market work, this part is non-market work now we assume for now just that the husband is able to work from home to simplify what we're doing. And we see here that the husband can do market work and market work that's work from home, but that might have this cost sigh for various reasons that we've talked about. Now the less likelihood of a promotion or the non wage amenity, but then you also have the benefit that we're thinking about work from home in terms of multitasking. So even if it might lose productivity or have perceived productivity losses, it allows you to multitask and get some non-market work home, which might change the maximization problem for the wife considering how many hours she'll do of market work. So to extend this model in the future, we do wanna maximize over the decision of how much work from home the husband will perform, which we think fits naturally with our connection to the flex index where we can look at the hybrid work requirements and how many hours people are actually required minimum to work from home. But we also know whether or not they actually take up this work from home. And we also wanna look at adding some minimum non-work constraints and see how they affect our model and extend this to a collective bargaining model. Seeing how, and we look at not just unitary household but also a collective household, how this might affect our results in terms of empirical extensions. So we definitely wanna extend this to parents versus non-parents. Currently we can do do that using our a CS dataset, but we are also applying to extend to having all of the parent child links and census. So we'll be able to do this for our whole data set. We also want to look at return to office, see how this compliments other results that have come out today. Because we can look at firms who allow work from home in 20 22, 20 21 and don't in 20 24, 20 25. And we have those flex index specific firm policies that we can look at. Yeah, so in summary we have this rich census data, employer employee matched panel with firm level data on work from home from the annual business survey, but also this private data from the flex index. We have these two main results that that went along with what we expected that wives experience positive labor force effects from their husband working from home. And in contrast, husbands are more likely to change jobs when wives gain workplace flexibility, maybe suggesting these different implications for for different future work by gender. And we're also developing a model to think more about the trade-offs posed by work from home. We think thinking about work from home is the benefit of multitasking versus the cost of promotion et cetera might be interesting to explore further. Thank you. Oh sorry, I have to call on people. Yes. Oh yeah, I should first.

- Okay. Yeah, very interesting. So one question I have is, is the work from home also has this impact on relaxing the dual career issue? Know that you don't have to be both at the same location and because the wife is often the second mover and might be have, you know, have restricted opportunities because of that. Is it something that you can sort of comment on or look at?

- Well excitingly we can definitely look at that that originally my co co-author and I Shelby were thinking about doing that as a, as a first paper. But we wanted to think more about the labor market implications first, more stagnated. But we do have addresses that's partially how we form our couples. So in the future it would be fascinating to look at what happens when couples move. 'cause we completely agree if it restricts the constraints for one member, it might also just increase their labor market outcomes, the them having a better fit at a job through work from home.

- Yeah, super interesting. As I mentioned we thought about doing this but too much work. I'm so glad you did it. And then one question is that, well within the same firm right, different team, they may have different remote work policies. So I think for your IV you could combine occupation information from a CS with a firm level of flexible policies to create a stronger iv.

- That's just such a good suggestion. Yeah, we are all, I didn't mention but we are also trying to use the AC or going to use the a CS for the detailed occupation information. But I hadn't thought of it in that way. So I think that's that's brilliant. Yeah.

- Just to follow up on Michelle's point, it, I think you have enough variation in local density to get at that a little bit. So if the, if the husband starts to work from home in a small community, that may advantage the wife more in terms of expanding her opportunity set than if they're, if they're already in a very dense urban area Yeah. Where she's got a rich opportunity set. But I, I didn't get all the data sets in my head. You're able to follow people longitudinally. Yes. Even if they move location.

- Yes.

- So you can, you can check all, you can look at all that. So

- Yes - The, the thought experiment is husband shifts to work from home, rural area versus urban area. How does that differentially affect what happens to the wife?

- I love that. I was also thinking about geographic variation in terms of social norms. So I think it'd be interesting too how much I think geography will be rich for this. So I should definitely look into it. Thank you. Yeah,

- Super interesting. So you are basically condition conditioning your sample based on couples who survive the pandemic. So

- Yes,

- So, but they're definitely different than than the others. So I wonder because they're already flexible and they're supporting each other and they survive the pandemic. So to what extent, what happens if you run the results in the full sample and to what extent you are worried about this type of limited variable bias that you can't even capture?

- I mean that's a great point. We can definitely, I, I have, I have not but we definitely have the ability to run 'cause we can, we can see at least that these couples change addresses so we have a pretty positive indication if they're, if they're leaving each other. Yeah, I think I would have to think about it more. It might be interesting to look at couples that have been together longer before to try to mitigate some of those concerns. But then of course you might be less likely to capture the childbearing age of couples. So, but I will think about that more because there's of course selection with that. Yeah, that's a great point.

- Yeah this is super interesting. I'd also be curious about sort of the direct effects on the individuals themselves as opposed to only like the cross spillovers of like how does it affect women when their company has an RTO, that sort of thing.

- Definitely that's, yeah that's a great point. And we can do that interestingly enough, like some of those results and we we release some of them are in fact opposite. So we have even more strength to believe that this is directly from the spouse if your individual goes in the opposite direction. So I think it only strengthens our results to, to look more into how those may might be going in the same direction or different directions.

- Okay,

- Thank you very much.

- Good afternoon everyone. My name is Hmong. Thank you for having me. I'm a PhD candidate at at London School of Economics. Today I'm gonna present my work on fathers working from home and household labor division. So in this work I'm studying whether a temporary exposure to working from home can persistently shift household behavior. And I studied this problem through the lens of COVID-19 where millions of fathers are forced to work from home regardless of their prior preferences. So back to the topic, why should we care about household labor division? Let me show you the gender inequality in the us. So this graph plots the decomposition of gender inequality in the US from 1980s to 2010. In 90 eighties men earn 58% more than women and by 2010s this figure dropped below 50%. But when we look at this pink area, the child related gender inequality, it's not only increasing in absolute scale from 28% to 32%, but it's also increasing as a share of the total gender inequality. So in 2010s, child childbirth can account for 70% of gender inequality. And why does that happen? Because of household labor division. So when the first first child is born, the father's well the father's trajectory spent remain roughly unaffected. Mothers step back, they work less, they seek flexibility, they take more leaves. Those are all at wage costs. And the natural response would be, well can increasing fathers household involvement like encourage maternal employment? There are some policies but the effectiveness of these policies are limited because of its voluntary nature. So when these policies are optional, only the parent who is already bearing these care caregiving needs, there's this childcare preferences tend to opt in these policies. And in practice these policies often reinforce gender gaps. And the COVID-19 brought us something different. So with the school shutdowns and the workplace shutdowns, fathers are forced to work from home where their children leave. Let me show you the magnitude of this shock. So this blue line plots the share of the US population affected by any kind of work from home orders. So starting in March, 2020, within four weeks all the states have issued some kind of work from home mandate and most of them are lifted by June, 2021. And this shaded area, I will refer this as COVID period throughout my talk. And then let's see how many worker actually work their full day at home. So pre COVID, this number stabilized around 7%. I'm sure we are all very familiar with this surge during the COVID and this stabilize around 20%. So why does this shift matter for my study? Because this shift is involuntary. This gives me something really rare in the household economics that is the exhaustion variation in working from home. And that allows me to study what happened within household. When fathers are forced to work from home, let's zoom in the household and see what happens within the household. Some descriptive patterns. So here I'm showing you the daily time parents spent with children for the last 20 years from American time youth survey. So the daily time spent with children com consists of two kind of childcare. The first one is primary childcare where the childcare done as the main activity such as feeding, bathing, reading to the children and secondary childcare, which is the childcare done while doing something else like cooking or working. So what we see on the graph is that for mom they spend seven and a half hours per day with two children and will fathers only spend five hours per day. And this two and a half hour gap is very stable, barely moves pre COVID during COVID due to school closure, both parents scrambled to fill in the gap, they both increased their time with children by one hour. But what happens post COVID is interesting where we see with the school reopening and the restrictions lifted, mothers should return back to the pre COVID baseline while furthers maintain this elevated level of childcare. So that brings my question, my paper ask, can this involuntary exposure to working from home shift labor force, labor division within household persistently and specifically I'm asking these two questions, does work from home culturally increase father's involvement in unpaid labor? And if so, do this effect persist? And the second question is, does increase paternal involvement promote maternal employment? In other words, when father are providing more support in the household, can mother be relieved from these household burdens and encouraged, be encouraged to participate more in the labor market? And on the first question, to address indulge 90 concerns arising from further selecting themself into certain occupations based on their like underlying childcare preferences, I'm using a shift share EV and well it's constructed by the state industry level, work from home rate hopefully use this term to capture the policy shock and also the share, which is the pre COVID occupation level feasibility score by dingo and naman. And hopefully this will capture the exposure to these policy shocks. And on the second question, I'm using the C, leveraging the panel structure of the current population survey and run some differences. Differences in event studies and also explore the heterogeneity effect by occupation flexibility. So before I run out of time, let me show you some mean results. So here I'm gonna present you the effect of fathers working from home during COVID on fathers on on fathers time allocation and mother's time allocation. And I'm going to show you the results pre COVID because I use this as a possible effect because we should not expect any results on pre COVID outcomes by the during COVID work, working from home if my AbbVie is valid. And then I'm going to show you the COVID results for contemporaneous effect post COVID for the persistent test and the unit, it's minutes per day. So first one, first results I'm showing here is on the commute time, which is really intuitive. When father work from home during COVID, their commute time dropped by 29 minutes or 94%. And then this working from home enable fathers to spend three and a half hours per day more with their children. This this is mostly secondary childcare but this still matters for mom because it enable mom to reduce their time spent with children and also unpaid labor by roughly two hours. And this also enable mom to increase paid work time. But what happens post COVID on fathers, we see they in, they persist their increase in unpaid labor but transform from secondary childcare to active caregiver. This 76 minutes unpaid labor post COVID is consisting of 43 minutes of primary childcare and 33 minutes of household household work. But what's puzzling here is on mothers, despite fathers sustained increase in unpaid work, we do not see any corresponding reduction in mom's unpaid work and nor their increase in their paid work. So why does this happen? My hypothesis is that mother have to be employed in a flexible occupation to capitalize on father support their occupation, have, have their occupations, have to allow them to utilize the support when the father provide their support. And let me show you the, I run some difference in difference event study. I will show you the difference event study graphs and I will also show the difference in difference estimates in black number here. So when I run the event studies on the whole sample, we see that when father work from home, mother mother's labor force participation does not respond. But when we zoom into the mothers that ever have ever experienced in the market, we see when further work from home, their labor force participation rate increased by four percentage point during the COVID and five percentage point post COVID. When we split the sample by occupation flexibility, we see this effect mostly concentrated in flexible occupation. Mothers, when father work from home during COVID, these mothers increase their labor force participation by seven percentage point persistently and that accounts for a 10% increase while mothers in rigid occupation does not experience any positive effects. So the key takeaway from this paper is that a temporary exposure to working from home can persistently increase paternal involvement. However, for mothers to utilize these support, they have to be employed, they have to have this occupation flexibility. And I'm gonna skip this policy implication 'cause I'm running out of time and feedback comments. Welcome,

- I, I want to give a different interpretation of some of your key results. So here it is, fathers suffer a huge child enjoyment penalty because they work so much and mostly outside the home they get a raw deal. This child enjoyment penalty for fathers diminished persistently because of the COVID shock. And so more child enjoyment for fathers after the pandemic, less commuting time. So fathers get a less bad deal in this respect than before the pandemic. So I put it that way very provocatively as a complete kind of contrast the way you put it. And I guess I, so two points, one, I don't see what in your data analysis really differentiates between those two. But the larger point I want to convey to you is you're dealing with a very interesting topic, but there's like a, how to say it, an ideological overlay to the interpretation

- That - Seems somewhat distinct from the evidence itself. So even, and I, I understand this motherhood penalty, child penalty term is widely used in the literature, but you know, it's got this connotation that somehow spending time with kids is a bad thing, which may or may not be the case depending on your kids and the person and so on. But anyway, do, do you, I hope you get the point and I don't, I don't mean to be, I don't mean to give you a hard time other than just take kind of a more neutral tone in the presentation, I think would serve you well.

- Thank you for your comment.

- Any, there was also question, is there something that favors the interpretation that you kind of offered us versus the one I offered us? The mere the mere fact that the women aren't, aren't gravitating towards more work suggests even when their fa, even when some of the fathers have more flexibility, doesn't suggest they feel like spending time with their children is such a bad deal. So I you understand what I'm saying? What, what is it, what, what's wrong with the characterization I gave or how, why, what what in the data favors your characterization over the one I gave. I,

- I do find that during the COVID when the father work from home, they tend to spend children spend more time with their children while during the daytime. So that means, I suppose that means like the fathers. So your story is that, if I understand it correctly, so your story is that fathers do not enjoy No, they like it, they appreciate the opportunity,

- They appreciate the opportunity to spend time with children. This huge child enjoyment, penalty they had been, suffering was diminished.

- That's

- All I'm saying. So and but you're, you're, you come at it from the idea that spending time with kids is bad and mothers are escaping this.

- Well that, that's that's not entirely my idea. It's not, I

- Know it's not, I know it's not, it is the prevailing idea, I understand that. But I'm pushing back against that and at least asking for evidence that distinguish between those two views.

- Okay.

- Right. So I'll stop here. I'm setting

- All right.

- Yeah, I mean maybe related to this. So I think the penalty indeed refers usually to the earnings penalty. I'm not sure, but, but there are, there are, there's I think a study in in Denmark that looks at the increase of parental leave and that part, part of it is actually has to be taken by the fathers. And so they also have this exogenous variation in how much time the father spends at home. And my recollection is that they actually found again that fathers did actually learn that they enjoy spending time with their children and that had a permanent effect as well. So you may want to tie to link with that literature as well on parental leave policies and changes in that.

- Yes, yes. I, and I do think there are also some part literature on the parental leave showing that these are also, well even though the paternity leave is compensated well, there are still low take up. And also even considering these like parental leave, we still don't see any well significant improvement in like the labor, labor for household labor division. That's how I All right. Thank you so much for comment.

- One other, I mean I, I think the penalty versus, and Jonathan thing is interesting. One way to look at it is, do you know the age of the children? I have to say like spending time with a 13-year-old versus a 3-year-old. There's a, there's a, you know, there's a different sense of pleasure or not pleasure but do you know that if you had the age of the children in it?

- I do have the age of children. Well the, I do find because American time use data is quite small, I couldn't like split that sample but I do that sample split in the current population survey analysis and do not find any difference impact on that.

- Amazing, thank you so much.

- Thanks.

- Welcome. Afternoon everyone. Today I will be talking about remote work and fertility. I have a lot of slides so I'll try my best to finish all the slides. Okay, so in the US history, every time we have economic don typically we'll see a temporary fertility decline. But this decline is only last for about like two to five years and then we'll bounce back. In other words, economic downturn tends to shift the timing of fertility but not changing the total number of children a woman will have in in her lifetime. However, something is different for the 2008 economic recession. As we can see from 2007 or 2008, fertility has been declining but not bouncing back. So many scholars are very worried about this trend. This may signal of a permanent decline lifetime fertility. So currently the US fertility rate is below the replacement rate, meaning that we're not having the enough number of children to replace the people who are passing by. And this posts a very important policy challenges as as exacerbating the population aging issue. You so many people have proposed that flexible work family policies may offer solution to this low fertility issue, especially in OECD countries where the majority of parents actually want people actually want to give birth to children but they don't have the opportunities so resources to achieve their desired fertility. However, US is behind other high, high in combinations in providing institutional work family support as the figure shows this displaced a ranking of the generosity of parental paid leave at the national level. If you look at at the bottom, US is actually the only country that doesn't offer any parental policy at the national level. But you may ask like what about the state policy? So in 2023, only 27% of workers had access to paid family leave and 90% of workers had access to unpaid family leave in the us. And as we all know, the access to family leaves also varies disparately. As a result, many many American women opt to delay fertility until they are established in their careers and many others they end up having fewer children. While contrary to previous findings that economic downturn leads to decreased fertility. We observed a remarkable pandemic fertility rebound in 2021 to 2022. Especially this rebound was largest among younger, highly educated women Hispanics and for first birth. And many scholars have hypothesized that while remote work might be the reason specifically remote work may help women stay employed and mitigate work life conflict issues. And to this audience I don't need to explain remote work is here to stay. And we know workers design more work from home days than the employers are willing to offer. There has been emergent literature on the association between remote work and fertility intentions and the results are all over the place. And in the fertility literature we know fertility intentions do not always translate to completed fertility. So these two are very different. There are very few studies have examined the association between remote work and completed fertility. This one study found that regular remote work access, regardless of usage actively linked to first birth in the UK but is fertility for long. In another study they found that the widening availability of broadband internet boosts higher all births among highly educated women in Germany, possibly through increasing remote and part-time work and better work-life integration. But this second study they didn't directly look at remote work and fertility. So they look at the availability of broadband internet and remote work and and fertility separately. So the two most important gaps in the literature. First we don't have any studies in the US context. And second the previous studies overlooked issues within ingenuity. And depending on the specific endogeneity, the relationship could go either way. Theoretically the relationship between remote work and fertility is really ambiguous. On the one hand, remote work may increase fertility for example, remote work may reduce resource constraints by helping women remain in the employment and maintain work hours and may also reduce the time constraints by cutting commuting and grooming time, et cetera. There's some suggestive evidence showing that a remote work may promote family information. On the other hand, remote work may reduce fertility because there has been a lot of work showing that remote work, anies work family conflict with multitasking and longer work hours especially for child women with young children. And remote work may also diminish promotion opportunities and slow wage growth particularly for young workers. So theoretically is really ambiguous. I would say the biggest challenge for this project is to find suitable data in the US We just don't have many good data for this topic so we will need to find a data set that is big enough for us to observe fertility, complete the fertility. Also ideally for the identification we want to have individual panel data. So the current population survey individual level panel is our best choice. It's not perfect so we are doing a lot of well cook we're cooked based on the materials we have. So we use this panel data from CPS from 2019 to 2021. And this panel is a little bit awkward in the way that it follows households over 16 months in the four eight fold rotation patterns basically interview household for four months and then the household rotates out for eight months and then they come back again for another four months. In other words, each household can have at most eight observations. So we restrict the sample to women age 15 to 49, whoever employed before pandemic, excluding self-employed. And then fertility was captured by observing whether there was a new household member on the 1-year-old in the current month. And we just use a difference in different models in individual student fixed effects model to follow. So our treatment group is defined as well. A respondent's major pre pandemic occupation is tele workable and we use this updated index developed by the A team in the Bureau of Labor statistics. So this industry be updated version of the dingo. I'm surprised not so many teams are using this. So it's a done variable version. So we merge this index with the C-P-S-C-S data. We also do robustly checks using all kinds of in indices including the ones we construct ourselves and also dingo andon indexes. So model specification, this just look at, looks like a typical different diff model. I'm happy to talk more details about this. I would like to highlight two analytical decisions which are very important in fertility research. And I see this mistake appears in many previous studies. First we decided to backdate the treatment and time 60 facts by nine months because they were truly trying to capture the timing of conception because we were interested in the behavior mechanisms behind this fertility. Second, for all analysis we stratified by parity, first birth was as high order birth. It's very important because the considerations behind first birth is completely different for the considerations behind higher order birth. Now let's look at the results for first birth. We, we have seen overall positive effects in the total sample and across all subgroups that we, we look at especially the effect is particularly larger among the age group, 30 to 34 highly educated and married. Married and partnered women and for higher order births we again, we see overall positive effects but the magnitude of effects are much smaller compared to first birth and we don't see any significant subgroup differences. We also did some analysis trying to tap into the couple dynamics using a sub sample the CPS panel. So this results shows the shows the first birth as we can see for both men and women, remote work boost fertility. But if you look at among women, the group of to a group of women actually increase experience. The largest fertility increase is the women with tele workable husband. So the effect size is much bigger compared to women with unemployed husband and women with nont workable husband. We don't see any subgroup differences among husband with all kind of work arrangements for their wifes for higher order birth. Again, we say overall positive facts but we don't see any significant subgroup differences. We also look at a lot of possible mechanisms. So among all this mechanisms we look at, so we also look at found information which is not listed here. The only mechanism we have solid evidence for is employment. So we found remote work reduced risks of unemployment and increased the risks of part-time employment. So we didn't find any evidence for the hypothesis that remote works promotes family information or remote work promotes work family integration. We also don't find any evidence for the argument that people simply have more baby together because they have more time spent together, they have nothing else to do during the pandemic. So to summary of the results. So overall we found positive, consistent positive effects between work and fertility. And the effects are particularly large for first birth and are greater among college educated are married or partner women and the effects are likely due to remote work helping women stay employed especially part-time. So at the end I also want to give you first taste about an ongoing project that we're doing using much better data in Europe. So we actually have administrative data in Finland, Sweden and Denmark where we have information about every people in the country because of HH one B band, my collaborators were not able to travel back to Finland, Sweden to export the results. But luckily the dentist results are exported as we can see. These are results from a synthetic different diff. We do see across all the three countries positive, positive effects between remote work and fertility. Thank you.

- Super interesting. Two small things. One, I wonder if the nine month lag is the right amount because there are delays between intention to have a child and conception and you also have miscarriages obviously. So I don't know whether there's, and that probably varies by age as well. So I don't know whether you, whether there's solid evidence on this and what these lags between intention to have a child and an actual birth look like. But if there is, you might wanna modify the model. I guess the, the main substantive issue, and this is whether what you're picking up, what to what extent what you have is a rebound effect

- Because

- There was a shortfall in in fertility during the pandemic, not surprisingly. And then there's a recovery. So maybe you can speak to that.

- That's a very good question. So for the first one, the nine months backdating is kind of a common practice in the fertility literature, but we also do robust interest using like the eight months or 10 months. The miscarriage information is just not available so it's very hard to deal with that kind of analysis. And then the second thing, that's actually the major limitation of this paper because we don't have long enough of the data to really say what is just shifting fertility timing or it's shifting lifelong fertility. So we need to wait on, wait until these women reach age 50 to track their cohort fertility. To answer this question, yes,

- Have you thought at all about just differences in in childcare costs or who's, who's there to take care of the child? 'cause you can see like the husband and wife being part of this, but then there's other family members. So I dunno if you can look at any geographic differences on the availability of childcare, things like this.

- That's a great question. So the only subgroup analysis we did is whether the state has paid parental leave and we didn't find any significant subgroup differences. But yeah, we could look at more, most heterogeneity analysis. Yes.

- So super interesting. I mean we, we talked earlier, we've been looking at, I think the, the hard thing for us to figure out is direction of causation. 'cause there's Two stories. One is I can work from home and therefore it's easy to look after kids. So fertility goes up. And then the other is we're gonna have a kid, we know we're planning, we're trying to have a kid and maybe don't quite get there, but we're trying to have it. And so in advance I move into a work from home job. And those pro in all honesty, they're probably both happening. So the question is how do, do you have any way to just, that was the big challenge that we're facing that we see the correlation,

- Right?

- It's not entirely clear what share is each of the two directions or some other third thing like as Steve said, one is the rebound is another thing is during that period there's a lot of job loss. And I remember in 2008 there was some claims in some industries there was a baby boom because for women, if I, I was in the UK I heard this, but for women, if you're pregnant you are protected under employment law. And Nina, it's hard to know how big this was. You know, there's various other effects going on as well. But

- That's such a great question. So we didn't look at the switching pattern, right? So we look at the jobs, whether they switch from treatment to control groups right before and after they give birth to the, to a child. I think this like can intend to trade model specification and deal with some part of it. But then we also conduct a bonding analysis trying to bond the proportion of women who actually switched treatment group statuses. And I mean the bonding analysis, basically this concern is not big enough to completely eliminate our results for the first birth at least. And then your second question, sorry, what's your second question? Well,

- The name was the causality, I mean, right? We were trying to look at occupation, but then even occupation switches. So then it would say, well, after what age does occupation not switch? By the time you get to 30, it looks like it's relatively stable. Then you're looking at very thin slither of people, you know, 30 plus.

- Right.

- And they have to be graduates probably. 'cause that's where they're liked. So I, yeah, I think the hardest thing is looking at the causal effect and I

- Right.

- Totally on board with the correlations. It's just

- Right, right. I mean, as we briefly talk about before, the switch from treatment to control, from control to treatment actually is, can be equivalent comparable in terms sizes. It's now, as people from women mostly switch from non tele workable jobs to tele workable jobs before, right, before or after they gave birth to children, which we found surprising, but at least it's in the CPS data. So that's why I think the European project will be more exciting because it has more information for us to tackle this causality issue. We actually know exactly whether this individual person work from home or not. Well, the us the data are really, really limited. And now it came back to me about your question about unemployment. So we did look at the fertility of unemployed people and yeah, so for the unemployment people, they also have a fertility increase actually, but only when the women unemployed, not the men. So when the men unemployed, actually there's a fertility decline and then we kind of compare the effect sizes between the unemployed, unemployed groups with the remote work groups to tackle this hypothesis that while they're having more babies, because they're spending more time together, the magnitude of remote work groups is much bigger than the unemployed groups. Many. It's not only the mechanism of time. Thank you

- So much.

- I don't really need to do much introduction after, after Emma's presentation, but, so let me just say one thing, which is that, so this is obviously new work on remote work and fertility. We're gonna piggyback on a bunch of new data that we've been, well even not that, new data that we've been collecting in sway our US based survey and, and in our global survey of working arrangements. And, and so this is a joint paper with jva, Nick, Caitlin, Steve Mathias, San Pablo, and Caitlin unfortunately was supposed to present, but she's sick. And so you're stuck with me giving a very unpolished presentation. So hopefully I won't do too much of a job, too much of a bad job. So yeah, the motivation for this paper is, yeah, we know fertility has been declining for several decades, more so perhaps in the developed world than in, than than say in Africa. But given the, the recent shift to work from home that we've seen, this really raises the question can potentially by alleviating people's time, budgets change their propensity to want to have kids or not. And, and so that's what we're after in this, in this project. And so what we're gonna piggyback on is a huge data collection effort that, as I mentioned, we've been doing in Sway and in g Sway for several years now. And, and, and so basically we're gonna be able to get a bunch of variables about people's fertility experience ever in their life in the past 10 years or so, as well as their intentions going forward. And so in those data that we can link to work from home outcomes, because both of these surveys are about working arrangements, we're gonna find a positive correlation between own work from home, partner work from home and these fertility outcomes. But so as, as we were just discussing, it's not clear what it really explains the correlation. So, so one story is this purely causal effect whereby suddenly the opportunity to work from home appears. And so people, for example, have lucr time budgets and they can, and, and, and so they might change their choices in response to that. The other story is, is maybe now that we're in a world in which work from home is possible and, and that facilitates to having kids, maybe there's this occupation switching into, into the sorts of jobs that allow you to work from home if you are already planning or intending to have kids. And, and so yeah, that, so what we're gonna do in the, in the second half of paper is, is show you some initial explorations that we've done with US data and specifically the CPS and the a CS for, to try to tease apart kind of what evidence do we see of each of these stories. And so, as I said, the the data sources that we're gonna use, we're gonna look at our global survey of working arrangements, which in which we have about 20,000 responses from 38 countries collected in 2024 and 2025. So if, if you've looked at some of our previous work previously, we were focusing on, on a sample that was, that of people who were closely attached to the labor force. We're going to relax that for this project because fertility decisions are very correlated with, with labor force participation. And, and so, and so we're gonna expand the sample both in sway and g sway and, and, and bring in a bunch of these fertility outcomes. And like I said, for the second half of the paper, we're gonna use the, the American Community Survey to think about, think about the causal effect of work from home on fertility and, and look at occupation switching in the CPS to see if there's any evidence for, for people basically trying to find work from home jobs, work from home mobile jobs if they are considering having kids. And so in sway, and, and I've spent, so, so this is the part of the paper that I'm most comfortable with because I, because I I've worked with the sway data and, and to some extent with the G sway question design directly, we've included a battery of questions that included battery of questions that try to elicit different types of fertility outcomes. So we ask how many kids you have ever had in your life? How many kids have you had in the past 10 years? So we specifically ask about 2015 and later years and ask for the year of birth for each of those children. In the past 10 years, we, and additionally we ask about planned fertility. So whether you want to have any additional kids regardless of whether you've had them in the past and, and, and at what ages do you at the latest you plan to have those kids. And so the key result that we get in both sway and g sway, so, so I believe these are the numbers from G sway. So, so in our global sample of 38 countries is that if you compare people or who, who say that they cannot work from home to people who are able to work from home at least one day per week, the people who are able to work from home have higher total fertility outcomes. So, so this, this includes past and desired future fertility or, or planned future fertility. And, and, and the relationship is even stronger when not just the respondent, but their partner also reports working from home at least one day per week. And, and so this is true for women and men. So, so these are the full regression outcomes. It's actually not super different. We, we find a significant coefficient regardless of which fertility outcome we look at, whether it's children since 2023, planned fertility, the sum of or or the sum of some past and, and future planned fertility and something similar when we look at sway, where we actually have a much larger sample spanning, I believe from late 2022 up until the latest month. Now how big are these effects? So, so here we're running a thought experiment in which, so on the top row what we have is the average total past plus planned fertility that we observe in our sample as as, as reported by the respondents. And so if we think about how big are these associations between the respondent's own ability to work from home, the partner's ability to work from home, and the sum of the two. So if you think about it, the, the, for a couple in which both can work from home compared to a couple in which neither can we get something anywhere between a nine to 22% increase infertility interpreting the coefficients causally, which, which we'll talk about in a second relative to the baseline. And, and so, so these effects are not gonna probably stem or reverse the decline in fertility that we've been seeing for decades. But, but they're not nothing. So to give you an idea, if, if you combine our estimates with measures of women's labor force participation with the change in work from home that we've seen over the past several years, then again interpreted causally, which we can argue about our coefficients would generate something like 2% more births per year in the us All right, so again, the key question is, so, so we have this empirical relationship work from home, at least in our sway and g sway data seems to predict higher desired and past fertility. The question is how much of this is a causal effect? How much of it is an occupational switching effect? If I really wanna have kids, I can. So in a post pandemic world, I can now look for a job that allows me to, to work from home and, and, and thereby be more comfortable making the choice that I would've made anyway, right? And so, and so we're gonna look at, we're gonna try to tease, tease out whether there's any evidence for either of these two explanations in, in other data. And so, so what we're gonna try to do, and, and here let me, so, so let me skip this. This is, this the specification that I think we're interested in when it comes to the causal effect. So, so what we're trying to do is, is run a regression on the likelihood of you having a child in the a CS on on on whether the type of job that you have allows you to work from home or not. And, and we're gonna allow the coefficient on on your, the work from home ability of your job to vary across times to vary across time. And so basically we want to test whether after the pandemic relative to before people in work from home mobile jobs are now more likely to have kids relative to those who don't have work from home mobile jobs. And, and so, so we've had many, many debates. So, so between, between the research team about how to measure work from home mobility. So, so we've tried the dingle and neiman measure that many people have been using in the conference. So for, for now what we're, what we're using is, is the Hansen etal measure that Nick and Steve and co-authors have, have built based on job postings. So basically this measure is gonna be higher in occupations in which, and, and correct me if I butcher anything in occupations in which a bigger share of job postings are are for jobs that can be done remotely to some extent. Yeah. Alright. And so, so, so we're gonna try to run this, run this specification and so, so we'd love to I think look in more detail at, at Emma's paper because we're not really finding anything. So, so, so there might be a lot of measurement there in our work from home mobility indicator or, or index right now we're still working on this right now. We don't really have any clear evidence in the a CS for causal effect of work from home on fertility. The other side of the story is again, whether people might be changing jobs into jobs that allow them to work from home, especially if they're interested in having kids. And so here we're gonna look at some CPS data. So why is this potentially a concern, potentially not a concern? Well, so when you look at people's occupation, this is something that tends to be very stable later in life, not so much early on in life. So people who are entering their prime fertility years or are still in their prime fertility years, say in their twenties or early thirties, it's gonna be quite easy for them to potentially change occupations in, in the direction that, that we think they might be later on in life. So something like above 95% of people don't really switch between major occupational groups af after the age of 35 or 40. And so is there any evidence of this that, that there's job switching in in the direction of work from home? It kind of looks like there might be. So, so what we, what we are looking at in these graphs on the vertical axis, there's is an index for the work from home ability of the occupation. People report in the CPS during the second wave in which they're interviewed compared with the first wave. And it does look like for people in their twenties, both for men and for women, they seem to be moving in the direction of more work from homeable jobs in when they're in their twenties but not when they're in their thirties. And so there, there does seem to be a little bit of credence for this reverse causality. So let me leave it there. We find a pretty robust relationship in our survey data between the ability for you to work from home or for your partner to work from home and fertility outcomes and intentions. Teasing apart how much of this is causal and how much of this is, is this selection or occupational switching effect not so straightforward we're realizing and so happy to discuss more

- Very interesting paper. So just to confirm, we didn't find anything a CS either, but a CS is cross-sectional data and we, when we try to run the exactly the same model specification, but with maybe education and year interactions, the pretre we can just now get rid of. So eventually we teach a CS data and stick with CCPs CPS individual panel. And, and another thing I think it's very important to do parity specific analysis if you use CPS individual panel because if you just loop all the birth parities together, you, you're not finding anything. But once you specify by first birth and hire all the births, the effect shows up.

- Yeah. So, so absolutely as you were presenting, I jotted down first birth versus subsequent births, which is I believe not something that we've looked at. And and I, I think yeah, we should definitely look at that. I think with the c with the CPS, it's still, correct me if I'm wrong, not super easy to figure out if somebody has had a child in the past year or so, depending on how the variables are reported. We, we've, we've, yeah,

- Small points. So one Danny Tanenbaum has a paper about showing that after women have a child, they're more likely to be in a work from home job. So just in the, in terms of thinking about that mechanism, that paper might be a helpful reference. And then the second idea is like something that Matt Conn and I did in our paper was think about college degrees as something that is like more fixed and like that as an indication of whether you have a skillset that could be done from home versus not. Maybe that would be useful in your setting in the a CS

- As in we I totally agree. We've been looking a lot of it is like trying to find your, in doing the diff of diff Exactly right. It's like what's the thing that's fixed and partly occupation's not that well measured. And I'm trying to remember some of these things is not, is really not that well work

- For home mobility conditional occupation is not that well measured.

- Yeah. Occupation itself is not that well measured and work from home ability, the and Neiman thing just does not, you know, to be honest does not look fantastic because it has some things like elementary school teachers has been work from home and certainly my kids' elementary school teachers are not work. And then so we moved to looking more education, which is a rougher variable but is well measured and it's fixed over time. So yeah, we, I I totally on board. We, we kind of trying to figure it out and it's going in the right direction, but we just didn't get it together for today. Sounds like the job market candidates, they're always like, they have such a great paper, they're not quite together. Yeah. Go for it.

- On, on the measurement error issue in the classification, which I think it's a serious concern there, there's two levels of measurement error here. One is whether you have a measure that is accurately measuring the work from home frequency or propensity within an occupation. We, we think our, our measure of the Hansen and all measure is a little better than Dingle and Neiman on that front. But even if you had a perfect such measure, we, we know, and this is in the Hansen and all paper, there's tremendous variation across firms in the same industry, same occupation in the rate at which they actually offer work from home opportunities. So at the individual level, and these are all individual level regressions, there's still a lot of measurement here that's probably generating an attenuation bias effect in, in the key coefficients of interest. And yeah, I we're out of time so

- Thanks so much.

- All right, well great. Thanks so much. As you're gonna see, there's complementarity between our paper and all the other papers, but particularly the papers in the previous session. So I think we're in the right place. This is joint work with my colleague Yana Gallin at Harris and Stephanie Carroll and it at Treasury. The standard disclaimer applies, these are our views and not the views of the IRS or treasury and an apology and caveat that we can't show you our latest results because the government has shut down and we couldn't get our latest results of, but the re, the results I won't show you are even better than what I'll show you today. Okay. So obviously every parent in the room knows that childcare is an important compliment to work for parents. The burdens of childcare child rearing often fall on the mothers. So especially for mothers and childcare costs and availability is often cited as a major barrier to maternal employment and whether or not there's market failures in this industry. This is a justification if we think female employment is a pro-social outcome, justification for childcare subsidies and credits. But I'm gonna show you what I think is a new fact that there's a weakening link between childcare and maternal employment. And here are two measures of childcare usage. The first, the dotted blue line is from the CPS. They ask a question about whether your family uses any paid childcare. And you can see unsurprisingly that it's down in 2020. In 2021, we know there were huge disruptions in the childcare market, but what is surprising is that it has not bounced back paid childcare usage in 20 24, 20 23, 20 24 is still 10% below its levels in 2019. The second line here, the solid red line, is a new measure of childcare usage from the tax data. I'm gonna talk more about this, but it's showing very similar trends, declines in paid childcare usage relative to pre COVID. At the same time, maternal employment is at an all time high. These are data through 2023 and 2024. It shows the same thing. This is a tax data measure, any wage employment or self-employment. But you see the same thing in the CPS and you also see it for hours. And the puzzle that's in the heart of this paper is that these are diverging used to be the case, that there were positive association between paid childcare usage and maternal employment. And that seems to be broken since COVID And in particular, we're gonna be focusing on 2023 and 2024 when the COVID disruptions are over and childcare centers have have been reopened and people have returned to work. So the key research questions we're looking at here is whether there's been a shift in the arrangements to care for children while working and then what this means for employment and career progression of women and particularly mothers. So in the paper we are through a lot more detail of multiple explanations for these divergent trends. Today I'm gonna talk more about what we've landed on as the main, the main driver of this one. You might be wondering if there's changes in reporting, particularly for our tax data measure. Maybe people are using more under the table care. No, that would be inconsistent. There's no incentive not to report that in the CPS measure of paid childcare. We've looked at time use survey, are there changes in childcare arrangements? You've seen some of these figures already today. It looks like fathers are doing more. And I'm gonna put that with a caveat until I get to the next point. Another, I don't think this has come up today, but grandparental care, there was a huge increase in early retirements during COVID. What we're gonna find is that's true, but grandparents seem less likely to help out. Actually, I know that applies to my family, but that's a different story. Changes in the childcare market is another possibility. So we've looked a lot and in the paper we're doing a lot on the supply side, childcare prices, availability. There have been some UPK expansions including here in California. None of this looks first order to us and relative to other things that we're seeing, the next set of possibilities are changes in the nature of work. And there's two ways we're going to look at this. One is that there's been a rise in work in the gig economy and that's associated with the most flexible types of schedules, at least for gig work in particular, this rise is gonna be too small to explain the magnitudes of the trends I showed you the motivating trends I showed you. And then finally, we're at the remote work conference. Remote work's in the title. This is what we're gonna find. That seems to be a primary explanation for the decline in pay childcare usage, the single most important factor for mothers as well as for fathers. Stop

- There.

- Yes. So what data we're, one of the advantages of what we're doing here is we have a longitudinal link via IRS tax data. And in these data we can see paid childcare usage because when you have paid childcare usage and you have a tax liability, you have an incentive to file for the child independent care tax credit. You also have a filing obligation of the, of form 2, 4, 4 1 if you have A-D-C-F-S-A. So you might do this or your accountant might do this, or TurboTax is doing this behind the scenes for you. We have links to other tax forms. The 10 40 self-reported occupation is going to be important for us. We know what firm you're working at on the W2, we have links to whether or not you're a gig worker or other self-employed. We're gonna link this to the IRS has actually linked the self-reported o occupation. They've mapped this to on net codes, which will then map to, we're gonna map to the dingle and Neiman measure that many other papers have been using. We also can see your grandmother. We can construct a three generational link between the child, the mother, and the grandmother. We have SSA links and we can also see who's writing who on the 10 40 for dependents back to 1999 to the present. Okay, so let me, and then, so we're gonna, first part of the paper, we're going to look at trends for people who are in occupations that are tele workable. And then the next, the second part of the paper we're gonna look at and do an RTO design around, around the return to office, comparing office first firms hybrid and remote, completely remote firms. So the first piece of this, the data piece is this form 2, 4, 4 1. We have, we started working with these data to better understand childcare markets in the US and then we saw those trends I showed you in the beginning and then we started to write this paper. But it's a pretty cool tax form in that you write down the care provider's name, their EIN or SSN and the amount paid. And you, you, we can see this for every child in the household occupation on the 10 40. So when you fill out your 10 40 or when TurboTax does it for you, it might ask you your occupation. It goes at the bottom here of the 10 40. And so this is where we can see what occupation the primary and the secondary and and their spouse has the IRS has mapped these to own IT codes. It's an effort Bruce Sasser dot has had been working on and and a team using machine learning methods over the past few years. Side note, these are the occupations that the IRS has released that are eligible to have no tax on tips. So it comes from these same data and we're gonna map these. The IRS has done the work for us to map these to OEC codes and then those easily mapped to the dingle admin measures. So here are what those occupations look like. I think this came up in the previous talk. You know some of them you might have some quibbles with. We were thinking about focusing on a subset of these occupations. For now what I'll show you today, we'll just take this off the shelf. These are the most common occupations that are tele workable for mothers with kids and and non tele workable. So here are just trends in what is going on with paid childcare usage by whether or not your job is tele workable or not. This is much noisier in the CPS, but here I think you patterns really pop out. There's a bigger decline. Well, two things to note. First, you're much more likely to be using paid childcare if you were in a tele workable occupation in the pre period Y. These jobs tend to be more likely to be full-time and high paid. The occupations I showed you on the previous slide. So there's higher rates of usage, but these are where we're gonna see the biggest declines in paid childcare usage relative to the jobs that are, are not classifiable as tele workable. You might be wondering about sorting, this also came up in the previous talks. We're not finding a lot of sorting among mothers of older children, but there is a one percentage point increase in sorting into tele workable jobs that mothers of the youngest children ages zero to four. There are other changes in the labor market that I've alluded to occurring in this time period. There's a huge increase in mothers are doing platform gig work for mothers. This is often gonna be delivery or grocery delivery or food delivery. It was half a percentage point in 2017. Now three and a half percent of mothers are working in platform gig work. The second trends, there have been trends for a long time and the rising availability of a grandparent home, a lot of this is coming just from the fact that they're more likely to be alive. But also also retirements increase during the COVID period. Changes in father's arrangements. Fathers saw decreases in employment during the COVID period, but they have higher rates of employment in the post periods. Not quite as high of an increase that we saw from, but there's still an increase here. And there's also a similar slightly more sorting by fathers into tele workable jobs than we saw from others. So the first exercise I'm gonna do here is a statistical decomposition of that decline in paid childcare that I showed you. We're gonna compare pre period to post period. We're gonna look at 2022 to 2023 to remove the main COVID disruption of 2020 and 2021. And the question here we're asking is just what is the association between paid childcare usage and each of these factors that I've been talking about, whether or not you're in a tele workable occupation, whether your spouses, whether or not you're self-employed. We've looked at self-employed and self-employed and platform gig work, which I'll focus on here, whether the grandmother is retired and separately the grandmother, the grandparental distance from home. And so we're gonna regress whether or not you use paid childcare on these factors. We'll add some controls in here too for mother's age, youngest child age, number of kids, mother's earnings, and the county childcare price, which turns out not to matter at all, we're gonna decompose Oaxaca style, the decline in paid childcare into changes in levels because the levels of these things are changing around and changes in the load of these factors on paid childcare usage. So that's going to be from the post coefficient. And so here's what that decomposition looks like. So the very first thing I'm gonna show you here is mother tele workable. What's on the Y axis here is the share of the decline explained, and I'm separating out by changes due to the level and changes due to the beta. And so changes due to the level, there was some sorting into more tele workable jobs. I showed you in that figure that those tele workable jobs have higher rates of paid childcare usage. So it would predict, the model predicts that there would actually be a red, an increase in paid childcare usage. So where you have negative ex explanatory power, but the load factor, oops, the load factor is what's gonna matter a lot. So even though you were more likely to use paid childcare before, there's a a big reduction in paid childcare usage among mothers in tele workable jobs. And so that alone is gonna explain 56% of the decline. We can attribute mothers and tele workable jobs to 56% of the decline in paid childcare usage that I showed you. What about fathers? Well fathers have, there's two things going. One, I'll focus on the one further on the right first, whether or not the father is working. So fathers are slightly more likely to be working and that generally means you need more paid childcare. But the load factor on fathers working, and this is in non tele workable jobs, is showing that even when fathers are in non tele workable jobs, you're less likely to need paid childcare. So one story here is maybe they're getting more flexibility in regular jobs, but we see a a slightly bigger effect on an additional effect for fathers and tele workable jobs. And so together each, each one of the fathers and tele workable explains about 21% of the decline and fathers and working explains another 11% or so. And then the total explained by fathers is 35% gig work. There's an increase in levels, but it's qua not quantitatively large enough to explain the decline. And then grandmothers is actually, grandparents are a mixed bag. Here we're focusing on grandmothers. They're more important than grandfathers. Retired grandmothers do less. So you can see a couple of things are going on. In fact, people get a little bit further away from their grandparents during the, during COVID. As they move they tend to move a a slightly further away only about a mile. But that contributes to a decline in in in that share that's explained. There's an increase in the load factor, but grandmother retired actually there's a big negative it, it explains much less. And so one story here is that grandparents are retiring but they're not retiring to help out. Maybe they're retiring for health reasons and consistent with the Brookings paper by Abraham and Rendell. So we're able to explain with these factors, 81% of the decline in paid childcare usage. Okay, so just some quickly, some supporting evidence from the ais. And then I wanna get to the R-D-O-R-T-O design. This is a figure people have shown. The one difference here in our blue line is that we're restricting to parents during working hours. We can see that there's been this persistent increase and as of 2024, about 30% of working hours are spent doing remote work among parents. It lines up pretty well with the SUA measure sway measure. Sorry, we see, and I think that we saw this in the previous session, at least the fathers one we saw in the previous session. Fathers are doing more on average about five more hours a week. My figure is a little bit different than the one we saw before because I'm restricting to working parents. But you see actually mothers look working, mothers appear like they're doing more as well about the same as fathers. The other thing to note is that fathers just have a much, much lower level than working. Working. Fathers have a much lower level than working mothers in terms of childcare they offer. And then we can also see an AIS that grandparents are doing less. So all this is consistent with what we're seeing in the tax data. Okay? So tele workable jobs appear to be a large part of the story of the client in paid childcare. We know telework ability and flexibility is something highly valued. And perhaps more so today. There's mixed evidence on productivity and careers by people in this room and papers you, you've seen and or will see. And so there's this trade off, fundamental trade off here between savings and child co cure costs and maybe career outcomes. And so our next design we hope speaks to that. So here we're gonna exploit an RTO design. Hopefully we haven't reinvented the wheel but we hand collected or our, our A did RTO policies, we have the Fortune 500 now we've gone through public news article searches and transcripts of earning calls. We're extending this to more companies as we speak. We're gonna classify firms into three groups. Firms that stayed remote through 2024 firms that stayed office first. We have two types of hybrid, but today I'll just, I'm gonna group them together and our, the simplest design is one where you assign mothers to their 2019 firm and follow them through 2024 ITT. The best case scenario for us is that these RTO policies are exogenous from the perspective of worker outcomes. We've talked, we've seen papers a little bit that suggest that it's up to the CEO and manager preferences matter a lot and based on public statements that CEOs have made that that seems to be true for or there's, there seems to be some truth to that. But let me in the interest of time show you what we're finding. First, we're seeing big declines in any paid childcare among mothers who work at Fortune 500 firms in 2019. This is bigger than what I showed you before because these are people who tend to make more money. And so you see they have higher rates and and bigger declines. So our main event study design is just going to compare firms that stayed remote and firms that stayed hybrid relative to firms that stayed office first. We'll add in trends for the N code in case there's differences in shocks that are facing different industries. And we can add state fixed state trends as well in case there's differences across states. As these fortune 500 firms have people working across multiple states, all this is defined as of 2019 and we'll cluster our standard errors at the company level. So the first outcome here is the event study for any paid childcare. And so first look at the red one. That's going to be firms that stayed fully remote and we're seeing there's no no difference kind of as you'd expect in 2020 and 2021. Almost all firms are remote at this time. But then the firms that can stay remote through 2024 by 2023, they have a four percentage point lower usage of paid childcare relative to the office. First firms we're not seeing as big of a decline. It's not statistically significant for the hybrids, but this is grouping all the hybrids together and the next set of results we clear will have some gradation among the hybrid firms and now onto wages. So first look at the red one. So these are the fully remote firms relative to the office first firms. And maybe there's a a, a little bit of a penalty here. Our results are noisy, not statistically significant, but actually we're not seeing that there's a penalty by 2023 and 2024 for hybrids, it's a similar story, although it is significant early on. But we're not finding statistically significant differences. Although some caveats here that that this is noisy. We have some other designs we're doing comparing parents to non-parents in the firm. Then we can have the firm fixed effects or fathers and mothers within the firm. We're not seeing as big a reductions in paid childcare usage if you're a father at the Fortune 500 firms. So those are some other designs that that we're working on. Here's a tabular version of our results. Looking at the average in 2023 and 2024. So you're seeing that decline in paid childcare usage. The average reduction in expenses, although this is going to be underestimated in the tax data, we're not seeing any differences in the probability of working. Again, this is the ITT design. So we're not conditioning you working in other periods. We're not seeing statistically significant differences in wages. The remote one wants to go positive, the hybrid point estimates go negative. We have also looked at fertility. So the sample here are working mothers, these are all women at the firm and then we're looking at whether they have additional children. And here we're finding s small but statistically significant negative effects. So one story here could be that if you are spending more time with your kids at home and trying to work at home, maybe you're deciding not to have an additional child because you're already spent, you're already pretty time constrained as it as is one other set of results we have. We looked at attrition at these firms and is it differential by whether or not you're at a office first or remote or a hybrid firm? It is. You're much more likely to leave the firm if it's an office first firm. Only 42% stay compared to the remote where 48% stay. And so one story here is that the office first ones are almost mechanically going to be sorting to if they, even if they picked a firm at random, they'd be more likely to pick a hybrid or a or a remote firm. And so we might be biasing us as downwards. Now this, there's some caveats with this, but this is, we're focusing just on the job stares and it's suggestive that we're seeing, we're seeing suggestively much larger effects here, bigger reductions in paid childcare usage and we're not seeing a wage penalty. In fact, it looks like it's, again, there's a almost a premium or this is suggestive of a premium, not a penalty on wages. Alright, so I'm out of time looking forward to the discussion. I'll end it with just this quote from Golden in her in the 2014 a AR paper. She said to close the last chapter of the gender pay gap, something needs to reduce the conflict between inflexible job demands and motherhood. She wasn't thinking about remote work at that time in 2014, but maybe remote work might, might be it. So looking forward to comments and discussion. Thank you.

- Super interesting talk. Thank you. As we, I mentioned during the lunch, don't forget to control focus private equity in the local. And another question I have for your paid childcare, do you include like nanny share or just like any cash transfers for the paid childcare?

- So what this is, is any paid childcare that you use, you can deduct that on your taxes. So you would indicate our, our measure here is an indication of any. And then we had the intensive margin measures of how much you're listing. We think that how much you're listing is probably underestimated because you might have other forms of childcare on top of what you tell the IRS about not saying I do, but so I don't know if, does that answer your question? What is being measured in the 2 4 4 1 And then the CPS measure should be comprehensive. There's no incentive or disincentive to under report under the table of childcare. For instance.

- Nick, it's great you, you have, this is, I wanna ask a question. It's almost like another paper you could write here if I understand it, which is you look at the effects of RTO on attrition and you have all the employees in these Fortune 500 firms. There's a huge debate on this and typically people really don't have the data of they do, they go to like LIO labs or something. But that's very incomplete. So it would be fantastic to look at these impacts of RTOs on attrition and split it by like owners, managers versus non-managers rate of earnings progression, which is stars versus non stars occupations, genders, gender with kids. I mean you could write an amazing paper it said it's a really big debate 'cause the claim are that these RTOs high flyers and various people exit the firm but nobody knows are com you know, I know that's not this paper but I think 'cause I've never seen data you've matched as far as I know, the RTO data up to the RS tax returns.

- Yeah we're able to do that because we see the firm name of where people are working.

- Yeah and there's, there's actually data on up to kind of the, you know, 2000 firms. So I can talk to you about it after this, but it's a separate totally different paper.

- Yeah, I'd love to talk more about that. It's a

- Really big policy question. Well that's policy but firm.

- Yeah, I'd love to talk more about what the, where the literature is on that and

- Because right now you're finding RTOs are causing attrition of what women with kids is it just to understand, so

- Right, so this is for, oops, this is for mothers. But we could look at, you know, we could look at women more broadly or the different groups that you mentioned.

- Okay,

- Yeah. But this is for parents, mothers,

- Yeah. Chris. Oh yeah, just I had a separate point but just to add onto to that, you also see where people ended up because you can also track mobility if you see addresses. And so you can see a people who moved as a result of their firm's work from home policy or more, also more likely to not be the ones to go back to the office. I had a point about the decomposition, which is I think that when you are looking at fathers who are in jobs that are non remote, there are big changes in the wage structure for those jobs. And I'm wondering if higher pay at the low end of the labor market, which is where most of those non-REM remote jobs are coming from, allow you to have sort of a joint household labor supply decision to where mothers are more likely to stay at home if fathers get big wage bumps if they have to go into central or frontline job.

- Yeah, I threw this under the rug a little bit. We are not looking at the extensive margin in that because we need to know your occupation in 2019 and so we need to know if you're in tele workable occupation in 2019 or not. So those are all mothers who are working in 2019 or women who become mothers are, are, are also in there too. I think we, you know, we could do a bit more on the extensive margin. There were some others, other papers focusing on that that maybe had better designs in order to do that. But that's interesting. And we could also look at wages. I think if, but wages are increasing, you'd expect people to be able to afford paid childcare conditional on them increasing faster than prices in paid childcare. So there have been increases in prices. We've, we've looked more at the supply side as well. I didn't have time to talk about today, but the interplay of yeah, wages and and prices at the state that we could, we could think about at, we can think about more. Yes, please. I I was curious

- About the role of informality.

- I understand childcare you're talking about on the supply side of the childcare market. Yeah, so this is something you, you probably don't list your informal care to the IRS. So that's a limitation of the paid childcare measure. If I go back to the the first slide, you can see there is a level difference between what the CPS shows and what the tax data show. That's not the question. Yeah. So we think like part of this is, sorry, part of this is yeah gonna be informal care that you're not telling the IRS about. It's not, I think when I first, when we first made this we said hey, the IRS data is not that bad. It, I think what a lot of times people have informal care, they also have formal care on top of it and that can report that formal part to the IRS. What's most important for us though in this paper is that we're seeing similar trends across the two, the decline in paid childcare between 2020 and 2023 in the tax data and the CPS is 10% in both of these. We don't have, we're still waiting on the 2024 tax returns to be complete with the late filers. We just had our RA make the 2020 other people probably just started digging into the 2024 ais which just became available. That one seems to go up a little bit in in 23 and 24, but they're both still far below the, the the, sorry this is from the asec, not the ais but there's, it's still considerably below the 2019 level. Steve,

- You, you controlled for a bunch of things in your regressions but it also be just be useful to see like for this chart, how things vary across states and things like that. They had very different yeah pandemic policies and so on.

- So I can tell you we have a state figure and it's cleared so I can talk about it. Every single state has seen a decline in paid childcare usage. It is bigger in some states relative to others. And so we've been trying to think about, you know, what might cause some of these differences. So California has a big decline. California is a place that has a lot of remote work, also had UPK. So that's why we can tease out some of that in in the regression by controlling for these different things across states and and different state. The share back to the question about informality, actually we haven't done this yet, but it's on deck to do, to control for the share of immigrants, the increase in the share of immigrants which differed across states. There was a big increase in I immigrant shares in the last few years of the Biden administration and that could affect childcare markets.

- I wondered if you could get information on even state policies that give resources for homeschooling and things like that.

- Yeah, I think so. For it to be an issue for us it'd have to be correlated with telework ability. It could be, I don't know if there anyone has looked into that. It is an interesting point. So we'll see if we can get data on, maybe we can look at the shareholder. There must be some way to figure it out because we know how many kids go to school, we know how many kids live in every state. So maybe one minus that is is a good proxy for the homeschooling share that it's, yeah, thanks for that. We'll we can, I think we can probably do something there to see if it's related but if anybody knows any other papers on that, happy to just cite that too unless it's something we need to control for and then we'll do that. Great. Well I think we're out of time. Thanks so much everybody for the feedback.

- Okay, thank you. Thank you for having this paper on the agenda today. I'm presenting joint work with Sela and Den Lakin where we are looking at how flexible work drives occupation choices. And this is gonna drive change the gender gap. So women continue to have lower labor force participation and lower earnings than men. And this has been largely attributed to the child penalty, which is the decline in women's earnings after the birth of their first child. And the literature argues that occupational structure is key, specifically the presence of greedy jobs where there's nonlinear returns to ours. So men tend to be overrepresented in these jobs. So this might be being a physician or surgeon, being a firefighter, being a lawyer. Men are overrepresented in these jobs and these jobs are such that they reward long continuous hours and inflexible work schedules which are incompatible with childcare responsibilities that typically tend to fall on women. This leads to lost wages and lost human capital accumulation for women over the lifecycle. The question we wanna ask in this paper and more generally in our research is how will large changes in labor market flexibility affect this division of labor? The app application I'm gonna show you today is gonna be working from home, but we could also look at other disruptions such as AI adoption. First of all, I want to show you that there is occupational sorting into linear versus non-linear occupation. So if we take the occupations by hours in the CPS, rank them according to hours from highest to lowest and we just take above median and that's our definition of non-linear occupations, then we find that women tend to be in linear occupations rather than non, non-linear. And this hasn't really changed over the years. Remote work adoption can be seen as a permanent change in the structure of jobs. We study how this is going to change household occupational and labor choices. Specifically how women are going to possibly go more towards high return occupations and then how this is gonna lead to long run adjustment of the labor market. What are the implications gonna be for gender gaps in hours worked human capital and earnings over the lifecycle for everyone. How are we gonna do this? We're gonna build a hydrogenous agent model with occupational choice and labor supply. We are gonna highlight the role of occupational sorting in the persistence of the gap and we're gonna study the effect of change in work from home flexibility. What's gonna be novel is that we are going to allow for occupational switching throughout the lifecycle as well as initial choices. It's the labor supply and occupational choices are going to be a joint household decision. So there's gonna be an impact on the choices of men and we are gonna look at the general equilibrium. So wages in non-linear and linear occupations are going to adjust as we've seen before, there's not a lot of occupational switching specifically in the category that we want to look at. So if you look between ages 25 to 55, there's not a lot of changes in occupations across linear, non-linear categories. But specifically we are interested in women switching from non-linear to linear occupations and this is about 3% in the data. And Nick, this is also over stated because in the CPS, if it's a different interviewer who goes to the respondent, they might classify the same occupation differently in the following year. So what this means is that we are going to see effects only in a very long run, right? And for that reason we're going to write a model. The model's going to be a lifecycle model with three distinct stages. There's going to be an initial period where before observing your productivity you make an occupational choice and there's couple formation which is going to work through assortative matching. Then we are gonna enter the working life of the couple where they make choices about consumption savings and whether they want to switch their occupations. Each person during this period you have staca stochastic arrival of children. We don't model fertility and at the same time your human capital updates with some shocks so you might accumulate human capital. The retirement period is fairly simple where you solve a consumption savings problem with some probability of survival. I'm going to talk mainly about the working life and then a bit about entry. And we are gonna skip retirement 'cause for time. So during the working life you've already formed a couple. There's a male, a female, they're both of the same age, they work for JW periods. We assume the household is unitary. So it makes decisions together about consumption savings. Our supplied by the man and our supplied by the female in every period we have A-C-R-R-A utility. So K is the number of people in the household utility scales by K and the consumption's divided by all the K people in the household. There's a a risk aversion and we have this utility from our worked which is also convex and we assume it's the same for men and women. Now at this point, every household member already belongs to an occupation. We assume that there's one non-linear occupation and there's one linear occupation for now. And there are some switching opportunities that arrive through your working life with some probability that depend, that is determined by where you are in the life cycle and the occupation and switching is subject to a utility cost. Households have to choose their labor sup labor, supr supply. These we keep as not working, working part-time, working full-time and working overtime. When you work, you earn, you earn a baseline wage in the occupation which is different across non-linear and linear occupations. You have a lifecycle return in the occupation, a lifecycle trend and then you have convex returns to hours, which is different across linear, non-linear households. In addition, you have a productivity term. This is AR one with positive drift which and the drift depends on the hours worked. So what this means is in the static, like when we look at the static setup, we already have higher returns to hours for non-linear occupations. But because the positive drift of the productivity process depends on our as well, there's also gonna be dynamic returns to working in non-linear occupations. Households enter this period with either no children or a young child to match the data. Children arrive statistically in our model we don't track the ages of children. So what we do is we basically say that might be in the young state or the old state and there's some switching probabilities again to match the shares in the CPS on retirement the children leave. You go back to having a household with two members and we just track one child, which is the youngest child for consumption equivalence purposes. So you can have kids two if you have no children, 2.25 for a young child, 2.5 for an old child. Older child childcare has a utility cost. This depends on both parents' hours and the child's age. We don't say whether it's costly or whether you like to enjoy time with your children. We are neutral about that. But you can take it as Steve said, that you like to have time and that's a utility cost. The people like spending time with younger children, more both men and women. Women like to spend time with children a bit more. And then what we have is for the whole household we have a parameter which captures the childcare costs and we are gonna assume that as one in the baseline pre COVID. And then we are going to have that scaling down the whole household's childcare costs in the work from home economy after COVID. Going back to the initial career choice, before forming this household, each person has to choose an occupation. They haven't observed their productivity yet, but they face some cost barriers. This is to capture anything that's not the child penalty in the different, in the wage gap, in the gender gap. And they agree on the expectations across different productivities, different occupations, what's gonna be the initial value function and also your choice of career is going to affect your match. There's gonna be a bit of taste shock so everybody doesn't choose the same, doesn't make the same choice, but non-linear is going to positive, have positive assortative matching with non-linear and linear. With linear we don't really model the production side, we just keep some representative firm because we want the labor market to clear so that we can have wages adjust. But basically the firms have some elasticity of substitution between non-linear and linear occupations and men and women are the same to the firm. We can change this later on the on the calibration. Are there any questions about the model at this point? Save question till the end. Sorry, save it till the end. Okay, so we then take this to the data. We keep standard parameters for preferences and demographics. We match to the data. We could have these internally calibrated but we don't right now. And we match the switching probabilities for productivity to match the PSID either returns from working full-time or working part-time or if you lose your job and you come back to the labor market, the kind of wage loss that you face for the substitution between for firms, between non-linear and linear occupations. We go with what Emma U suggested, which is taking college degrees as a premium. If there's a better suggestion for this, that would be very helpful For the internally calibrated par parameters, we have the cost of hours in the utility function. So this is the disutility from working, which is the same for men and women. We match this to the share of men and women in part-time, full-time and overtime status in the CPS. And then in the childcare cost functions we have the, the cost of working for men and women for young and old children, for young and old children. We match these to the employment shares in the data of employed work, employed men and employed women with young and with old children with these age cutoffs. But actually we in practice, we jointly identify all of these and then we have this deterministic lifecycle profile of the occupation looking at men to back out the parameters in the non-linear and linear sector. So there's one which is a level difference in linear, non-linear. Then there's an H profile and then there's some to match the data. Essentially we have concave functions for the second parameter for the square term. So the moments we target. So we don't target any moments which are male versus female, but we are going to look at them. We do have for, for all of these, we look at part-time male, sorry, male part-time, full-time overtime and so on. Same for women. And then we have these two wage premium like I said. So we have an over-identified system, we have about 13 of these moments and we have about eight parameters for the cost function. If we go take this to the data for the baseline model where the fire, so work from home hasn't kicked in fire is one we could do better on the model fit, but we basically capture two things. One is the share of men rather than women is sheriff men in non-linear occupations is higher than the sheriff of women in non-linear occupations. And the second thing is men's wages are higher as in the data and wage growth is higher as well. Now this is a parameter, which is our key parameter for the presentation and for this paper we want to estimate what is the reduction in childcare cost coming from work from home. And we thought about how to do this and the first thing we went for is what is the decline in the motherhood gap from work from home? So we started with MIO paper where we took the elasticity of the motherhood gap to work from home and then we tried to multiply that out with the a CS change in work from home. But that was about 5.7 in 2019 to 13.3% in 2024. But that gave us like a decline in the motherhood gap of 18 percentage points. So then we didn't go with that. So we should talk more about how to use the elasticity correctly. What we did instead is that we started, we looked at the CPS and we said that around these years how much did labor force participation for women, for mothers with young children change? That turned out to be 2.27 percentage points. So that's if you look at 20, 23 and 2019, that's the difference there. And our baseline gap in the model, which is a little lower than that in the data, is about 11.9 pre COVID. So we take this and essentially from this we get the work from home cost parameter, the to the to the household, the reduction in cost. And what we're going to do, what I would like to do is causally identify. So of course this changes including the the work from home part as well as any long run trend decline in the motherhood gap. So I'd like to have a causal identification of how much work from home change the motherhood gap in COVID. So any suggestions on that would be helpful. So after we do this, we set a fire to the new work from home world and we first see what's gonna happen in the short run. So we keep the initial occupation choices fixed, we keep the initial positive assortative matching fixed and we, but we still do allow for switches during the lifecycle during the working age. What this does is that this does already lead to a short run increase in female employment and women in non-linear occupations. So you can see from the first figure on the left that women of different ages start participating more in the labor force. The share of women in non-linear occupations goes up as well and men switch out from non-linear to linear occupations. One because the childcare costs are lower. Second is because of the income effect. Okay, so that's in the short run. Now these effects are not so large because occupational switching is slow. So what we do then is we go to the long run where we allow initial occupation choices to adjust as well and positive assortative matching to happen again where non-linear pairs with linear and linear pairs with non-linear again, what happens now is women tend to, there's a higher, there's an increase in women choosing non-linear occupations. Men start choosing linear. And because of this we are going to have, when these two combined, we are gonna have an increase in the joint linear households that you can see in the left graph on the Yelp first yellow bar. There's gonna be more linear linear households. Now because more women enter non-linear occupations, this is gonna push down the wage premium of working in non-linear occupations to linear occupations in the long run. And actually linear wages relative relatively are going to increase. Men still have higher employment. So they're more likely to work in non-linear occupations over the lifecycle. So this is the third set of bars, the third bar there. And so essentially once we allow for this full reallocation to go through across time, we are going to get much larger long run effects of work from home through occupational choices. There's gonna be a significant shift for female employees. They're gonna have higher employment, higher hours, they're gonna be working more in the non-linear sector. Wages are going to adjust and this moves reallocation a bit except that men still go from nonlinear to linear occupations here. So we can see that here women tend to increase hours particularly during the years in which during the ages in which they have peak childcare costs. And we also see on the right panel that there's a new pattern of occupations over the lifecycle. Women shift a bit to from linear to non-linear occupations and men relax maybe a little too much in our model. They relax towards having lower shares in non-linear occupations and more in linear. So I'm gonna skip the conclusion so I can get your feedback. This is work in progress and thank you for your attention and I look forward to your questions Nick.

- Megan, that was great. One, one question to add is what's you had utility in there you must put to calculate, oh maybe I missed it, but a welfare effect of all of this. 'cause you, they're maximizing your utility function. So

- Yes,

- It'd be good to have a welfare 'cause with this set up you can do welfare and then you can also do various policy. I mean

- Yes, - The big thing I, I think first step would be like as you're doing matching it and look at match the micro to macro data and once you've got that you can do something that's really valuable, which is what's the impact of remote work and welfare and given that what a policy thinks, like how could I improve welfare? So in this I presume remote working, the, the option of it improves welfare.

- Okay, - I just as a question, I'm guessing it does,

- Yeah, we could do like the utility, some of all the households basically.

- Right. And I, in terms of output, I think it also increases output. This is a big debate that we've had 'cause of, you know, there's no productivity effects here, but I think there's labor supply effects. So my sense were remote work actually quite large output effects in the US not through productivity that people are obsessed with, but actually through labor supply.

- Okay.

- I mean this

- Is, she has induced productivity effects because it changes the mix of jobs.

- Yes,

- Yes,

- Yes. Yeah. But there's, there's so many things you can do. I guess I would, I would look at some of the big macro policy.

- Okay, thank you Steve.

- There's a lot going on here. So I wanna make sure I understood the, the only difference between men and women on average in your

- Yes

- Way you set up the model is women have a preference, a greater preference to spend time with kids.

- That's one. And the second one is that they have a cost that's higher in non-linear occupations, which captures like discrimination or you know, sort of just not having enough women there. I see. So that to match the data. So there's two parameters that, do you

- Need both of them or, or can you do it all with one?

- I, I'll try. I

- Mean I, it's, and and the other, the related question I had is I I I wanted to get some sense of how big a difference you need on those two dimensions to get the differences in your outcomes. 'cause this is related to discussion we were having during the break. There's an endogenous specialization here over the life cycle and what kind of job you choose, right?

- Yes.

- And, and so do you need a lot of heterogeneity difference between men and women on average to get the outcome differences? Or do you only get a little bit need a little bit of difference?

- So we are matching, maybe this answers your question. We are matching these cost parameters with part-time, full-time overtime shares. What are the exogenous

- Differences in movement? Where are those parameter values?

- Those the exogenous differences are these parameters, right? So it's the phi, female, old five female young versus men and then there's a cost.

- Okay, what I'm, this is for the next version of the paper. Yes. But somehow I want a way to help me think about whether those exogenous differences are small or large.

- Okay. Exogenous differences. Okay.

- Okay. El I like they very much this gender blind setup, but I was curious whether you've built into it any notion of a absolute comparative advantage between market and market activities between one and women. And the flip side of it is the earning earnings and stability on nonlinear jobs, also conditional employments. These are jobs where, I mean the attachment of women is, is much lower than for men And even conditional being employed, there's a lot of variability. You know, variable pay plays a big role Yes. In some of these jobs and how that factors in, in the persistence of the gender wage gap even once you low,

- Okay. Okay. Understood

- Or wrong because a certainty equivalent calculation would be very different for men. The certainty equivalent calculation.

- Yes. Yes. I understand.

- Even for same risk aversion parameters could be different. Yes. For the two.

- Yes. Thank you. That's a good point. Yes. Crystals awesome.

- Yeah, so very, so yeah, so it was great to see a model like this because I think the, we, we need a lot more kind of structural labor to disentangle some of these things. Yeah, I would just to double all Elena's point about the earnings volatility that we see performance pay fixed wage. I wanted you to go back to the human capital production function 'cause it looked like it was, is it just that there's this linear versus non-linear or is there some, like how are people accumulating human capital? That was something that I missed and then I have a follow up after you clarify that. I thought it was like one of the early slides, right?

- Okay. I mean there's a pro, basically there's a productivity process and you can switch across the values of the productivity grid and that depends on the hours that you put in.

- So if - It's AR one with the drift depending on hours and there's a row somewhere Yeah.

- But more hours means higher.

- Yes. It's a positive drift. Yes.

- Okay. Yeah, because you know the, the kinda debate like the learning by doing Ben Perth.

- Exactly. Yeah,

- Exactly. Okay. So as long as there's some way of capital,

- It's human capital accumulation. Yeah. Yes, exactly.

- Yeah. Oh, and then this is a quick comment was about the, you asked a question I think to us about is there something in the short run that you can use to capture? So I forget how you,

- The causal effect of work from home on the motherhood gap or the gender gap to match for, for phi.

- Yeah, so we can talk about this more, but there's a lot of variation. I mean this came up in prior co talks as well, state level variation and within state variation on childcare regulation. And I mean some of it is about just like reopening, not reopening, but there's also just like ra, the ratios that some states mandated on a number of kids to stu to staff. There's a lot of variation there. And so I can share some data with you that

- Okay, thank you

- To that regard. Thanks. Yeah. Awesome.

- Emma. So I, I, a couple of thoughts. I really appreciate that you tried to use our estimates. I, one argument we make in the paper is that maybe this like intermittent availability of work from home is really important. So it's not necessarily crucial to work from home all the time. But having access to even a little bit pre pandemic was helpful in increasing women's labor force participation after they had children. That might be like a harder number to get your hands on of how much that increased With the pandemic it certainly increased much less than fully remote work

- Did. I see, I see.

- But one other thought is there is a paper from someone at Yukon that was their job market paper last name song that is trying to use the pandemic itself to estimate sort of how much it, it changed mother's labor force participation. So that might be like a more portable estimate for your paper.

- Thank you. Thank

- You. Yes.

- It's super interesting. I don't

- Think it's, I heard you,

- I think it's a super interesting project. Ah, okay. Thank you. So it's more of a clarification question because maybe I didn't understand, I didn't get it. So, but the amount of occupations that are linear and nonlinear, it's also an endogenous like object in your model. So you have this general clip

- After the baseline. So in the baseline, no. Okay. With no work from home, it matches the shares of men and women employed in non-linear and linear occupation.

- I see, I see.

- But then after that we let it change. Right. And then they change the shares and that's how I get the men going into linear.

- Okay. No, this one I, I didn't mean. Yes. So, so and and what is the effect then? So we see this, these increase in nonlinear occupations.

- We see that women tend to go a little into nonlinear occupations that changes the wages, got men have an income effect and childcare costs going down. So they switch more to linear occupations as well.

- Yeah. So thank you very much. Thank you. I was,

- Thank you. Thank you. Thank you very much for putting the paper on the program. I'm really happy to be here. And I'm gonna talk about indeed what happens when parents work from home. This is joint work with Pascal Asha and Anu Valier. Ma acknowledge, we started this project in 2018, so not very fast, but we thought this was actually a bit of a niche topic to work on teleworking, but now we are actually, you know, still actually it's still in progress I have to say because we got a lot of feedback since the working paper has been out and we want to actually still add, add more data. So you'll see, I think we have great data, but I think we can do e even better. And also before, so thanks for putting together this great conference. I think also want to acknowledge the huge public good provision that you guys are doing and sort of first putting together this, this conference, but also all the data collection that you're doing on sort of work arrangements. So that's really, really great. And I'm actually gonna go straight there. So I'm gonna skip most of my introduction because we've already had the introduction many times about how remote work is important. So this is indeed from, you know, the data of some people in in the room where we look at the prevalence of remote remote work. I'm showing this figure first because the study is gonna focus on the Netherlands and this is indeed sort of showing the prevalence of remote work post pandemic and the Netherlands is, is basically around the sort of average of the sort of developed, developed countries at around 1.2 days a week. Now this said our study will actually focus on a pre pandemic period, and there the Netherlands was actually a bit ahead of their time. So in 2019 it was not something that was completely unusual to work from home. They were around, you know, 20% of people would report occasional tele teleworking and, and around seven 8% would actually do it on a regular, on a regular basis. But so as we've discussed on many times already in this conference, the question is where work from home could actually be, you know, this sort of chance, this policy potentially that could really help people achieve it. All this sort of family friendly work policy with the sort of potential to improve work life balance. What I want to again acknowledge is that there really sort of two key mechanism. So one is that yes, you might actually be spending more time at home, it saves on, on daily commute time, about an hour a day from what I understand from your, your work. But also it gives a lot of flexibility in when to work. And when we look at children, I think this is really relevant because if you think about, you know, how you may organize your day, you may actually be much more able to free say the time from school pickup to actually, you know, bedtime and sort of shift some of your work to the evening, for example, which I think is really interesting. And so the potential is indeed more time for child related activities. And so really related to severe sort of earlier paper, we think that must, you know, potentially indeed affect children. Now there's a downside potentially is that as we know, and particularly during the pandemic, you know, the idea of working from home sometimes can raise, you know, challenges of, you know, less separation between work and and personal life and also possibly more conflicts as well at home. And so we sort of approached this study a little bit thinking, well it's not so clear that it should actually go in the sort of a positive manner necessarily. You may actually have, you know, challenges that you didn't have before. So, so there's a huge, you know, actually I thought that we were done with the literature review and I'm gonna go home with like 20 papers that I have to add because it's like really amazing to see the amount of, you know, work that is done in this space. But of course we've looked, the literature has looked, you know, at the impact of work for, from home arrangements on productivity, wages, career prospects, job satisfaction, work-life balance, it's sort of subjective measures of work-life balance in particular quits wellbeing and health, residence location, access to amenities. Our contributions is really being, you know, in this dimension of looking at externalities on family members. And because we are talking about, you know, this should be a work life balance, family friendly policy, it makes sense to look at sort of other family members. So in terms of the possible channels, again, so this is gonna be difficult actually to get great data on this. We hope to make some progress on that today. We don't have much actually to say on this. I'm gonna do a little bit of speculation, but you know, there could be several things going on. One is that yes, you might be actually able to sit with your child, read to your child, do sort of direct, you know, tutoring of your child and that might be beneficial, but it would also be that you're just there, right? And so you can make sure that they don't, you know, do silly things and that they may actually be sit down and sitting down and do their homework without necessarily really being present, you know, next to them and reading with them, for example. So what do we do? So we look at the impacts of, you know, working from home arrangements on children's educational performance. So again, the sort of this idea of looking at at human capital formation. Now the identification strategy that we have, I think is quite interesting and quite, you know, different from sort of using the pandemic is to rely on variation in working from home provisions in collective labor agreements in the Netherlands. And so I'm gonna tell you more about, you know, what, what the strategy is. But what we think is really nice about this experiment is we are really going to look at a variation from people being like mostly fully remote to move to a situation where they might actually be in the sort of hybrid setup with like one day at home that they, they spend instead of being in the office. And so we think, you know, when we look at the impact on children, we think there are lots of challenges with associated with using the pandemic as an experiment because as we know, schools were closed, we talk about, you know, long lasting impact of, of the pandemic potentially for children's educational development. And so we are a little bit, we, we think it's kind of nice to be able to actually look at this pre pandemic period. Again, I don't want to raise say this pre pandemic, but having in mind that the Netherlands was not a country where, you know, nothing was happening. And so there were, you know, there was some some sort of use of of Teleworking before. Before, okay, so context, the Netherlands where, so the teleworking provision are actually sort of taking place in collective labor agreements. They are different types of collective labor agreements in the Netherlands. Some are, most actually a lot of them are at the sector level and that will include, you know, a lot of, a lot of firms, but some of them are actually at the firm level and these are usually, you know, larger firms. We actually like to focus on those because when you focus on the sector level, you have to have, you know, the variation is gonna affect a lot of firms that are similar. When we look at firm level, we can actually look at within the same sector firms that look similar, but that actually had either a provision or, or not. And the nice thing about the Netherlands is that it's one of the countries where the data is great, where you can link, you know, firms to employees and then you can actually also match them to information about the children and their educational outcomes in particular. Okay, so the key outcome I'm gonna sort of focus on in, in this this study is the, is called aceto, which is a high stake exam that most children actually take at the end of primary school at the age of 12. And so this, this, this test is really important because it gives, you know, an idea of actually which track you should go into in secondary school and it's really kind of high stake for that reason. It's gonna determine pretty much whether you are able to actually go to university or not. So the score determines, you know, is a, is a guide of whether you should actually go to a, a specific track and then the teacher will actually, usually based on that score, but also sort of private information they might have, will make a recommendation. Again, none of them is strictly binding so parents can still decide to actually do something else, but the correlation is very strong between the CTO score and what the children end up doing. And just to show you a little bit, sort of more detail. So in the Netherlands there's a sort of primary education that goes on until the age of 12 and then you have basically three possible tracks for secondary education. So they track quite early on. And it is really quite important because if you want to attend university for example, you have to be in the higher track and it's called pre university education. The middle track, you can still do some higher education, it's called higher vocational education. And then the other one, the lower tracks, the, the type of education that you can do afterwards is not, is not gonna be sort of higher education. So the period is 2006 to 2019, which seems like an other era now, but the collective labor agreements, we actually collected information that was gathered by a private company called, called called Expert hr. And that was a lot of work actually to codify all these labor agreements. Then we matched this with employee employee data by provided by Statistics Netherlands. We have the CO test scores, again from Statistics Netherlands. And then we also use a labor force survey because unfortunately, even though the data is great, there is no data on occupation in the admin data and there is also not a direct measure of teleworking. So we use the labor force survey to have information on hours and other sort of other things. My identification strategy, I know it's late, so I'm going for a visual, you know, representation. So we're gonna have some firms that at some point in blue, so means they adopt a collective agreement that formalizes, you know, the, the ability to work from home, we're gonna match them. So we're gonna match these firms a year before the collective labor agreement is implemented with other firms that look similar. And again, we do a strict matching on sector and year and then we use a closest mahan nobis distance on a, a whole bunch of variables that we think will matter or are likely to matter for. How interesting, you know, teleworking might be things like, you know, share of highly educated workers, share of female workers, share of part-time female workers and so on. And then we zoom in these firms and we try to find the parents. So we find the parents who have at least one year of tenure because we want to avoid, you know, having people who self-select into these firms and who have a child between the age of eight and 18. So we have repeated observations obviously on the parental income, so we can do a sort of traditional definitive for them for the children, obviously each child is only taking the CTO test once. So what we have to do, and what I think is quite nice is that we can compare the children who were basically too old to be affected by the, the change in provisions to the children who are actually young enough to be able to, to to be affected. So the ones who are younger at the time of the reforms are gonna be or sort of treated, you know, group while those who are actually too old are, are gonna be untreated in the treated firms. Okay? So this is, this is a strategy that we want to, to implement. We end up with a sample of 28 firms and and 86 control firms. I think we actually, you know, increase this sample in the coming months because I think we found a way to actually get it more at more firms, but at the moment we have 28 firms and 86 control firms. I don't want, and this is important I think especially relative to all the other papers that have been presented. And we are gonna look at, you know, the effect on on also labor outcomes for parents as well. This is not a representative sample of the population. This is gonna be a sample of firms that are quite large. And here I'm showing you a bunch of variables where we look at the mean of the firms that are not at all in our sample, either as treatment or control. And then the difference between, you know, the ones who are in our sample and those who aren't. And so in general we see that these are firms that have less, you know, women, they have less part-time women, less part-time men. The, the, the, the people in these firms tend to be higher educated, they are larger firms and, and they are earning more. Okay? So we are looking at positively selected firms. Now we, we actually, when we look at the parents and the children is the same story. So we also find a positive selection of, you know, parents and children relative to the, the firms who are not in in our sample. So the, they are higher earnings, they work more and then the children are also not doing better on all these, these tests. This is pre, you know, pre any, any change in in CLA. Now the internal validity, obviously we match over a number of variables. So these variables we do see that they are similar between our control and treatment. So this is showing you the mean for the controls and then the difference between the TR treatment and and the control. And, but what is maybe also interesting is that despite the fact that we don't match on variables characterizing, you know, the parents at the individual level, we do find still a, you know, very good balancing. So in in we have a number of characteristics for parents but also for the children we just, I'm just looking here gender and age and number of siblings. But so we are, what we have to think is that we are managing within our sample we have very good internal validity. Now I just want to show you this event study for first for the parents. So this is a sort of classic definitive because we have repeated outcomes for the parents and we look at whether, you know, they are staying in the same firm, whether they participate in the labor, labor market earnings and work part, part-time. We actually don't see much and this is something, you know, we've been looking at, you know, various ways of looking at the data. It's always quite, you know, similar. We actually don't see much happening to these, to these parents. Now for the children. So I'm showing you here first the event study where we look at, you know, that score for the math, you know, outcome in the CTO exam we look at Dutch, we look at the teacher advice or whether in particular they recommend the university track and then we look at the CTO score and whether they are actually eligible for either the top two tiers or where they are eligible for re leader the university track. And so on the X axis what's important to realize this is this, this is really the ones on the right are the ones who are gonna be treated. So we're act actually taking her her their ceto test after the CLA has been changed. So one year, two years, three years, four years and so on. And what we see, what you can see is that we do see some evidence of a positive impact on these test scores. And particularly the margin where this seems to be happening is actually on the margin from the lowest tier to the middle. The middle tier. I'm gonna show you now sort of regressions now where we can control for a lot more things. So again, it is like a different D defect except that the post here is really looking at the age of the child, whether they're young or or old. So the first two columns are looking at the Z store score. Again, I'm just reporting the sort of, you know, D definitive coefficient when we have no controls, controls sector fixed effect and then a matching fixed effect. And what you see is actually quite large effects of around 0.1 standard deviation in increase in test score. These are actually quite large relative to what we know from literature and other education interventions. We don't find a direct impact on the university track recommendations. But we do actually see again, so this effect that we saw in the event study that this seems to be really at this moving people into the sort of middle tier from the lower lower tier by about five percentage points. So they're more likely to go there and you can see that also controlling for things even having, you know, this sort of matching fixed effects, we actually still find relatively robust estimates. Now the heterogeneity is interesting here. You might expect, okay, so if I, I can't look at occupation as I as I said, but I can look at you know, above median wage, below median wage, high educated, low educated. We also look at whether this, this, the change was actually early on in the period or not. We find all the effects we find go in the direction you, you will expect, but we actually find no significant differences between the groups. So it's not the case that I can really tell you it's only, you know, for the ones who are high educated above median wage and so on other heterogeneities by you know, gender of the parent, I would think maybe mother girls senses and father boys may, may, may maybe more sort of have more of an effect. We actually see again the effects that go in the direction you would expect. But no the significant differences across any of these groups. So coming back to the mechanisms, which we think we can speculate a little bit, but it seems to me that the story doesn't seem to be that the parent is really sitting with their child and doing homework but maybe might actually be, it goes more in the direction of maybe they're present rather than actually doing a lot of direct, extra direct tutoring. I want to talk a little bit about the placebo analysis because of course this collective labor, you know, agreements change not only one thing at a time often and there could be other things going on. So we look at other provisions, particularly provisions that give more flexibility to employees but that are not directly related to or are not increasing the amount of time that parents may be able to spend at home and we don't find anything. So whether whether they can take, you know, a leave or taking care of another family member, whether they can split the parental leave. So again, this is holding the number of months constant we don't find anything and then we also looked at whether, well what if we assume that the change was three years earlier, can we actually see anything? And we don't see anything this is in in panel D. Okay, I just want to finish by looking at the impact on teleworking. So this is coming from the LFS, so much smaller sample here on the labor force survey. But we do see that in our sample of parents in these firms that are part of our, you know, sample, we do see quite a significant increase in the probability of teleworking by 17 percentage points. We don't see the, our work is actually increasing but it's not static statistically significant. And then we interact basically in the blue box. I'm looking at only treated firms to not show you a terrible table, it's already a quite complicated table but we just want to look at, you know, before and after and interact this before and after with a bunch of variables. And as you would expect maybe we see that it's particularly dis effect of increase in teleworking is particularly pronounced upon people with children and not so much actually women or educated and so on. Okay, let me finish. So we see evidence of positive impact of work from home provisions on children, no strong evidence of hetero heterogeneous effects. And we also see that the labor outcomes, things like hours earnings and all the things we look at for parents remain unchanged. Yeah, that's it. Thank you.

- Yeah, can I just ask that methodologically, it was great, it was really interesting. So I understand it. This is just having the agreement so this is intention to treat.

- Yeah, exactly. Yes, yes. So that's why the LFS is quite useful to actually look at what happens to the probability of teleworking, whether they actually telework and also we were worried,

- Wait, what the LFSI thought was labor force supply,

- No sorry, labor force survey. Okay, so this is actually where the only place where we have information on

- Telework. So the coefficient on the, you should then scale up I guess the treatment effect by that if you wanted to look at the direct of just magnitudes

- Yeah, we, yes, but then the effect looks really large and that's why we are a little bit concerned about the doing this. Also the LFS here is asking about, you know, whether people are have done, you know, teleworking in the last week. And so it's not exactly, we don't have no, I can't really show you what you would like to see, which I, which which I think is really looking at the parents and how much they increase their teleworking over the last year before the TTO for example and so on. So we want to be a little bit careful, but it's true that in principle the effects are even larger. Yeah.

- Got it. And then the other quick thing is you mentioned it looks like it's supervision. When I looked at your results, the one coefficient that looked particularly big was educated parents on math, which if anything would be, I mean I see it around here that like, you know, parents who've got us some kind of maths undergrad degree tend to help their kids out more and yeah, that it's, yeah, it's the math with above medium wage, doesn't that suggest there's also some direct teaching?

- Yeah, no, I, I actually think we can't exclude that there is a difference, but I don't know, it's interesting to see how people look at this evidence. I i we, yeah, we don't at this point we can't reject that the two are the same.

- I would just be curious on the supervision point, whether you could also think about like some quantile estimates or something of like, are you really pulling up the bottom and the test score distribution or something like that?

- Yes, yes, yes. We have to do more on this because I think it's, it's indeed interesting to, as I said, we, we think we are actually affecting people in sort of the middle, the sort of middle margin and we should, yeah, we can do more on, on this. Thank you Chris.

- Those effect sizes are huge. Yes. If you sort of look at anything in the education literature. So I guess I had two questions related to that. One is, do you think that telework allows people to move to different or better schools rather than just having parental involvement, which would be kind of consistent with one of the papers earlier today. And then the, the second sort of unrelated point to that is in the moving to opportunity literature, it's all about sort of the exposure time. And so looking at kids who are eight only gives you like two or three years of pretest data. But I wonder if effects maybe accumulate from earlier in the life cycle if from age four or something you have a parent at home that is kind of keeping a kid on task.

- Yeah, so the, the no, the great point on the first point we have actually looked into moving and especially we we're curious whether the the, when you have sometimes two siblings, whether the older and the younger sibling are in different schools, which would also indicative that they really changed schools and we don't see any evidence of that. So we don't have very strong evidence going in that direction. And then the second point was

- Younger kids.

- Yeah, so looking at younger kids, yeah, so we are a little bit bound by the period we have actually, so the data that we have, so we have to have them reaching the age of the CTO and being affected by the CLA change. But so, but you know, I I take your point of course, you know it's younger kids are, are likely to benefit as well.

- And the same issue as Nick's first question. It, it seems like these, these effects are actually much larger than what you reported, but can you tell us a little bit more about what's actually in these bar, these agreements? So are they the option to work from home one day a week, five days a week? Are you strongly encouraged? Is there a wage incentive? How does it work?

- Yeah, so the, the collective labor agreements are quite, it's gonna be quite generic. So it's basically formalizing, you know, the, the option of actually being able to work from home, but the actual, you know, practicalities and how, how many hours and how often and so on, this is going to be very often left to sort of, you know, the employer to, to negotiate that with their, their employees. So there is really this option. So you this formalized in the, the agreement but it's not like it's written, oh you can work, you know, three days a week.

- So it sounds like it's sort of a soft intervention.

- Yeah, yeah, yeah.

- And still you're finding what looks like a very big effect.

- Yeah, but I think it's also a shift, like it's really, I think really a signal, you know, to the employees that now it is acceptable to actually work from home. And so that could actually shift quite significantly behavior. Yeah.

- This is just a question about the educational system is this zero sum, right? So if my kid gets into university, does that mean that another kid can't go to university?

- No. I mean, so there are constraints at some point, but no, it's not at all sort of a competition in that sense. No, there are not a limited number of spots in the, in the top tier. Yeah. Great question.

- So are you concerned that there might be self-selection of parents that feel that their kids really need help with school, which would be consistent also with your results on stronger effects for those kids that are moving from the lowest career to the middle and university career,

- You are concerned that they are

- Like a self-selection of parents into working from home when they feel like that their kids really need some help with their

- No, of course, of course. But that's why, I mean the results that we report our intention to treat and so we are not looking at only the ones who actually change their practices. Yeah.

- So I guess this, I guess this means on the margin holding our wages constantly spending a little bit more time with our kids. My, my question's on, as we go back further in time, remote work looks different from, that's off, sorry. Okay. As we go back further in time, remote work looks different than what it is today, particularly in the early part of the 2000. And so my understanding late implementation means it's happening like 2010 to 19. Is that what that is here? So the effects seem bigger there and I wonder if, could it be something like the firm is giving you, you get a better computer at home because you can work from home and then the kids are using that computer. Could it be something, could, could that have something to do with it? And is there anything about technology in these collective bargaining agreements that you could use or the different types of industries or occupations if they might be using a computer or telework means they're doing it on the phone or something like that?

- Yeah, I've, that's super interesting point. I, I really don't see much of a difference between the late and, and early, but I think it's certainly an interesting point of, you know, how the technology changed and so on. So again, I think there was no, the internet was around for a, since like the end of the nineties, right? So it's not as if that meant, you know, no computer when you were working from home in the early two thousands. But yeah, I, I think there could be something more interesting going on, but at this point I don't see really a difference between the early and late. So I don't want to make too much of that.

- Yeah. Or maybe interactions with broadband. I know people have used that and some of this literature.

- Yeah, yeah. Yeah. That's, yeah. Good. Thank you. I think I'm out of time. Thank you.

- We have a half hour break and we'll be back at four 15.

- All right. Thank you so much for having our paper on the program. This is joint work with Alan Kwan, who's also in the audience and math is and Rick Johnson. And the paper is pretty self-explanatory. We show remote work sponsor entrepreneurship. So the paper was originally motivated by these two important macro phenomenons since the pandemic, the rise of remote work and then the surge in new from entry. So we're interested if there is a link between these two macro phenomenon and we're gonna test this at the micro level. Best study whether and how remote work increases workers transition from wage employment to entrepreneurship, something we call entrepreneur spawning. Now I don't care, obviously entrepreneurship is important, but one key factor is that the majority of entrepreneurs actually come from prior wage employment and therefore frictions within wage employment can impact talent flows to entrepreneurship and ultimately growth and innovation. There's also a lot of debate as we see in the papers today, that remote work might inhibit innovation and idea generation, especially through collaboration. And therefore as we continue to evaluate remote work policies, it's important to consider this POR effect through entrepreneurial spawning. Now to empirically test our question, we need a good kind of firm level measure of remote work cover a wide set of firms we need to be able to observe spawning ideally at the individual level and finally have some identification. Now all we're gonna do our paper is that we're gonna leverage this novel firm level measured remote work that's kind of created by Michael the Allens companion paper that we measure this basically using employees internet activity data to construct kind of firm level measured remote work at different frequencies. We're gonna observe spawning from LinkedIn and for our empirical strategy we're gonna start with a crosssectional analysis, basically comparing the spawning shear of firms that have higher versus lower level remote work during the pandemic and and with rich set of controls. To get at some identification, we're gonna use an instrument variable approach. Our primary instrument is firm's pre pandemic average employee commute distance and we also use alternative instrument which is county's local business closure orders issued during the pandemic. And then we're gonna move to dynamic firm panel analysis, kind of like an event study, compare firm's annual spawning share how that changes before and after the pandemic across firms with high versus low tendency to adopt remote work. For this analysis, we're also gonna fix the composition of employees, only track those that were at the firm before the pandemic and check the same individuals over time to address potential selection or employee recomposition effect. So lemme preview our main findings. We'll find firms with high level of remote work during the pandemic are more likely to see their employees subsequently living to start a new business. And this effect holds both unconditionally as was conditioning on job turnover. So basically when we focus on people who have left their pre pandemic employers and we find that there's still a disproportionate effect going to entrepreneurship relative to other labor market destinations such as other wage employers or unemployment kind of suggesting this is not just a general turnover or separation effect but there's something unique about shifting to entrepreneurship. And we find that the marginalized bond businesses tend to be of higher quality than the average new firm in the economy. And then we dig into the mechanisms and we're gonna rule out that this is not caused by employee selection change in people's preferences or forced turnover such as layoff. The mechanism most consistent with evidence is that remote work provides the time and downside protection needed for experiment with a business idea. So basically first as we're seeing you have, you don't have to commute, you don't have to groom in the morning, you might be more productive so you have this extra time to experiment with the business idea at the same time because there's less employee monitoring. So you can explore some a side project without being discovered by your bosses and without risking your current job. So you're better able to use your wage job as a fallback option when you explore a entrepreneur idea. And in the end of paper we also kinda confirm the similar results hold at the aggregate level when you look at industry or counties using census entry outcomes. And we also did back of envelope calibration and show that at least 11% of the post pandemic increase from entry can be attributed to spawning from remote work. So let me dive into the data and our measure. Our data comes from this data partner that essentially operates a consortium of thousands of media publishers and we observe about 20% of all the IPV four activities happening on the internet. We have about 1 billion observations per day. So in our data we can observe a user and their employer the IP address, what website they visited, as well as the time span stamp. And we can also infer the location from the IP and the data provider Gonna link these anonymous users to their employers through a variety of methods such as their IP address, whether they used a work email to log in, say watch the journal as well as third party identity resolution data. On our end we are gonna classify the IP addresses they're coming from into four categories, business or office, residential, mobile or VPN. And we're gonna consider the latter three as remote, but we're gonna show even if we're just focusing on residential versus business dropping VPN and mobile, we have similar results. So essentially we're able to collapse this to the firm year or firm month level. Basically measure the fraction of employee internet traffic coming from this remote IP addresses and we're gonna measure this during work hours Monday to Friday, nine to 5:00 PM And this guy shows you how this average firm level remote work measure evolved around the 10 pandemic. We see a big jump at the start of the pandemic, it kind of persisted into 22. There's less of a reversal relative to kinda the survey based measure because we tend to capture these knowledge workers that's pretty active online and we did a lot to validate this measure. For example, first we showed that when you measure this remote traffic during non-work hours for example, we can nights, you don't see a jump at the start of the pandemic and it also correlates strongly with individual's mobility pattern measured from safe graph data which can be considered as kind of the ground truth for for for this. It also correlates strongly with the remote ability of occupations as well. Firm level measure of remote job postings. Now our outcome spawning, we're gonna measure it from the LinkedIn data individual's job history. So we're gonna define a spawning event as somebody who reports new job with a different firm. The title contains founder, owner, et cetera. And the individual is one of the first initial five employees by job start date. And the job start date is within one year. Affirm founding date. This is just to make sure that we're actually capturing two founders and obviously not everybody's on LinkedIn but we think it does a good job capturing those that sort of at risk or would consider becoming a founder. And our LinkedIn data ends in kind of end of 23, but we're gonna track responding until December 22 to mitigate potential truncation from stale cvs. So lemme start without crosssectional analysis. We're gonna work with both firm level and individual level sample for firm level. We're gonna focus on US employers that has this now missing remote work measure but had at least 10 employees as of February, 2020 right before the pandemic minimum 10 employees. Just to ensure our measure is not too noisy and we're gonna call this FAB 2020 firms and we're gonna call their employees as of that time as FAB 2020 employees as our individual level sample. So this is our individual level specification. The outcome, again this is a cross section analysis outcome, is whether one of these FAB 2020 employees later left their Feb 2020 employer and startup firm between March 20 and December 22. The key right hand side variable is the firm's average remote work level from 2020 to 21 and we're gonna control for a variety of worker and firm characteristics such as industry and location of the firm. And for individuals we wanna control the tenure seniority salary which is inferred from lio based on their role and location as well as the person's past found experience or measured exam. We can throw in additional controls like age, education, job role, et cetera. And the result is similar for firms we're gonna control their size age as of February, 2020, but more importantly also gonna control for the lacked spawning year as well as the remote work level in 2019. So kind of controlling for these past spawning year individuals past found experience as was as pre pandemic remote work level help us absorbing some of these tummy variant unobservable such as firms culture, the type of employees or just generally whether they have telework infrastructure pre pandemic. Our firm level specification essentially collapsed from individual level where some of these indi individual controls become firm average shares is so to isolate some kind of exogenous, more exogenous variation in remote work we're gonna employ this instrument variable. Kind of similar to what Greg showed yesterday is what I use firm level commute distance, which we can measure based on our data as well. And we're gonna measure this in 2019. So basically pre pandemic, what's the average distance commute distance for their employees. The idea is that for two similar firms, when the pandemic hits those whose employees live farther away from the US office can be more incentivized to adopt remote work and for a longer period of time than those whose employees live nearby. And we're gonna normalize this measure within employees location as well as within the office location. So the idea is that kind of conditioning on location and conditioning on pre pandemic remote work level. The remaining variation hopefully it's largely predetermined and idiosyncratic for example might pick up office proximity to public transit or availability of parking spots. We can validate that conditional controls this pre pandemic immune distance actually doesn't vary, kind of does doesn't correlate with pre pandemic spawning rate or workers found experience. And in our dynamic analysis later I'm gonna show that firms sorted by this commute distance also trend similarly in their spawning rate before the pandemic. An alternative instrument is simply county level business closure orders. It should during the pandemic you can think of this as sort of a forced remote work and lot lot of it are kind of in installed due to local pandemic situation. Now lemme show you the key crosssectional results. This is the instrumented results. So at both individual and firm level we see that the pandemic area remote work has a strong effect on employees kind of spawning for entrepreneurship in terms of magnitude, one standard deviation increase in our remote work measure increases firm spawning share by a 0.18 percentage point, which is about 40% of the Ming. Now this is a large percentage effect but it's actually in line with the effects we typically find in the spawning and entrepreneurship literature which just have a very low ming and a large percentage effect. Now we did a lot of additional robustness checks, different specifications in the paper, but I wanna highlight two alternative explanations we considered and grew out. One is that maybe this is just picking up a COVID specific demand shock to say food delivery SE sectors or certain sectors. And we showed that when we drop the top boomer industries during the pandemic, the results also go through in our subsequent dynamic analysis was gonna include industry year fixed effect to absorb some of the sectoral level shocks. And we also show this is not driven by the remote ability of the spawn firm. This is wrote about remote work of a previous employer. We first show that this affects, holds even within workers' location, say two workers working in the same area area but working for different firms. And if that that county has similar occu occupation composition, this is kind of different out. We also show that the newly formed firm are actually equally likely to be in person as remote. So it's not that these newly respond firms tend to be predominantly remote. As I mentioned earlier, we also show that these effects also hosts conditional job turnover. So what we did is that we restrict a subset of employees that left their FAB 2020 employer for various destinations and within the subsample look at the effect on spawning and we still find an effect. So this kind of suggesting that remote work disproportionately shift these employees towards entrepreneurship rather relative to other destinations like unemployment or other wage employers. And this exercise also helps to roll out potential tation buyers from LinkedIn because this is conditioning on observing a job update and it also helps roll out some of this demand based explanations which should affect job reallocation in general across sectors. And next we're gonna move to our our dynamic event study and the this is gonna be a firm year level where the outcome is the fraction of employees that spend and become entrepreneur each year. And we're essentially gonna do very simple definitive compare before NAFTA 2020 across firms with different adoption of remote work, either based on the realized increase of remote work or one of our instrument that proxy for different tendency to adopt remote work. In this analysis we're also gonna control for the interaction between this post 2020 dummy and a variety of ex anti fromm characteristics to observe some of these potential time variant unobservables. As I said earlier in this analysis we're also gonna use a fixed employee sample. So rather than tracking all the employees over the years, we are only gonna focus on the FAB 2020 employees and track the same set of people over time regardless whether they still work with the FAB 2020 firms. So basically we're gonna use their spawning from other firms, their future or past spawnings to kind of difference out their underlying entrepreneurial attendance And this kind of try to hold the composition of employees fixed to make sure that this is not because of a recomposition effect, say firms increased remote work level and then attract more entrepreneurial employees. So we can roll that out. We're also gonna use the specification to look at not just the spawning, the transition, but also whether you're an entrepreneur in a particular year, which is gonna capture sort of how long you stay in entrepreneurship in addition to the transition. So here are the, the graphs here. The treatment is pre pandemic commute, distance left, the outcome is Spanish share on the right is the founder share whether you are a founder in a particular year. So as you can see there is a kind of steadily increase effect for the founders here, which is, you can think of it as accumulating this transition effect but also suggests that these bondings are not transitory. These people do not quickly revert back to wage employment. And in terms of magnitude, one standard division increase in the commute distance increases bonding rate by 9% and founder rate by 6%. So in the remaining time lemme talk a little bit about the mechanisms. So first thing we consider is potential preference change. For example, as you spend more time working from home, you may start to appreciate time with family or develop this preference for flexibility or preference for quiet life. If this is the case, then this should predict the spa business should be kind of this low growth, more kind of hobby based self-employment that's compatible with this preference for flexibility. So we're gonna speak to that looking at the quality of these marginalized spawn firms. And we're gonna measure quality with either initial employment or whether they subsequently received VC financing. And what we find is that these firms, if anything, are actually higher quality than the average new firm we see on LinkedIn. So this is kind of helps speak against this preference for flexibility story. Now another potential explanation is that maybe this picks up forced entrepreneurship in the sense that maybe high remote work firms also had more layoff and these employees subsequently become entrepreneurs out of necessity because they were laid off. I think in the data is actually those that adopt more remote work actually adapted better to the pandemic and actually performed better. But regardless, we're gonna try to roll this out by focusing on the set of firms that experienced continued employment growth every year over those, those 3, 5, 3 to four years. And these firms essentially unlikely to have had mass layoffs and therefore what we observe there is unlikely to be driven by layoff of forced entrepreneurship. Now the mechanism that's most consistent with our evidence is this experimentation channel. And this is kind of best summarized in this quote from Vox. It says that while people have always worked nights and weekends to start their own business remote work, give them the time and flexibility to do so and a better hedge against failure. So as you can see there are two elements here. One is this extra time because saving on commute time because you might get more productive but more importantly you have more flexibility with your time. You can just finish your job in the morning and have this big chunk of time in the afternoon to explore your site project. The second element is this downside protection, which is basically related to less employer monitoring. So before the remote work it's hard to just do something on the site at work without being discovered. And sometimes for to take the jump you might have to quit your wage job to experiment, to build a new business. Now you can do both at the same time without risking your career. So this means you can better use your wage job as the backstop when you explore a new business idea. If it's works well you can take the plunge. If it doesn't work you always have your wage job. So we're gonna try to speak to both of these channels to speak to this downside protection. We're gonna look at the risk of the industry that you enter when you launch a new business. So the idea is that this downside protection, this option value should be higher when you're entering into riskier industries rather than safer industries. And we indeed find a stronger responding response into those industries with higher yum from failure risk relative to those safer industries. So that kind of suggests that this downside protection, this option value is playing a role here to speak to the time channel, we're gonna kind of lean on this series of papers presented about childcare and we're gonna look at basically whether the local K 12 schools are mainly doing in person versus virtual learning during our sample period. And we find the response is stronger for these working parents for for areas where the this K 12 school are mainly in person. And interestingly we find this heterogeneity when we look at kind of working age parents, but outside of that like really older young ones, we don't find such a difference. Kind of suggesting that this time also plays a role. Basically when you have a child doing virtual school and home, you would have less slack time to explore entrepreneurship. So in the last part of paper we kind of replicate this finding at the aggregate level using census data. This is BDS. So if these are employer firms, we can do this either at the county or at the industry level. And our treatment is gonna be the remote ability of the jobs in that area industry. And we find a similar positive effect for these remote, more remote ones. But here a key assumption is that kind of this spawning tends to be happening in the same industry as the employer industry. But in in the data we do verify the majority of spawn in the same industry. So lemme wrap up. So here we show that remote work facilitates workers transition to entrepreneurship. The mechanisms that remote work provides the time and downside protection to experiment with business idea and the marginalized bond businesses actually higher quality than the average new firm. We did a back of envelope calculation, at least 11% of the post pandemic increase in new firm entry can be attributed to remote work spawning. We want to caveat that we can't speak directly to allocated efficiency, but if we think remote work relaxes constraints in exploring your outside value in entrepreneurship, it could potentially lead to better allocation of our human capital. Right? That's all I have and look forward to the comments and questions.

- Yeah, really interesting. So I'm wondering whether one mechanism could also not be like you, you have more time to think really about maybe your big ID going for your dream and so on. And I know again there's another paper in JHR that was published last year about the effect of maternity leave on self-employment. And again it seems to be that there's some evidence that women who take maternity leave are more likely to go into self-employment. So it goes a little bit in the same direction.

- Yeah, that that maternity leave is my other paper. So it's similar mechanism that there is about job protected maternity leave. You have this downside protection as well. For the ideal question that's possible, it could be that you already have an idea now you can better able to explore without career risk. It could also be that the time gives you time to brainstorm new ideas and to the extent kind of conditional entry we do observe these tend to be higher quality ones, both can be happening. Yeah, you could use the time to find idea as well. Yes. Steve,

- A couple clarifying questions. In the spawning of new businesses statistic exactly what criteria define a new business?

- Oh, how do we define spawning?

- Yeah,

- I had some

- Questions. So what, what's your definition of a new business in the spawning measure? 'cause you, you had other measures that I understand what they are but what's that one?

- So that's basically you report, you're a founder of a new firm and the firm is kind of

- In a survey on

- LinkedIn. On LinkedIn, sorry, on LinkedIn. This is Observ on

- LinkedIn. Yes. Okay. But then, so we don't really know there whether it's an employee business or not

- Or do you, what

- Do you mean? They say I'm a founder.

- Oh I see. - I founded my, my consulting firm last week and I am the one person who works at that firm for example.

- So yes, we're also gonna look at whether there other employees associated with the business. So one of the quality measure is whether you have other employees, whether employer

- Okay, but in your, okay I but in the core results is that, is that

- In the core result it could be the only person that's possible.

- Okay. Yeah. Alright, great. So that, that's one question. And then in the 11 when you, you this 11.6% statistic you have there, what's the denominator in that statistic? What's the measure of a business of the new business in that statistic?

- So let's just look at all the new firms created by

- What measure is it The Census Bureau one you showed at the Outsets. Okay. But that, that, that is a mix of several things, including, including not mostly it's not employee businesses but then it's broken down to those that are likely to become employee businesses down the road. Right? So that that part I, I like all this stuff but you need to be a little more precise about

- Yeah, that part we use the one that's likely employers

- Likely s Okay. Okay. Can

- I ask another, just another I I thought it was super interesting. I was a clarifying question. Just how you measure remote is if I'm at Stanford, I have their laptop and I'm using their laptop at home. 'cause that's how you can tell I'm a Stanford employee. But you can also tell the location is not on Stanford campus. Is that correct?

- That's right.

- Okay. And so you, so if you see a, a company's laptop using somewhere far away where you know they don't, you can confirm they don't have an office and also the person's not traveling for business or something.

- So traveling it's possible. But if you just kind of log in from home, then we can capture that. So we might have a bit of management error with like consulting jobs where you just travel all the time.

- And on Steve sing on the LinkedIn, couldn't you look when they say they're a founder of, you know, I dunno, blogs company. You can also look where there are other employees at blogs company on LinkedIn at least.

- Yes

- You do. Out of interest you do that. I mean, again, I I'm fine with what you do. I'm just interested as to, you could also look at whether they found a business with other people on link. So the reason I ask is one question is people that are working from home may spend more time online, including on LinkedIn and they kind of tend their social media presence.

- So

- Are they posting more on X, are they more on LinkedIn, are they more Instagram? And if I'm on LinkedIn, I'm, I guess the question would be you could control for that. Do they for example, put down more education? Like, 'cause that historically education shouldn't be affected. But if I'm just on LinkedIn all the time and I'm deciding to polish up my, oh, I

- See

- My whatever homepage, then I may also fill out my educational details. So that would be another thing just is the more detail on LinkedIn you just wanna control for they're not just actively on LinkedIn for hours every day and like updating it. 'cause LinkedIn definitely did very well from the pandemic as an aside, they saw much more activity. But that's a great paper, so

- That's good point. Well the the census result kind of helps alleviate some

- No, no, the census result says there's, there's definitely an explosion entrepreneurship and John Holter has stuff showing us out in the suburbs, which is everything you've, your stuff is consistent and we've talked about it. We, we, we don't have any papers on it. I just noticed the same fact. And I think your story is, is the,

- Is

- A good story. I, I believe it. I would just, you want to for a journal I guess make sure it's

- Not right. Conditional results might speak a little bit against that. So that's kind of looking at, you already have an update, it's just whether you report as a founder of whether or just working for a new firm. So, but you might just view report as founders as more positive signals, you're more likely to select report that than alternative wage employment. So so I need to think more about this selective reporting. Yes.

- Can you say more about the industry composition of Disney? Can you say more about the, oh, sorry. More about the industry composition of these new businesses in the sense that because you know, you, you can't follow these people, you know their previous job. Yeah. Whether they basically found these businesses in the same industry or they basically kind of have new ideas. Yes. And they basically fund this,

- A lot of them actually go into different industry. So we just look at knight to digital level. Only one fourth are in the same industry, others are in different industries.

- And how do you explain this is,

- So part of it is just kind of you explore your outside value in entrepreneurship might, which might not necessarily be in the same industry. It could also be that those that eventually entered basically sort of a bad matches for the original employer or industry that's, that's not the possibility or the innovation they wanna do is not best down in-house. Yeah.

- All right. Hi everyone. Before I start, I would like to thank the organizers for an amazing conference. Thank you. So let me talk about how the networks among employees in the firm affect the quality of innovation and what that has to do with remote work. And this is joint work with Mike Gibbs from Chicago and Fred Rican, Manuel from the University of sx. So I don't think I have to motivate why innovation is important. I think we all get it. What I might have to motivate is why we're interested in a network perspective on the topic. And one view there is that innovation is often a recombination that is, it's not just something new that has never been seen before, but rather there are two existing concepts or ideas. And if you put them together, it's, it's kind of something novel. And if you take that view then it makes sense that if you put different employees of different backgrounds, of different skills together, then they may be, may be able to come up with something that they wouldn't be able to come up with on their own. And hence the network structure in a form, in a firm or in an organization would matter. And it matters who you bump into, who you talk to, what kind of innovation comes out. So that's the idea here. And we are then therefore interested in how does an innovator's position in the firm's internal network affect the quality of ideas or innovation? And in particular, we want to ask, do more collaborators lead to better ideas and more collaborators in the network? Terminology means higher degree. Is it the case that if you connect different groups of employees or people, is that beneficial to the, to the one who's connecting those individually? And is this beneficial to the organization? So in the network terminology, that's called bridging. And then the reason why I'm here, how do these networks in the way the networks change under remote work and work from home? Alright, we're touching upon lots of strands of literature. I'd say there's at least three there. The first one is the network literature, which there's a lot of economists nowadays doing network stuff, but I think it actually started more with the sociologists who had a bit of a headstart there. Especially when it comes to networks and innovation. Ron Bird is a big figure in that literature. We're also, of course, touching upon and contributing to the remote work literature, which you all know very well. And then the innovation literature, and of course one of the organizers here is a major contributor to that one. So very wide net that we, that we have here. Let me talk a bit about the setting and the data. So this is data from a single firm, it's called HCL Technologies, which is a very large global IT firm, which is, has a very heavy focus on innovation. So this is reflective in the culture, reflected in the, in the processes there and also reflected in how they manage innovation. You may not have heard of this firm because they're more business to business rather than business to consumer, but they're pretty large. And I've seen the confidential client list. So a lot of firms that you know are on that client list. So the data that we got is basically the database behind the firm's formal employee idea, suggestion system. And they call it the value portal because you know, innovation creates value. So the idea is that any innovation that any employee has in the firm, if they want to get that implemented, they have to sort of submit it. I mean they have to write it up, they submit it, and then there's some sort of a review process and some senior managers look at it and either they, you know, they get an r and r and they need to revise the idea or straight acceptance that actually happens. Or you know, rejection. If the idea is in infeasible, that could happen too. This innovation is rewarded in the sense that if you get the idea accepted, you get a cash reward. I like this data for many reasons. A lot of innovation research from the psychology angle comes from these more artificial creativity or lab tasks, which are nice in the sense that you can get data very easily, but you know, they are artificial. Whereas here we see kind of high stakes innovations in an actual firm innovations that are actually used in the market or in the firm. So that's very nice. And how do I actually, and the thing is that we observe the universe of all the ideas submitted over a span of 5.5 years and that spans both work from the office period as well as a work from home and a hybrid period. And the network will come from co-authorship in these ideas. So if we observe all ideas and all the co-author lists on those ideas, then we can recreate the entire network who is innovating with whom. So that is very nice and I think it's important in, in a network study that you don't omit a few nodes or a few employees and then you're sort of missing crucial links where information could travel through. So we see where really all those links, so the data includes all the ideas that were accepted, but also those that were rejected. I think that's crucial because if you look for example at patent data, which is really cool data, you have it for many different countries, many different firms. But a patent is really at the very end of the innovation, right? Many ideas were had within the firm and they didn't pan out or maybe they did pan out, but then you didn't get a patent for it for whatever reason. Here it's really much more granular and small. So we really see any kind of small idea. And of course the employees are incentivized to submit those ideas because they want the cash reward. So that's kind of nice. We avoid that sort of selection bias in that sense. Overall, this is a very large scale. We have 48,000 observations and 28,000 unique innovators. So, you know, it's a big firm and not every employee innovates, but many do. So how do we measure innovation quality? That's usually tough. But I think here we, we have two very good options. The first variable is called idea accepted. And so over a certain period, an employee submits several ideas and every idea is either accepted or rejected. And so we just take the mean over this indicator, so we get an acceptance share, if you will. And so this is a measure of the internal value. And like I said, these are large stakes. So some of these innovations could mean millions for the firm in a positive sense, if it's done and then it realizes value, or maybe in the negative sense, if they don't do it, they forego it or if it's a bad idea and they implement it, then they, they lose millions possibly. So large stakes, and that's why we think the firm has every incentive to get this decision right. Second outcome variable is client approval. So for every idea we observe whether the client, so every employee tends to work for a specific client, whether the client raid the idea three stars or four stars out of a scale of one to four. So if you get a good rating, that's the indicator one, otherwise zero. And again, we take the mean over that. So we get a appro, an approval share. So two possible outcome variables. How do we measure networks? Like I mentioned, every idea will show who co-authored with whom. And then of course you can see who's the co-author of a co-author who's the co-author of a co-author of a co-author and so on. I think the maximum we go is I think seven or eight degrees there. So it can reach pretty far. So that's how we construct the network and we aggregate the data for six months. Oops, that was the wrong one for six months periods because you work from home period last six months. So it lines up. Well that means we observe the same employee in different network positions because that will vary over time, which is kind of useful for us. And we think of this as largely random in the sense that maybe your long-term co-author that you usually innovate with is, is on assignment elsewhere for that period. So you can't really coauthor with that one. Or maybe they're busy with their other duties or maybe you are busy with, with this a temporary assignment or that random water cooler moment that you have where some visitor you meet and there's innovation, maybe that happens one period but not the other. So there's a lot of variation over time and, and it's sort of in the estimation we're going to use that. Networks are very, very high dimensional. There's many, many sort of network measures that people have come up with. We are going to focus on three. The first one is very straightforward degree. That's just a number of co-authors that you have in a, in a period. So that's a very simple measure of connectivity. It's obvious to look at that. The second one is the network size. So that's the total number of people in the network. So in other words, all the people you can reach as a co-author, as a co-author of a co-author and so on. At the same time, that could be multiple distinct networks in the firm because they're just not linked by anyone collaborating. So that's measures the, the overall reach of that network, the overall size. And then the third concept is maybe a bit more complex and that is called bridge centrality. So we have a formal definition in the paper, but I'm going to tell you graphically what that represents. So it measures the extent to which a single person connects otherwise disconnected employees or groups. So if we look at this graph here, oops, this is an actual network from our data. every.is an employee and every line is a co-authorship. So what you can see here is that there tends to be two clusters, right? There's the, the right group here and these guys, they tend to co-author a lot amongst each other. And then there's the left group. They also tend to collaborate amongst each other, but then there's really only one link sort of between them, right? And so graphically you can kind of see, you know, we have the right island, we have the left island, and there's literally a bridge between them, right? And so those employees with a high bridge centrality, that will be those guys here at the bridge, they, they sort of, they form the bridge formally a lot of, if you pick two random employees, a lot of the links will go through them. So that is bridge centrality, you connect otherwise disconnected groups. The empirical approach, we're going to use linear regression with employee fixed effects, making use of the aforementioned variation in network position. Over time, we're going to be interested in the effect of these three network measures. So degree, bridge, centrality, and network size. We're also going to control for long-term time trends, we're going to assume that these are linear so that we, that term here, and of course that could be seasonal shifts, it, it lines up, well we have six month period, so it is the summer period and the winter period. So we control for that via a dummy. So effectively what we do is compare the same in different network positions, correcting for any time trend. That's how we're gonna do that. Alright, results a regression, well two regressions, one for each of the two outcome variables of the innovation quality. And what, what we see here is that first the effect of degree is positive, significantly positive. The magnitude lines up for both outcome variables. So if you have an additional co-author in a period, your your acceptance rate will be 2.4 percentage points higher on average. So that's nice and I'm going to give you an interpretation of what that, what that is, what we think that is now. And I'm going to add a few more progressions afterwards that add evidence to that interpretation. So why is degree the effective, degree positive? We think that is a shared effort channel. So in our little model that we have in the paper, the idea quality can increase either because you have more information that comes in through the network and that makes it better. Or it is because people put in a lot of effort to refine the idea, work on it and make it does better. So we think degree is actually the effort channel more pa you know, think of it, you're writing a paper alone, you have to do the data analysis, literature research, you have to write it up, do the data analysis, submit the paper. It's a lot of work. So if you have a couple of co-authors, you can sort of share the workload. But if it was just a splitting of of effort, it doesn't really make the, the, the paper better. What makes it better is that because you have more people there, your, your endogenous optimal effort provision means everyone does less than they would if they were alone. But in the sum, the idea just gets more effort and that is what makes the idea of the paper better. So that is like we think what is happening here and I'm going to add some more evidence to that. Second, the effect of bridge centrality that is significantly negative also for both outcome variables. Those who are familiar with this sociology literature will be probably surprised because Ron Bird and so on, they're actually emphasizing that, wow, bridging centrality is really good. People who do that, they're good for the firm, they're individually successful, these bridgers, they get promoted. Why would they get promoted if they're not doing well? One of the reasons for the discrepancies, of course we're running a regression where we control for degree, whereas these sociology studies, they don't actually do that. They're just showing more bridge centrality better, but they're not actually holding degree constant. We do that here. So that's one of the reasons. So how do we interpret this negative effect of bridge centrality holding the number of co-authors constant. We think there's a cost, a communication cost, a translation costs when you bridge, 'cause bridging involves talking to people of different backgrounds, different, you know, the marketing guy talking to the IT guy, they just don't speak the same language. Or I'm in the economics department and I wanna write a interdisciplinary grant with the sociologists and then I go to these sociologists and all they wanna do is talk about how bad neoliberalism is, but I just wanna talk about the grant. So there's a lot of issue. Whereas if I just do that with economists, everything flows. So there's something here where it's just, there's higher communication costs when you talk to people of different groups. So that would explain the negative effect there. Network size is, there's no significant effect there that would suggest that basically if there's information flow through the network, it's not traveling very far or very thickly. Otherwise, if it was, then we would, we would see that a bigger information, bigger network size would have a positive effect. Okay, now let's think a bit more about this bridging business, previous sociology literature and so on has shown, okay, bridging is is kind of good for the organization. So we want to pick that up to, so far we've only looked at the individual effect to the bridger and that is costly because it takes a lot of time to talk to these sociologists. What we're doing now is we're going to identify, we're going to define two dummy variables. Someone who is in the top 10% of all degree in the degree statistics. So someone who has a lot of co-authors and the top 10, a top having a top 10 person in the bridging coefficient. Okay? So someone who's, who's really good at, at bridging across different groups and the dummy denotes whether I have someone like that in my network. Okay? And if I have someone who, who has a lot of co-authors in my network, that doesn't really help me. Okay? So my, my accepted idea share or my client approval doesn't really go up just because there's someone now in the network who has a lot of co-authors doesn't help me. But if there's someone who has a high bridge centrality in the network who's making information flow through the network, suddenly that actually helps me. So that is this evidence that bridging is not just a cost, it actually does make information move. At least that's what this would suggest here. And in terms of magnitude, we can see that having this high bridging individual in the network is in terms of magnitudes comparable to having an additional coauthor yourself. So that's helpful. We're also interested in the nature of the effect of degree and bridge centrality. Is this, is it long? Is it long lived? Long lasting? So to that effect, we're going to add the lagged degree to the regression. And we see that my success today, so my, my innovation quality today doesn't, isn't affected by how many co-authors I had last year. Doesn't matter. So that again, reinforces our interpretation that this is a shared effort kind of effect because the effort that mattered for my innovations last year, but it doesn't really matter for the innovations today. Whereas if the additional co-authors, there was some sort of information sharing among co-authors, I would expect that the information sticks also helps me this year. But since we don't see an effect there that speaks against this sort of channel, that is what brought us to think that this is probably the, the effort channel in terms of bridge centrality. If I have a lacked bridge centrality here, so if I was bridging last year, how does that affect my innovation quality this year? The effect is no longer negative and that in a way makes sense because the cost of talking to these sociologists and so on, that was last year. The cost is born, it's paid, that's done. So it's no longer bad for my, my innovation quality this year. So that is sort of all consistent with this idea that bridging because of this externality effect makes information flow, but also that there is a cost and that cost is short lift. Okay? What now happens during remote work? So limitations, this is not a field experiment we would like it to be, but unfortunately it isn't. What happened was the firm was working from the office, pandemic hit, everybody was sent home, was mandated by, by law I guess. And later then people were coming back on a hybrid basis. So what we're doing here is we compare the same employee during the work from the office and during work from home and same work during the office and in, in, in, in hybrid, while we control for time turns. And what we can then look at is how do these network structures change because of remote work and work from home and hybrid. So work from home, there's, there's basically no effect. So networks are roughly the same based on these three measures compared to when people work from the office. So that's okay, work from home, no significant effect there. However, the hybrid effect is unfortunately somewhat negative and it is negative on all three dimensions that we look at here. The network size decreases by 1.4 people in the network. The average network I think during the work from the office period has only four people because there's a lot of networks that just have one guy in there, namely the single author. But then of course there's a lot of networks that are really taught really big. So the decrease of 1.4 is kind of large in magnitude compared to the base rate of four degree also deteriorates. So if I'm in in hybrid work, I tend to have a third of a co-author less and bridge centrality. Similarly going down indirectly by implication it would mean that because degree tended to be good for, for my own innovation quality. And because bridging is good for the organization as a whole, because of the externality, it would suggest that innovation is not as good during hybrid compared to work from the office. We're not estimating this directly in this paper here. We have done this before in another paper with a slightly different sample. You can find that here. So what we're doing here is just documenting how does, how do the innovator networks change during remote work? Alright, all right, let me wrap up. So we have very unique and I, we think high quality innovation data that is usually not available to study how innovator networks affect the innovation quality. We see that direct collaboration. So the number of co-authors is, has a positive effect for the innovation quality. The effect is immediate is however short term. So that likely reflects the shared effort effect. If I have more co-authors, more effort goes into the idea bridging across groups, there's a trade off for the individual. Individually it's costly but it's good for the group. So there's some management implication there that organizations might want to incentivize this sort of valuable bridging. And finally, the hybrid work arrangements unfortunately look not that good for the innovator networks. And that's it. Thank you.

- This is very cool work. Thanks so much. I guess I have a couple of clarifying questions. One is, why did you use the mean acceptance rate and the client approval rate? It would suggest that there's like a very high cost of having crappy ideas, whereas I guess in my mental model it's okay to have a crappy idea as long as someone puts the kibosh on it before we spend a lot of time on it and we really just care about having three really good ideas. Who cares about the seven really bad ideas?

- That is a good point. So I guess you, you would've proposed something like the sum, let's just count the number of good ideas and and then it doesn't matter if I have a few failures. Okay, so that's, yeah

- I guess I just don't know how costly it is to have submitted an idea, right? Did I have to work on it for six months and that and therefore there is real opportunity cost for that? Yeah. Or it's, this is like a, you know, three day affairs.

- No, good point. Yeah. Okay. I mean I guess deliberately rephrased it as the quality of innovation. So in that sense it makes sense. But I guess you're right in terms of from the firm perspective, you don't necessarily only care about the quality, you care about a, where a weighted sum of innovation that comes out. So that's a, that's a good point.

- Yeah. I guess one other thing that maybe I missed but I didn't quite understand how they were doing hybrid is that that sort of my whole team comes in on the same day or it's a, you get to come in on any day that you want and you could imagine that one of those would have a bigger impact on who my network is than, than another.

- Yeah, good point. So this was immediately following the work from home lockdown. It, it wasn't as organized. There was no strict coordination as to everybody's in on on Wednesday or something like that. And you're right, I mean if they had managed a bit better, presumably it, it would look a bit better. So in that sense it was I think very, let's say fair.

- Did it translate into then acceptance rates? 'cause you showed us about your three network features, but I, maybe I missed the actual acceptance rates. So

- We have, we have another paper where we look at this. There is one sort of little issue there, which is that review takes time and hybrid came last. So the review rate among in hybrid months was lower than in other months. And controlling for that is a bit tricky. So there's this big caveat kind of, but it looked like the effect is either neutral or negative for, for hybrid on.

- Thank you. Yeah. Can you go back to slide 14?

- Sure.

- So just wanna understand exactly how you're defining work from home. So is March, 2020 in the work from home period?

- No. So that would've been, so it would would be April, 2020 20 to including September, 2020.

- Okay. But that, but then in April, 2020, as I understand your measure, it's constructed based on interactions in the last six months. So the April, 2020 period.

- No, no, no. So that would be all the interactions from April, 2000, 2020 to September, 2020.

- Okay. So it's a forward looking measure of

- The - Window is when you define the network measures. Yeah, exactly. I'm sitting in April, 2020 and you're defining these network me measures over a six month period.

- Yeah.

- And you're looking forward then,

- I mean I wouldn't call it forward because I'm not saying it's the April, 2020 period. It's, it's the window from April, 2000, 2020 to September, 2020. I see. So it's it's the six month period. Yeah, in the summer of 2020.

- Okay. So just, I wanna make sure, so in the, by the time we get, so if I take observations for July, 2020 or September, 2020, those are looking over a, so, so, so it's not a six month period, it's just this work from home period. I'm just trying to understand how you're mapping the definition of network structure to the time period of the observation on innovation outcome.

- Okay, so you, you take the six month period starting and including April, 2020, I look at all the ideas that were submitted during that time, that tells me who co-authored with whom. And based on that we sort of recreate the network. And in that case that would be the work from home period and the networks that were

- Got active

- At that time. Okay. Got it. Okay, thank you. Yep, yep. Chris. Okay. Or we go? Yep, yep. Go

- Ahead. This is great. Thank you. Thank you so much. It's very interesting. So I was wondering whether you know anything about the team leader and manager characteristics? Because I can think of a scenario, you know, COVID happens, you don't know how to deal with the situation. So your attribution of hybrids basically reducing the innovation network and it can be driven just by bad management. If you have team leader, probably you want to fix, effect them out. Okay. And then second thing, why don't you control for tenure in the firm? Because as I spend more time in tenure in the firm program having more ideas. So I think this is something that you should account for

- Good points. So the thing here is that we got the data from the innovation portal, which did not include any HR variables. All we see there, this is the idea, we have some idea characteristics, but all we see is this is the idea of the co-author. We know nothing else about the co-author except he was co-author on that idea and, and whatever other idea. So we cannot do, we don't know who was the manager. We don't know how long they've been in the firm. Should we have gone out and got the HR data? Maybe, but it's probably impossible because that would've been all HR data from the entire firm. The good thing about this project is that we have all ideas. We have the complete network that is way more worth, I think way more than having a 10 year variable for a couple, but then having to cut off the, the networks. So that's why we do this. If you're interested in, in HR variables in the other paper, we do have that. And that is why it's a smaller sample. Kristas.

- So, so yeah, this is such a unique setting. One thing that we did in our brack paper that that and Kyle had played a really critical role over was parsing the text of all the emails. And so we created these measures of novelty. I'm wondering whether you might be able to parse the text of the projects and to even maybe create a proxy for, I don't know, organizational practices that gets at some of what TVA was mentioning. Even if you can't do that, maybe in the beginning of the presentation you just walk through what some of these example project proposal ideas, like what are people actually suggesting? But the language and how they're describing could get at some of these measures of workplace perceptions. And that could be otherwise omitted variables. Thanks.

- Yeah, good point. So in the paper we have some examples of, of sort of titles of of, of ideas what they are about. It's like we even have the categories, you know, it's about cost cutting, it's about process improvement, it's about product improvement, you know, so different categories of the ideas, but we don't have the full text of the idea 'cause that's kind of proprietary. So. Alright, thank you so much.

- Hi everyone, first of all, many thanks to who were colleagues for organizing this and entire organizing team. This is great. So before I start, let me tell you what we do in this paper. So this is going to be a randomized control trial in a call center arm of a large multinational company. This is a, this is a company, they basically provide business process outsourcing to many other firms, large banks and telecommunication companies. And the setting is the following. So before COVID, everyone was working from the offices with the first national lockdown in March, 2020, everyone shifted fully remote work and as of today about 80% of the company work remotely. It has some advantages but it also has some disadvantages in the following for the following reason because company is struggling with managing those remote, remote teams that basically motivated this, this paper. And then the study, so this is a joint paper with Nick and Steve, they're here, Victoria, att, BRD, and Gemma at the OECD and Paris School of Economics. So this specific remote management problem is not specific to this particular Turkish firm, but this is also widely discuss in the media but also it's discussed in the policy circles. So if you basically get the UK parliament discussions, there's quite a bit of discussion around how to manage teams, hybrid teams and also fully remote teams. Why this is important, we know that, I mean Nick did a lot of work on this. Management practices, good management practices, it is important for productivity and for example, Gallup estimates that a potential about 9.6 trillion loss in global productivity loss in global productivity due to low employee engagement and poor management. So, and then the key challenge of course as I mentioned, the managing fully remote teams to maintaining, to maintain engagement, performance and retention. And this is what we do try to do in this paper. The question we ask is the following can occasional coordinated in-person contact, boost engagement, performance and retention in fully remote teams. The intervention will be, we basically bring people to the office once a month for nine months and then we track their productivity, service, quality, retention and then we, through surveys we basically try to understand to what extent did their communication and feedback structures improve over time. The nice thing about this project, we've worked with a call center data and the advantage is that you get a very precise measure of productivity and service quality as well as retention. And that's what we will basically leverage in this, in this project I will show you some results on main outcomes and then some of the, the mechanism analysis is in progress. So this is, I will be, I will basically mainly look at productivity and service quality and retention today and some evidence on communication and team cohesion. So lemme tell you the, the, the headline results. So first one, first one is the following treated agents, agents who came to the office once a month, they became more productive over time. So the, I will show you the, the just descriptive graphs and you will see that effects basically kicking after four months. The obviously then you can think like to what extent this productivity gains come at the expense of service quality, right? Maybe they're becoming more productive but they're just doing sloppy jobs and we have two measures of service quality and we find null effects on those measures. Retention was higher among those who attended the office. They stay longer on average nine days. And at endline, when we do, when we the, when we analyze the endline survey, what we find is that treated agents reported greater team cohesion and they're, they were more likely to say they receive feedback. So the, I won't talk much about this, but we know very little about management practices in hybrid and fully remote settings and our paper provides one of the first evidence using the setting where we have ized service context. Lemme tell you a bit about our company. So this is tempo, one of the largest BPO companies or firms in Turkey. We will basically work with a call center arm of this, of this company. They have about 3,500 call center agents. They have seven offices in seven different provinces and then the HQ is located in Istanbul. What they do, they provide several services, but the main business is about call center services to banks, mobile phone operators and s visa section. Some projects require in-person attendance like banks for example. But then for example, we will work with the telecommunication company. We will basically look at Vodafone inbound calls. That requires that, that doesn't require in any in-person presence. So if they are basically working fully remotely. In terms of the workforce structure, each team has about 20 agents. There are 175 team leaders in the whole company and there are super eight supervisors basically managing the team leaders. I will give you some details about the agents and how they work because they, they, they will become important for our results. Agents basically work five eight hour shifts per week and each shift include two 50 minutes break and 30 minute lunch break. Teams basically share the same work schedule with their team leaders and they don't get to choose their shifts, they're centrally allocated by the hr. How this works, you have a problem with your Vodafone billing, you call the call center and then your call basically goes to first available agent agents do not get to choose the, you know, the complexity of calls, but even then at the agent level we know that at in for every single month the call composition that they basically answer. So we have detailed data on that as well. All agents receive national minimum wage. There is no bonus scheme. The only kind of motivation for them to perform better in their jobs is basically becoming team leaders then their salary goes up by about 50%. So let me start stylized facts and why we basically decided is like one once a month office treatments before we basically started this intervention, we basically did a company-wide survey and what you see is that nearly 40% of temp post staff, they want to come into the office at least once a week. So there is some demands regarding coming back to the office. But then if you basically, so on the left hand side what I show you is the, our survey with tempo on the right hand side, what I show you is basically the current working from home levels from sway. So this company located in Turkey and their workers' preferences basically match with sway levels. So it's not, it doesn't look like, and if you look at the batch, it's, it's pretty close. So it's not like this company is somewhat like outlier in that regard. So people want some office presence. So this is basically one reason why we basically wanted to bring people back to the offices. Second one, we basically randomized all this to basically minimize any type of bias, but then we basically ask everyone in the company, do any of the factors below negatively impact your productivity when working from home? And then if you basically look at these orange bars, what you see is that a meaningful share of agents cite limited, limited manager support difficulty, difficulty in connecting with colleagues and isolation as key challenges. And then we want to come up with an intervention that could be somewhat low cost but then that would potentially, that could potentially address these, these issues that people report. We also talk a lot of people, and this is the second time I do this type of like semi-structured interviews and it turned out to be extremely useful. There are two things, basically the two things came out of these discussions. The first one is people like remote work, they like the flexibility, they like the fact that they are not committing but they feel isolated and they feel disconnected from the company because they never see their colleagues. They only exchange emails or they have WhatsApp groups when they have a problem. They basically exchange messages through WhatsApp. So drawing on all the survey evidence, inwes and consultation with the company. What we did, we basically designed this relatively low cost intervention where we basically brought people one one day a month to their office. The experiment took place in in Chand, Luva, Turkey. It is in the eastern part of Turkey where the company already has existing offices for one of the, the, the banking projects. So the, this infrastructure basically enabled the intervention without additional investment. We basically use the existing offices and as I mentioned, all agents work for Vodafone, inbound call services, customer services. So this is basically where the, the experiment took place and Istanbul is where the headquarters is located. So lemme tell you about the experimental design and recruitment flow. So as you can imagine, HR was worried about about like bringing people randomly into the company. And what we did, we basically informed 661 word ofone inbound employees about the program in ufa in March, 2024, about 476 of them basically volunteer for the experiments and then remaining people they didn't volunteer for, mainly for family and care responsi responsibilities and also long commutes. And we wanted to have two criteria when we basically did the, when we basically decided on the, the final sample, first we wanted everyone to have at least three months of tenure so that they will have some experience in the company. And then the second one is to basically make sure that compliance will be fairly high throughout the treatment. Because the nine month intervention is fairly long, we didn't want basically people to have long with long commutes. So we basically decided, okay, if you have less than 45 minutes based on Google map and then we know everyone's address and then we basically said, okay, you are in our sample. So we, we are left with about four two with 248 waterfront inbound employees. And then based on the company id, those with odd numbers, they went to the treatment group and then others they, they remained fully remote. What we did obviously then another concern you could have is that spillovers across teams. So we basically fixed the teams in June, 2024, we fixed the team leaders. So even today they basically work with the same team leaders throughout the intervention treatment and control teams remain intact. They basically work 9:00 AM to five 5:00 PM shifts or in the treatment group and also control group and all other conditions remain identical. Treatment group just came to the office once a month, control group didn't come to the office at all. That's basically the only difference. We do several checks. The, the, the basically treatment and control groups are pretty well balanced. I can also show you the summary. It's basically exactly the same thing. A couple of things. Prior performance didn't predict the take up of office work. Employees, as I mentioned, who volunteer tend to have shorter commutes. They tend to be single and not to have not have their study study rooms on office days. As I said, people basically share the same schedule and shifts, same technology pay evaluation structure and no overtime was allowed and compliance was very high because the company treated this as a work requirement. So it wasn't something optional. So everyone and company provided shuttles so everyone came. And the compliance according to the data we have is about 95%. So all the cost of this intervention, commute, meals, refreshments, they were fully covered by the company. So there is no cost to employees. And when the people came to the offices, they, they kept working with their tempo BPO laptops. But what they did, they had basically spent longer breakfast breaks, coffee breaks, lunch breaks and so on. The whole point was to basically enhance this, you know, interpersonal relationships within and across teams to see to what extent people basically continue basically kind of utilizing this one-on-one interactions later when they go back home. And we also randomize the seating plan. I won't be able to sitting plans, I won't be able to show you any results on that, but we, we will be also randomized where people will sit and in surveys we basically ask people the names of people they communicated often. So companies sign frequent text reminders to remind, to remind them basically they needed to come to the office. They did a great job basically pushing people to fill in all these surveys that we, we rolled out, this is how they worked before pandemic, this was since the pandemic when they shifted fully remote work. This is during the intervention, one of this, this, so this gentleman is the, is the team leader and then they basically use this company shuttles and, and we know which people basically took the shuttle. So we will also exploit that information. And this is basically during the intervention they kept working as usual, but then they spent plenty of time together. They had coffee breaks, lunch breaks and ex they, they were pretty extended. So lemme show you, lemme tell you more about the data and I will show you the results right away. So we have individual level data, we have production data, number of calls received, we have the duration of calls. We can split this duration of call into talk time, hold time and admin time. We also break time. We have service tool, service quality measures. The one is monthly audit rating. The company is independent audit department. They basically evaluate each 10 random calls a month. And then we also have customer ratings where basically at the end of each call it is voluntary, but there's some, some customers basically give a rating. We have call composition data, we have baseline and deadline surveys and as I mentioned we have the randomized sitting assignments and shelter plans. And an intervention started in August, 2024 and in April, 2025. So the first result we find that agents who attended the office became more productive over time. So what I will show you right now, I will just show you weekly means of number of calls per hour. There are no controls here, I'm just splitting the, the, the weekly means. And what you see is that there are, the first message to, to first observation is that this is a type of job that there is quite a bit of seasonality. If you look at the Ramadan month here, basically the number of calls drop, but then it picks up both for treatment group and then the control group. So if you don't have the IRCT, you would basically, you know, attribute to this productivity gains to to remote work. So what we find is that after about four months FX kick in and then you see this confidence intervals diverging and then treatment group is answering more calls, processing more calls. So monthly office visits leads to 0.7 calls, more calls per hour answered, which compared to the baseline mean it's about 6.5% increase in the last four months. If you basically look at over the, the, the experiment period nine months, but you that basically correspond to 3.3 0.3 additional calls per hour. So this intervention leads to higher performance. That's the first message. Then given that we can, we know the ation and, and we can split ation to to time hold time and admin time, we can also try to try to understand where this effect is coming from. And what I show you here is just the ation and what you see in line with the main result. People are answering calls in a much faster way. So they are becoming more efficient. And this effect is, I, I won't show you the, the the kind of breakdown of this call duration, but what you see is that this is purely driven by reduction in talk time. So people are becoming more efficient in handling calls. As I mentioned, the obvious concern is to what extent this productive in the gains come at the expense of service quality. People are doing sloppy jobs. We find null effects throughout with controls. Without controls there are no effects. So service quality doesn't decline. Then we basically try to understand to what extent there heterogeneity. We look at the results by gender, age, presence of children in marital status and baseline productivity we find no heterogeneity. So these gains are sort of broad based, not driven by any particular group. And I can show you the results with proper events, studies, saturated regressions at the agent day level agent week level agent, month level results are qualitatively the same. Basically third result we have is that retention is higher among those who attended the office. I will show you two results. The first one is just this graph, this is a descriptive graph and what you see is that month office visits lead to seven percentage points higher retention in nine months. Retention in the treatment groups is 94, 90 4% in the control group is 80 87%. That basically corresponds to 50 54% reduction in quits. We can also look at this in terms of the number of days they stayed and I won't be able to click on the the table this hyperlink. But then what we find is that treated agents stayed on average nine working days longer than fully remote peers first. The last part that our last result that I will show you is some results on mechanisms. Again, most of this is in progress, but what we find is that we ask a better way of questions on to what extent people feel a good fit with the company culture, to what extent they receive regular feedback from their managers to what extent team communication are effective and then valued in job and so on. What we find is that at endline monthly office visits basically lead to stronger team communication, manager feedback and cultural fit. We find limited effects on job and life satisfaction. And what we will do, as I mentioned, we will basically exploit this information we have on seating plans and shuttle plans and then we will try to see whether this is driven by people who basically sit next to each other and they, they remain in communication. We do a very simple this, we will improve this, but for now we have a very straightforward cost benefit analysis. This is just based on direct cost. So direct cost including lunch, transport and refreshments opt. This experiment over the course of nine months is about $10,000. Turnover, savings and productivity gains. Basically turnover, savings corresponds to about $6,000, productivity gains about to $12,000. And then what we find is that each dollar spent generated about $1.80 in measurable benefits. This is again based on the direct cost that we can observe so far. So what are papers suggested occasional in person contact matters for productivity and retention in fully remote teams and and survey evidence suggests that those attended to office, they report better communication and more cohesion in the company. And thank you very much.

- This is very cool work. Thanks. I guess I had one quick question that I don't understand quite the implementation of it. It sounded as though there were longer breaks and longer lunch period for people to bond, but would that have a spillover effect on the people who are at home having to handle more calls while I'm in the office socializing or was there some sort of offsetting something

- But so we have 128 agents, they come, so we had four teams, two teams came to the office and then we have other teams still working from home. So I guess even if there is some effect on the control group, that will be very limited. Okay. Does that answer to your question or you have a

- No, that was exactly, that was like if I'm working at home am I all of a sudden having a, a greater burden and so maybe I'm more stressed or something. So we

- Can, we can analyze the call volume actually,

- Right?

- Yeah we can actually that's a great idea. We should do that. Great. Yes,

- Super in super interesting. Just wanted to ask you, so what is the mechanism that you have in mind? So at the beginning you mentioned the fact that like managers or I, sorry, these, these people were like feeling a little bit isolated. So I'm, I'm thinking maybe you have some question related to health, like self-assessed health, whether like this changed between the two groups. Second like question is also, no, let's just start with this. Yes. If you have some measure.

- So we had a question on self as health in the survey we don't find anything but being on the ground and talking with these people, what we see is that they have been working in the fully remote teams and they never interacted with their team leaders and their team teammates, but all of a sudden they basically meet each each other and they're engaging in conversations. Actually being in the, in the offices, they had a great time basically meeting with their colleagues and it seems that they're learning a lot from each other. And from day one they were basically talking a lot about how they handle calls and how they basically kind of handle hard like tough customers and tough questions and so on. So I guess this basic, that's why we, we don't see any immediate effect but effect kicks in after about three, four months that once they spend more time together, that we, we see these effects,

- We deliberate that kind of looking at each other, how they handle

- Exactly. Yeah. We will be able to be very precise on this once we analyze the seat plans, Chris. Yep.

- See it's ambiguous because we're both go

- For, oh either Chris is fine.

- My question is superhero quick, it's actually more of a comment and an anecdote. Gallup has this work on like having a best friend at work and so you might just find anecdotes around like when you, when you survey people more and just being like, hey, did you get a best friend? And it just might be part of the mechanism.

- Okay.

- Anyways, that was just a quick comment, really amazing execution of a pretty big experiment over to the next Chris. Thank you.

- Yeah, great, great experiment. I'm most interested in the, why is there this positive productivity effect because also that matters the most for the, for the benefit calculation. I probably missed it, maybe you had something there, but the justification sounded a bit like people are doing more calls per hour and then you said the reason for that is the, the

- Full time goes down

- Is lower, which sounded a bit total logical to me. So, so so can you, can you find out why productivity increases more in the treatment group? I, I have maybe two with theories, which is one, somehow it stimulates loyalty in the company because they do a nice thing for me and then I put in more effort somehow. Could that be it or is it because there's some information sharing going on and maybe you can look at that because of the seeding plan and all that. So they, they, they share tricks and then they just get better. So one of the, I would be interested in that. And and another question I have is just, so in the experiment you tested one day per per month, was it? Yep. And you find positive effects, so, so that's good. One does wonder the, what is the optimal design? So if, if you could have run an experiment with, you know, once per week, once per month, twice per month. So, so we, which of those is best?

- I just wanna jump in here. What he didn't say is we have an earlier paper that looks at the shift from traditional office arrangements to fully remote that led to large productivity gains. So, so clearly we know now it looks like the optimum is neither of the the traditional or the fully remote and we are trying to, so the company stuck with a fully remote 'cause it's very good. It works out very well for them in many respects. So we're, we're taking a modest move away from fully remote.

- Okay. Just as a, the reason to do one day a month is if you look at tech firms in the, all the firms I know in the US that claim to be fully remote meet one day a month. So just by reveal preference in the market, Zillow, Upwork, Dropbox, Airbnb, all of them, they all meet one day a month. So I was like, you know, look, they're doing it. They probably GitHub's about the only exception actually. So that seemed to be also a natural setting. We, you're right, we could test other stuff but we just went for kind of market outcome right now.

- And Chris great points we will analyze, we will actually do more detailed analysis of the mechanisms. Yes, sorry. Yes,

- I was the same thing. Frequency. 'cause a lot of frequencies once a quarter is another kind of standard. So I wonder if you had any anecdotal feel or two days a month, one day, you know, kinda where, where do you think it's gonna land?

- We decided to do one once a month because it was both the most practical and low cost intervention because eventually company wants to adopt this.

- Did the people say things? It's been great. I wish we saw each other

- Actually, I I I have some results on the, whether people like the policy at the end line we ask this question and then we, we find very positive results actually people like the policy and they would like to have this like a permanent feature of their work. Yep. We ask, we ask this question in the end

- You should ask them how many days are they still possible.

- We actually asked that question as well. I should, we should look at the data. Yes. Oh,

- So on the same point of Chris, but maybe even a bit more extreme. So are you concerned that it could be that this employees are aware this experiment is going on so they're aware they're going to be compared to those that didn't get the treatment and then they're going to feel like, oh, I should really put effort and really work harder because then I'm going to be compared to the, you know, the control guys that didn't get all of these so they feel compelled. It's not really about getting better.

- So if it was the case, I think you would see the effect immediately. There is no reason for us to wait for them to wait four months systematically to see these effects. And second, the way that we approach this with HR is just we said, okay, this is going to be kind of rolled out over time. And that's the, initially that was the initial intention. But I guess my answer like first answer to you is just if that was the case then effect would have kicked in right away. Yes. I think there's one last question there. I

- Yeah, thanks. Super interesting. I wonder if the right benchmark is not going into the office versus people who are fully remote with bad management practices where they never talk to one another, but instead a virtual intervention where people actually get together with their teams and you could benchmark the face-to-face thing versus going into the office with some of the amenities. I, I don't know how that would look, but that would be sort of an interesting horse race if you could do it. Yep. And then I kind of second some of the mechanism questions. You could kind of imagine that being in person allows you to compare productivity with people where it might not be salient or some of the other things versus the learning thing, which I probably believe very strongly in as a prior, but some of those other sort of social comparisons would also be interesting to see if you can tease out.

- Absolutely. I think we can do this in a very careful way because we know everyone's pre productivity level so we can see whether if once you sit next to someone who's productive, whether that's, that's where, when you, whether you become more productive, the first is a great idea, maybe that's going to be a different paper. Yeah. But great idea. We, we should actually talk to the company and thank you so much.

- Thank you very much for organizing this conference and thank you very much for having me. Today I'm gonna present the productivity externality of remote work with dramatic increase in remote work. There is some concern that more remote work reduce in-person interactions and the society benefit less from onsite productivity spillover. However, to look at whether the IMP to look at the implication of remote work on the social welfare, it's also important to check whether new technology could generate this productivity spillover that benefits all workers. An example of remote productivity spillover is that with more remote workers, one can use remote communication technology more effectively and there could be more innovation about remote technologies and can improve one's productivity. So in this research I explore how large is this remote productivity spillover and what's the implications on social welfare for achieving social optimal level of onsite work in the in percent of both two type of product rollovers. I build a quantitative model to answer these questions. In the model, workers can choose whether and how much to work from home or onsite based on the relative productivity and relative amenity of these two options. And there exists productivity spillover between and within remote and onsite workers. However, when workers make those choices, they will ignore this productivity spillover effects. So this leads to a gap between the social optimal level of work arrangement and the market equilibrium. And market equilibrium is decided by individual choices. So I matched the model to the data and based on the quantification results I find that on the social optimum requires a bit more onsite work but the existence of remote productivity spillover can bring the economy closer to the social optimal level. Then I'll introduce how I model the spillover, how I estimate the spillover and the quantification results. There are several empirical facts laid. The foundation of the model on the first one is there exist variations across city and sectors in terms of remote work both for the employment size and for share of time working from home for hybrid workers as shows in this figure. And also we can observe the positive correlation between wages and employment size for both onsite and fully remote workers. And one of the possibility that drives this pattern is that there may exist positive productivity spillover for both onsite and remote work, which leads to a higher productivity and higher wages. So in the model similar to the literature I assume more remote workers or more on onsite workers leads to a higher onsite or remote productivity and two elasticity govern the strengths of these two spillovers. And what's new to the literature is that I also allow the cross work site spillover, which means more on onsite workers may lead to a higher remote productivity and a parameter go tall governs the strength of this crosswork size spillover. To match this to qu estimate this spillovers, I match the model to the data. I'm using the US survey data CPS and it says from 2022 to 2024. And in the data I can observe employment size for different work mode across different cities and sectors and how much hybrid workers spend their time on site or remotely. There are two parameters and that are new to the literature. One is this remote productivity spillovers. To estimate these parameters I use general method of moment that join estimate several key parameters in the models that here I will introduce the intuition to estimate this parameter. If there is no productive spillover, we'll expect a downward sloping demand curve. However, if there exists positive spillover, we'll have a upward sloping demand curve, meaning more remote workers lead to a higher productivity and a higher wages. And to estimate the slope of this, this demand curve, I'll leverage the shift from the supply side and lay and leverage moment condition where the supply side shifter and the demand side shifter are exogenous with each other. So the algorithm will pick a parameter such that the Covance of these two shifter are exogenous with each other. So when we have Covance of these two shifters closer to zero, it's as if using the supply side shifter to estimate a slope to estimate a slope of a demand curve. And on other new parameter is this gross work size spillovers and this parameter governs the slope of a relative demand curve where on the excesses we have the onsite timeshare and in the Y axis we have the relative productivity. And if the cross work site labor contribute to the productivity the same as within work site LA labor, then how much hybrid workers spend time on site or remotely will not affect the relative productivity of onsite and remote work. So similarly, to identify the slope of this curve, I use a shift from the supply side and impose a similar identification assumption where the shifter from the supply side and demand side are exogenous with each other. And based on this structural estimation results, I find that the remote spillover elasticity is relatively smaller than the on onsite one and the cross work size spillover is relatively small and based on those estimations I solve a social planner problems where social planners choose different employment and share of time working on site for different city and sector to maximize the social welfare and compare the social planner solution with the market equilibrium. And the market equilibrium is what we observe in the data and I find that the result indicates a social planner could increase their welfare by two per percent by encourage hybrid workers to spend 3% more time on onsite work and increase the onsite employment by 2%. And this can be achieved by providing an on onsite subsea equal to around 11% of hybrid workers gross income. And then, and then I run on experiment setting those remote spillover and cross work size spillover to zero. And I find that if there is no remote spillover to achieve similar level of welfare improvement will have more changes and it costs more. And then I also calculate how much productivity change if hybrid workers spend more time onsite remotely, which varies by different CT and sector and calculate the variations across CT and sector comparing the social optimal level of on onsite work and a remote work. So to conclude, in this research I used model to estimate the remote and ons onsite spillover effects. And I find that at the extensive margin which measures how much productivity change, if there is more ons onsite remote workers, the ons onsite spillover is stronger than the remote one. And on the cross work size spillover is relatively small And based on this as measured results, I find that according to the comparing the model prediction and the data, I find that the social planner will prefer a a little bit more onsite work on average, but it varies across different city and sectors. If there's no remote works spillover to achieve similar level of welfare improvement, we'll need larger changes. Thank you very much. I'm looking forward to your comments.

- This is a great topic on this. I'm so happy you're working on it 'cause it's super important. A few things it it, maybe this is in your paper, but it would be helpful to think about the nature of the remote work externality you have in mind. Two come to mind immediately and they're very different in character. One is I avoid the negative externality of colleagues who waste time or so if I go on site. So that's a static type externality. I'm not endorsing that as important. I'm just saying just setting up as a counterpoint to something else that that Nick and I and you just, just cova had a bit of evidence on. We showed that the composition of new US patent filings shifted towards technologies that that support remote interactions in the wake of the pandemic. It's basically the share doubled now that is presumably going to increase the effectiveness of these remote technologies in the economy as a whole. That's a very different kind of externality and I'm, I'm not sure what it is you seek to capture in your model. So maybe you can say a little bit more about the nature of the remote externalities you have in mind.

- Yeah, sure. The nature of remote spillover I have in mind is more related to the innovation effects you just mentioned and I think the decreasing in the negative spillover is also very interesting. The model does not really specifically specify the sources of this spillovers. I have that in in explanation of why the, there is possibility of existence of this remote spillover. I think it's always good to think about those micro foundation.

- I think maybe just building on Steve's point, I don't know if the model, if this would be tractable, but it'd be interesting to try to estimate those externality terms and particularly the remote onsite externality terms like over time to see if the technologies have gotten better. And so now you do seem more like agglomeration across these modalities.

- Oh yeah. In the model the parameter does not vary by time. I think it's al always very interesting to look at the changes of those agglomeration or spillover effects over time and we can, it coincide with the changes of the technology we have today. And I think I can do this by leveraging more data structure, although it's not a dynamic model but I think it's good to estimate this.

- No, but, but, but, but I think so, so one easy way of implementing MS suggestion could be so, so you have one estimation with 2022 to 24 24 data. If you repeat the estimation and, and you're not gonna have exactly the same data but, but like using work from home numbers from say 20 18 20 19, then you can see if you get a similar parameter from these two different estimations and that can be very informative.

- Oh okay. Thank you. Thank you very much.

- So sorry I guess rule on the same but I thought from the model side the other way to do it would we go back to like the classic endogenous growth literature and have learning by doing so if you think of some of these old models, you would have the productivity of onsite work. You know short, a classic shortcut is that the is a function of the amount of onsite work that's happened over history to some parameter. And that parameter is often less than one. So it's not explosive, which is like the early learning by doing, you know, like we saw Philippe, the Aeon and how nope it's, it's that this kind of thing. So you could have one for onsite and one for offsite and when you start in 2020 there's been almost no offsite work historically. So the rate of progress of offsite is going to be much higher because basically no one really remotely worked before 2020. Sure, sure. So the technology is terrible.

- Sure. - So the potential for improvement is much steeper. So it may actually be you want to have you, you get the social planner would say we should have more remote work initially to encourage the development of technology. Okay, so zoom gets better and the

- Yeah,

- AB gets better. Yeah. And then because we've had so much improvement of onsite in person stuff,

- Yeah

- But that, that gets a, you have two extra state variables there but that would be the one standard way of doing in the model.

- Okay. Thank you very much for those suggestions.

- That's kind of, that's really helpful what we have in the, the patent paper. It's the same idea

- That's it's what we appeal to.

- Yeah, yeah, exactly. Just it's, it is, it's like

- We appeal to,

- It's the whole learning by doing like going all the way back to the liberty ships papers, there's a whole literature and learning a massive literature that would be very much like this. Your productivity is a function of past activity.

- Thank you. Thank you very much.

- Yes, thank you.

- Good morning everyone. Thank you for organizers. I think it's a great opportunity. My name is IL I I got my PhD from Universal of Minnesota this summer at join there people online you might wanna talk and joined Penn State as a visiting custom professor here The title of of my paper remote work for implications for Optimal Income Taxation. A fraction of work from home is big enough to ate its implication over tax base and as we know like there have been productivity of work from home was was a hard topic and thanks to next contribution we even have information before pandemic. But there's a group of researchers saying the world for one could be associated with some productivity loss. I say that if they are correct and those effects somehow present. I claim this could matter for policy analysis specifically for optimal income taxation. And regarding that, I raised the following question, how does work from affect optimal income taxation? To answer it, I write a static general core model, diagnos workplace choice. In this environment we have discre choice like either work from home or on site. I nstitute two dimensional hetero for amenity value of work from home and the skill of the workers. And finally I solve other problem with Ramsay approach using a a TIC income tax function. So idea in this paper not excluding the those many nice flavors either for work from home or other channels. I want to keep, I meant to keep it simple as much as possible to get a comparable result with the traditional method. That's the part of the contribution. We have a continuum of workers indexed with the data and Kai data stands for skill. Kai stands for the multi value of work from home run from independently from some non distribution. I consider separate form of utility in addition to consumption. And ours workers choose their workplace which is indicated with R and we at red mark we see Kai of R represents the Emory value of your workplace that is normalized to on site. So ca kai on site is zero otherwise Kai has mentioned. So we have single consumption good here that's produced by representative firm with a linear technology and workers allowed to work choose their workplace. There is no any friction if they work from home there is some productivity loss with some non-technology and they pay tax on their labor income. I consider this parametric form of the tax functional. I will mention more about later we a, a nice feature of additive form of the amenity value brings this nice result which is very simple. What intuitive, so optimal workplace choice follows the threshold property. What is that an individual work from home if and only if their T are large enough. And plot simply shows that in the x axis we have skill and in the Y axis we have corresponding threshold. And as we see the curve is decreasing. The area of the curve represents the fraction of work from home for corresponding skill level. And what's the takeaway here is that, so we have a selection to hire the skill guys to work from home, which is the parallel in the data. So what like if we that much simple ingredients, if we think why does work from home matter for optimum for tax policy. So the mechanism allow us here, higher tax progressivity can lead higher work from home. And similarly higher tax progressivity mitigates the distribution of workplace by across workplace. That means the distribution could be more even and and what is the re what should be the result? And I claim tax progressivity on labor income should be lower with work problem and why the elastic of taxable income is higher with work from home environment compared to no work from home. One is an example, everyone is working on site. This is a key reference point in the optimal income taxation literature. What is that? The concept like simply measure how responsible of the behavior of taxpayers for a little change in the tax policy. The map could be confusing. I want to tell the intuition like so we have value of workplace that's not taxed. I'm not saying anything normative. That's a fact. And we have positive selection work from home and we do lower productivity. All those forces together increase the sensitivity compared to no work from home one and what is next? So we have finance, we need to finance the transfers and government expenditure and using the features of the work from home in calibrated economy with the US data. I solve plan problem with the Ramsay approach where we maximize the utilitarian welfare by choosing the optimal tax process. That style in this form of the tax function that has been a popular like in the recent years, like the due con contribution of the Atlantic Coast and Violante. So what is that? The key parameter here is in the power that's tau, that's the tax progressivity in the traditional form. Marginal tax stands for with the uc in this third bullet point, which is monotone in income and to get to benefit from comparable feature of the environment, I solve the same problem for the no work from environment like we are turning off the workplace choice. And here I wanna summarize the results over the welfare function. We see the tax progressivity agreed in the x axis and welfare work corresponding tax progressivity on the y axis. The blue one is for the work from environment and red one is the no work from home one. So we see that the optimal policy is significant, could be significantly lower. What is the interesting caveat here is that we see that welfare is lower naturally. Like this could be comparing apple and pieces like we, we don't do that because there are two different environments. But an interesting caveat here is that we have some amenity value of workplace and I force it to be non-negative. And what is that The efficiency losses from the, those top guys are dominating the, those all post to welfare in the utility. That's the, that's the interesting result I, I wanna share and one nice feature of the model, like we, we have the tractable measure of work from in environment. So the calibrated data for the progressivity was the, is the 18% that I used. And so we have higher optimal value and corresponding work from home shame is higher. So what is that? As I mentioned, the progressivity can decrease the threshold that means like work from home and leisure could be more attractive for people specifically for the richer. I can, I wanna come this, if if questions are rise sensitivity, I can conclude with the following slide. Work from home can reduce optimal marginal tax on labor income and this reduction is quantitatively significant. It is sensitive in scale since the environment is very simple but robust in the direction. So it is always lower. So if you have just 4% loss, 10% loss, the the the, the result we see in the the first plot is always true. One, one contribution of the paper simplified environment delivers us to compare with results with the standard that that can be updated, the standard approach. And finally, since the time we have time constraint, I, I don't mention here but in the paper I discussed taxation by workplace. It's a nice literature going back to 1970s initiated by ker in the literature. We have a nice application that the status taxation, age based taxation or general based. So I propose a, a taxation by workplace and if it this mechanism, since we are providing one more information set to planner about the distribution of this workers, this leads higher welfare optimal policies and thanks so much

- By the basic result. But you've tied it to controversial claims about productivity of the relative productivity.

- Yeah that's

- From home. That's entirely ential to the paper. And your point as I see it, you could just as well specify a model that allows for remote work to be more or less productive with some substitution between the two possibilities. It's like the functional form that we have in our paper with Jose and Nick on that, that's not the revised version's not yet out there. So there's no reason to tie your, your result to claims about what the productivity, relative productivity is. You can be entirely agnostic on that point. The same thing will still go through.

- Exactly. I agree.

- So I so I

- Selection itself is al already other angle that I'm currently cons. Yeah.

- So I would just encourage you to restructure the model so that it's, it's agnostic on that question about whether there, because it's hetero, the, the literature's all over the place, it depends on the job and so on. That's 0.1 second and less fundamental. But you might wanna think about whether there are other externalities, external effects associated with the choice between work from home and not. And the obvious one that comes to mind is greater load on the transport system when people, everybody works in the office than when they don't. So you need, you know, you either need a, a greater investment in transport infrastructure which is expensive or you're gonna have higher congestion costs. So that, that's kinda like a side issue but there are other externalities that, that you might want to think about in this optimal tax problem.

- Yes. Yeah, yeah. Thank thanks so much

- That thanks. Okay, interesting. So two thoughts. One is on the tax on the workplace, you mean that you would tax your, the amenity value of working from home?

- No, no I don't mean that. So we are having different tax policies if you, you're working from home or on site.

- Okay. - You know, you report your to when you are reporting like I I I work from home or on site.

- Is it okay implicitly you'll get I I was about to say it, it would be a weird thing to do but it's, there are other amenities that are taxed. For example, your taxed for a company car, your, I think some places tax for free lunch healthcare pensions are untaxed up to some level. So it is not out of the ordinary to tax for some non-monetary amenities.

- Yeah.

- So I I agree. I mean it just, it it would be an unusual thing to do if possible. The other thought is, and I I thought went from your title, you're going in a different direction which links to someone that presented last year, Daniel Aall, which is, you could look at here, which is there's an issue about tax avoidance based on working in one place and living somewhere else.

- Yeah.

- So it could also be, you could set up, when you work from home, your tax rate is different 'cause you're living in a different state.

- Yeah. - So in the US there's all these distortions, which is another angle which is kind of interesting as well.

- Yeah.

- Which is a, you know, a, a separate distortion you could pretty easily look at in this machinery.

- Yeah, that that, that's a, the local texts are really good point. In my third year paper actually I was working on this direction and later when I recognize the channel, actually this is more broad so it is a, and IIII thought this is more interesting and I want to this direction, but I have some preliminary work about that the local dispersion, the tax progressivity D across states are really interesting in the United States. Yeah,

- Okay. As much say I was very happy to see Minnesota meet to work from home. Yes.

- No, I was wondering if you can speak a little bit to the fact that you, you had something about the different valuation of the amenity and I'm thinking about you know, the populations that have sort of a, a bigger elasticity of labor supply like women and maybe you know, disabled workers who are actually brought in the labor force because of work from home. And maybe that how that interplays with your

- Well that that's so true. Like as you see here we have some loss and we assume that it is decreasing the output in the direct pay but it is actually not necessarily true. So what you're saying and even for those guys, it is not realistic. So this benchmark is only for the, we can think this is like the corner angle. So whatever we add there, it's gonna decrease these differences and positive angles like that also will contribute that. For instance, I don't change the labor elasticities by workplace there obviously it is the case because some of the papers saying that was defending the lower productivity case, they measure productivity loss in unit in time, but in total output they say output is decreasing a bit, not much. So I definitely, I'm definitely considering those posts like as well as the externalities. It's some of them are in my agenda and on purpose. Thanks so much. That's great idea. I is,

- You could have the stochastic fixed cost of work, everyone draws it and you're gonna then have labor supply effects. I think the labor supply effects are probably bigger than productivity. It may well be the productivity's negative, you can argue about it, but the labor supply is clearly positive and you'd find and that also has welfare effects.

- Yeah, definitely. Many, many nice micro foundations are like the, because

- People pay positive tax. So there's a positive welfare effective work.

- Yeah, yeah. Higher employment for mothers or other many microfund is definitely a matter here. I agree. Okay, thanks so much.

- Okay. Wow. Hi everyone. I'm very glad to be presenting my joint work with my amazing cots, Juan and Isabella from Columbia University. Let's get started. So I don't think I have to argue a lot that gender gaps in labor markets are very large even in developed countries such as the US which is gonna be our focus today. And because of con pioneering work by Claudia Golding, we know that a lot of this is due to the manhood penalty. Now me and my team like to think of this as fundamentally a problem of the household allocation. Let's think for my presentation that we are Gary Becker and that we can think of a child and parenthood as broadly speaking a set of new constraints and challenges imposed on households that heavily incentivized household specialization. And hence you see, even if we may have small comparative advantages at dealing with house hospitalization, you see these big changes in the whole of the economy, right? So see this is a remote work conference. We're gonna be trying to argue that remote work could be a venue to relax this friction between motherhood and work. And how we're gonna differentiate ourselves is by arguing that by looking at this remote work, how it could be affecting spousal decision making and how does it affect the household as a unit. So the idea is that to drive this idea of spousal decision making, we will be looking at how increasing the male spouses availability at home could be affecting the division of childcare. And then by enabling a more equitable division of childcare, we could be seeing at how this could be increasing women labor supply and thus we would be seeing how household specialization could be decreasing overall in the economy. So those are to be questions we're asking whether shocks in remote were options for dual parent households with young children could be improving specifically women labor market outcomes. And does this, is this happening because there is a more equitable division of household tax and very specifically of childcare? How are we gonna go about this? Is that we're gonna be leveraging large variation in available jobs with remote work options across the US using the pandemic shock. And we're gonna be using these shocks at the occupation level specifically to estimate a, an event study framework where we estimate the effect of an increased work from home options from one spouse on the other spouse's labor market outcomes. And just as a previous result we see that spousal shocks will be increasing women labor will be increasing labor, women labor supply. But the spousal shock on the woman has largely no effects on the outcomes that we are representing today on men's labor supply. And we also see that women and male spouse, both spouses are working from home more and we see some patterns of spouses substituting childcare responsibilities. We already suggest that there has been a decrease in household specialization. So what data we're gonna be using are made data set for household article is gonna be the a CS and for the household time use data, we're gonna be using the available data that is in the US which is the Aus. And then we're gonna be a bit of a rebel today and we're not gonna be using the dingle Neiman estimator for the telework ability index. Instead we're gonna be relying on Nick and Steve's job posting data, which computes the monthly share of job openings online that have remote work options at the three digit occupation level. We're gonna be using that as the basis for our shock. Our shock is very simple, simple. We're gonna be computing long run differences in this hour share of job openings with remote work options between the after COVID period in 20 23, 24, and 20 in 2019. To simplify the econometric analysis, we're gonna con transform this discontinued shock into a binary shock where we're gonna classify occupation as treated if the occupation shock is above the median. And then since we're interested in estimating the effect of the spousal work from home shock on women outcomes, we will be classifying a household as treated if the male spouse in that household works in one of our high exposed occupation. As defining this in this bullet, our strategy is very simply we're gonna be using an event study framework and we're gonna be using our work from home shock, which is at the level of the male spouse occupation. And we're gonna be interacting with time dummies in the years surrounding the COVID pandemic dropping using 2019 as the base year. We're gonna be adding fixed effects for state by time at the state, by time level and of course at the husband occupation level and a handful of household level controls such as age of education of the spouses and the number and the number of children. So essentially we'll be comparing the changes in outcomes for households where men work in occupations that saw very large changes in the availability of work from home options in the jobs that offer work from home auctions be with households that were the main working occupations that largely didn't see a lot of changes of which there are many, there's a lot of variation that we can see. We've seen graphs that show this variation in previous presentations. Of course the, the essential assumption is one of parallel trends between these, both these two household types. But even with that we'd like to clarify that even if those assumptions called AR betas are largely capturing two types of effect, which we call the spillover and the matching effect, the spillover is basically any increase in women and supply that is coming exactly from the men being more available at home. But then because as sort matching is a thing, then a shock that increases the probability that a man works from home is also bound to shock the probability that a woman works from home. And so if we see an increase in the women labor supply, it could be coming from the women being more likely to work from home even if the husband doesn't uptake work from home himself. We have a strategy to kind of differentiate the two. I'm gonna talk about, hold that in a second, but having said all of this, let me just show you our main set of results on women labor market outcomes. We see that the spousal shocks are increasing women employment at the extensive and intensive margins. We see an increase of one to two percentage up to two percentage points at the highest level for women. While wages seem to be increasing by some 5%. We also have a little bit of a quasi first stage. This is a reduced form estimation but we have a stage where we see how does the shock on the husband occupation seeing the shock actually increases their uptake of work from home. And we see that it increases a lot with, with this type of shock, it men are doubling maybe up to tripling the amount of work from home or the probability that they'll be working from home using the a CS data. In addition to that, to go more into the intensive margin, we see that the spousal shock is increasing women weekly hours by some half an hour per week using the a CS data and are is reducing the probability that women are working in part-time jobs. And also we to go a little, I mean this fi half an hour a week may look a little bit small but then we also did a similar analysis with the Aus data on how much time women report to be working. We have information even though the atos only has one spouse at a time, we have information by linking it to the CPS on what their spouse is working on. And we see a, a large increase of women being work working five to 10 percentage point a larger fraction of their day towards the end of the, after the pandemic period, which a day has 24 hours, that's around one to 20 additional hours per day according to our estimation. Now we have a lot of tests that we have made to try to sell the point that this is something that has to do with work from home specifically. We don't have to talk about all of this but we'll happy to discuss. But is there one thing that I would like to highlight is just the first bullet point here, like we said we are, are betas are identifying the speed lower and the matching effect. So we wanted to try to distill that and one strategy that we had was to, we got inspiration from Emma's paper on using the college degree major level shocks. Like how does that, how does work from home up the exchange at the college major level with the idea that this kind of shock at the women's college major level, the idea that this kind of shock would be able to capture the probability that the women herself works on home. So when we take that shock and we interacted with our spousal shock, we should be able to separately to break apart a correlation between women's own probability of working at home and her spouses probably of working home, home. And when we estimate this regression, what we see is that both effects matter for sure the direct and the spousal effect and the spillover effect. We still see something similar like around increasing for example the the extensive margin result by some one percentage point. Now let me talk about time use and household specialization results that we find. So the first thing that we find is that this spousal shock is increasing women's work from home uptake, which is to be expected since as we said we are identifying a matching effect. So, and that is bound to be increasing women labor supply because there have been previous words such as MS paper in the US or if you think more broadly you can think about Lisa hot job market paper in India that shows that yes, increasing remote were options for women naturally increases their labor supply. But then if women are more time at home arguably then you would think the same story as always. Then they're also having to spend more time with the child. But what we see in our IT data is that well at the beginning of the pandemic it does increase, we see that. But by 2022 and onwards it doesn't seem to be increasing. If anything it's slightly decreasing a little bit. But what we do see is that the man himself hit by his own shock or, or the men that are working in his high work from home high changing work from home occupations are taking on more childcare specifically while working around an additional one hour of their day that they're working. They seem to be devoted it to taking care of the child on average. And we also find a similar result to what Dimitri presented yesterday, that accordingly house homes are also less likely to be hiring and overall spending less in money in daycare and all of our results when put together, we think that there is some evidence that remote work could be driving down the, the incentives for household specialization in the US by making a more equitable childcare arrange making households more likely to enter more equitable childcare arrangements and increasing women labor supply it. That's about it that we have to present for today. Like we said, we're leveraging large changes in remote world options available to their women's spouses to ascent their own changes in women labor supply, particularly when they have young children. And we do find significant increases in women's labor market supply and we do find that it is the avail that the availability of work from home specifically and changes in childcare arrangements do seem to be enabling these type of changes. Now I'm glad to, I'll be glad to open up the space for questions.

- So again, really great topic, cool paper. I, you know the, so the overall labor supply in the household's going up, so some and something's and they're using less childcare so there's gotta be something else that's offsetting. Now part of it could be there's less time spent on commuting but

- We, we have results on commuting.

- Okay. So

- Yeah, so, but it'd

- Be nice to have an overall picture of your paper of the rebalancing of time use in the household because there's the household's spending more time in work, it's not clear whether they're spending, well it sounds like they're spending more time in childcare as well 'cause they're purchasing less from the outsider but, but there may be efficiencies in co you know, monitoring the kid while you're working so that's less clear. But just some overall picture of the time reallocation,

- That would be amazing. I guess part of the difficulty on that is that by construction the only chooses one member of the household. So it would be amazing if we had the both of them to see. So yeah, we're trying to make a story with what we have of there being some childcare arrangements, but yeah, it's true, it's true. I agree. And that's, that's a bit of a limitation.

- Yeah. So let's do one question and then move on so that,

- Okay. Thank you so much. This is so interesting. I, I think as I think about this topic a little bit more, I think more realistically, more realistically it's now one person shock. The other persons like the household jointly make decisions, right? So I think in the future a promising direction would be to join this model, this joint decision simultaneously rather than using one person's occupation chain to shock the other. Just like little comments on your SAS analysis, because I also use this kind of data, I think it's quite important to include the interaction terms between women and their men's education with time dummy because that will re yeah that will relax some of your parallel trends assumption.

- I have done that and we survived that analysis. I dunno we have more time or, but yeah, we should be modeling that and we're thinking of steps to do that. There's a practical reason why we use this spousal shock as well. We are interested in finding the extensive margin outcome, right? So if I were to use women's own education or, or women's own occupation, then I would be conditioning on a sample that is already working because if they're asking, they're only gonna ask you your occupation if you're working already. So that was part of it. Also a nice little trick to be able to identify the extensive margin outcome. I think Emma also discusses it in her paper.

- Thanks very much.

- Thank you.

- Good morning everyone. My name is Ottavia Aguila. I am a first year PhD student at the University of Michigan. Thank you to the committee for including me in the program. I'm very happy to be here and this is very much a work in progress, so I'm very open to all types of suggestions and feedback. Okay, so there's two motivating facts behind the research project. And the first has to do with this historical rise in disability employment post pandemic that's been driven by rising labor force participation. And the second fact has to do with that the almost entirety of this increase in disability employment can be explained by work from home. So in this paper what I want to ask is how has this rise in work from home among individuals with a disability affected industry productivity? Okay, so to do this, I combine industry level productivity data with the shift share measure that captures differential exposure to this rise in work from home among individuals with a disability. And what I find is suggestive evidence that this rising work from home is associated with year over year productivity declines ranging from 2% in 2021 to 4% in 2024. So for data, I'm using two public use data sets. The first that we all know about is the American community survey using their transportation to to work question, to have information on work from home. And they also have information on disability status. And then second, I'm using the BLS industry productivity data, which is gonna give me information at the annual level on total output output per worker and output per hour. So I'm gonna construct a panel of four digit makes industries from 2016 through 2024. Now for the definition of disability, the a CS has a set of six questions that can be used to identify individuals with a disability and they're shown on screen now and following prior work, I'm going to focus on the set of physical disabilities, which are shown in the first five bullet points. So I'm not going to include a cognitive disability in my definition, which I'll talk a little bit more about in a few slides. So for the identification strategy recall, my goal is to identify how industry exposure to this rise and work from home has affected productivity. So my approach is I'm gonna leverage two sources of variation to create a shift share variable. So this first source of variation is going to be these exogenous shifts or changes in work from home that vary substantially across occupations. And then the second is going to be the fact that occupations vary in their distribution of industry employment. So equation one is showing how I formally construct the shift share variable. So I'm taking the inner product between these changes in work from home at the three digit occupational level and the industry occupation employment shares, which are essentially weights, that's sum to one within industries. And notice the superscript D. So everything in every object is disability specific for workers with a disability. Okay. So just to take an example of what exactly is going on here. We could look at software engineers take the change in work from home between 2019 and 2022 and then weight this change depending on how many software engineers there are, say in the computer systems and design industry. And summing this measure across occupations is gonna give me an industry level measure, which I call work from home shock. Okay? So industries are gonna be more or less exposed to this work from home shock. And it's somewhat intuitive that these high exposure industries include computer systems, design software publishers, and these low exposure industries include child daycare services and restaurants. Now my main identifying assumption here is gonna be related to these shifts and changes in work from home and following the same assumption as bloom doll and Ruth, I'm going to assume that these changes in work from home are driven by supply and demand side factors related to non-disabled workers and not those with a disability. Okay. Now before moving into estimation, I just wanna quickly address three identification challenges. The first, as we all know that there is this overall or universal expansion of remote work during the pandemic. So there could be concern that my shift share measure is capturing this broad expansion of remote work driven by non-disabled workers. So my solution here is I'm gonna residually with respect to non disability work from home. And an alternative approach is I can just control directly for non disability work from home. And then these second and third identification challenges have been flagged in prior work. The first is changes in the composition of the population with a disability. So this is why I'm not including cognitive disability because there is a large inflow of individuals reporting having a cognitive disability post pandemic. So about 1.5 million individuals. This is potentially concerning 'cause these individuals could have more marginal disabilities relative to the rest of the population and they can make that could make them more employable, which could bias my productivity results. So I'm gonna focus on the set of disabilities that don't drastically change pre post pandemic. And then the third concern is with this increase in labor market tightness. So to address this in similar spirit as bloom doll and Ruth, I'm going to control flexibly by including a third order polynomial of non-disability employment at the industry level. Okay? Okay. With all that being said, moving into the relationship between work from home and labor productivity, what I'm going to do is estimate two-way fixed effect event studies. Okay? So just breaking this down on the left hand side, that's gonna be the percent change in total output output per worker and output per hour in a given industry from 2016 through 2024. And then the main coefficient of interest is alpha one, which is gonna be recovered from a set of interactions between my work from home shock and a set of time dummies. So it's gonna tell me the relationship between the work from home shock pre during and post pandemic. And I'm gonna be relying on the pretre assumption. And then everything else is quite standard controlling for unobserved heterogeneity by including year industry and sector by year fixed effects. So again, the source of variation that I'm picking up on is within industry over time. Okay. And then the results are gonna be interpreted as a one standard deviation increase in the work from home shock. Okay, so here are some of the main results. So here I I have three panels, total output, output per worker and output per hour. Okay. So starting with total output, there's no detectable significant detectable consistent effect post-event. Okay? However, once we look at output per worker and output per hour, what we find is a consistent negative effect post pandemic. Okay? So a one standard deviation increase in this shock is leading to a 2% decline in productivity in 2021. And that grows to around 4% in 2024. And so these are the residual results. If instead I decide to control directly for non disability work from home, the results remain consistent, which is to be expected. Now, next, as an additional exercise, what I want to see if there's a specific disability category that's driving this negative effect. So here I'm re estim the regressions, but instead specifying two groups. So having a sensory disability hearing slash vision. And then in navy I have the self-care independent AM ambulatory disabilities. And the takeaway here is that there's not one disability category that's driving the effect. And then the last thing I want to leave you with before concluding is what the results look like if I compare individuals with and without a disability. So in red I have the point estimates for individuals without a disability. And in blue is what we've been seeing up to this point is just for individuals with a disability. And really what we see is that this universal expansion or broad expansion of remote work that's that's mainly driven by individuals with us without a disability leads to productivity gains over time and what's being massed in this aggregate effect. And what I I've pinned down is once we isolate the disability specific channel, we actually pick up on this negative effect post-event. Okay. And I'll, I'll conclude here just so we have time for questions. Thank you.

- Okay, so, so maybe let me start, and I think the elephant in the room is, is what is the mechanism here? Is it, is it a measurement thing where we measure the hours of people who don't show up at the workplace differently? Especially, or, or maybe yeah, maybe especially or maybe not. Especially if they have a disability, it's, yeah,

- Yeah, no, I, I, I completely agree. I think for future work I want to see if I can replicate these results using other data sets and also really think through the underlying mechanisms behind the productivity decline. But right now I don't have a a good answer for you.

- I was guessing one, one thing is that isn't there a supply curve of if you think there's a supply of people with a disability rated by how productive they are, you make it easier to work, you move into that supply. So the marginal person coming in is less good than the

- Yeah, that, that's a good point. That that is, that is one thing I had in mind.

- Well, there's a demand side as well from the employer that they're slightly less matched. So we, I think in our stuff we don't see any effect on wage. 'cause the big question is that supply demand side effect. 'cause the supply side stuff, I think you see wages go down and the demand that you off see wages go flat. So

- Okay, - It's not clear, but you have a similar structure. Okay. Just presumably a distribution of how good people are. And anyone with a college degree never disabled may be working. And

- I agree

- By the way, in the day, typical person in this is like age 57 and has a walking issue.

- So - People have an idea that in the day it's someone in a wheelchair, it's not, it's actually typically an older American that has some vision.

- I see, I see. It

- Probably was working for 30 years, but a back from, but that's assume most of them.

- Okay, that's very helpful. Thank you. Yes.

- Yeah. Along along the same lines of Nick, the first part of Nick's comment, the, the simplest model I can think of that you could use to rationalize this result is one in which the, there's a competitive market, there's compensation for wages per efficiency, unit of labor, disabled people on average provide less efficiency units of labor services per unit time. The non-disabled person. And that market, the wage differential between the disabled and able people within a given occupation is the productivity differential. So then you could take that model, look at the wage differential compute, the implied change in industry level productivity associated with the shift, and you get something remotely like the magnitudes you've backed out from the industry level data. Correct. That's one way to evaluate a particular mechanism. It seems like the most obvious place to start.

- Okay, thank you. That gives me a lot to look forward to and work on. That's a great idea. Thank you.

- That'd be curious to try to use this machinery to also think about industries that are maybe more exposed to the expansion of like women in the workforce. Because there you might think they're a little less sorted on the basis of productivity. So it seems like maybe the effects are, are more ambiguous there.

- Okay, thank you. I think I'm outta time, but I'm happy to chat out there. Thank you.

- So good morning everyone. Thank you for having me. I am Nicole, er PhD candidate from the University of Naples Federico, and today I will be discussing about role of working from home in shaping equity behavior in the Italian labor market. I guess I must not convince you that COVID-19 accelerated the adoption of working from home arrangements. Perhaps what was contrary at that time to our expectation was this kind of persistence in the years immediately after the pandemic of this new way of approaching work. And this plot just to say you, that adoption and persistence involved the Italian labor market as well. In general, what I want to say about work for home is that it does not change only where people now are working, but has this potential to reshape basically workers quitting behavior. And so in other words, it can change how mobile they are into the labor markets. Moreover, by reducing the commuting constraints and the special friction where for home could also change how sensitive quits are to wage. And while this is the case, so basically, I mean traditionally commuting constraints are able to generate friction into the labor markets, which results in acqui elasticity less being less sensitive to wage differential. And so basically when people have to choose whether to leave or quit their job, they must take into account of course wages, but also the commuting lens that they have to face. And so the higher is this commuting lens, the lower will be the quit elasticity, meaning that quits are less sensitive to wage differentials. Of course, work for home creates this unique scenario because as this can in a way reduce, can relax commuting constraints and so can basically expand the workers outside option because now people are no longer geographically constraints to the local labor market. And so in principle can change how these quits are sensitive to wage change to to wages. And so what, what actually I want to ask for this project is whether work for home, first of all can shape in a way workers quitting behavior. And now can this effect translate into changes in workers quit elasticity? Of course, when you're trying to answer this kind of question, you must deal with the endogeneity that arises from the self-selection into work for home adoption. So basically people choose work for home because of some unobservable characteristics that can simultaneously affect their workers, their their their mobility decision. And what I do is basically to rely instead of the actual user work from home on the potential that each occupation have in order to accommodate for remote option. And so I'm going to basically exploit this occupational differences in work from home feasibility in a difference in difference setting in order to identify the causes effect of work from home on workers, quit behavior, voluntary quit and quit elasticity. So as I said, I rely on this potential for work for home, so I'm going to use this potential for work for home in order to define my difference in difference setting treatment treated and and control groups. But then I also have a look at the actual use of work for home, just for the descriptive purpose in order to see the, the incidents of the actual user work from home across the Italian sectors and occupation, but importantly in order in a way to validate the tele workability measures. So I want basically that the telework ability classification is able to predict the actual use over from home after COVID-19. And then I rely, I draw data on an Italian administrative data set that basically allows me to, to track job, to job separation. But importantly I can observe why this kind of separation happens. So I can see whether this job to job separation I do because of a, a voluntary resignation or quits, which is kind crucial in order because this is my main variable of interest in this setting. So first of all, as I said, I want to validate in a way this tele workability measure. So I want to see whether tele workable actual means the real adoption after COVID-19 of work for home. So I'm going to run this specification where on the left side I have these dummies, which marks whether an individualized going to declare to be in work for home adoption in occupational attempting. And then on the right side I have these tele workability measure interacts with the ears dummies and while controlling for occupation sectors and time fixed the effect, basically the, the coefficient that is going to be estimate is going to, to tell basically how much tele workable jobs are likely to actually adopt work for home after COVID-19. And it seems to be the case. So that basically we observe that there is this sharp increase in the actual use, in real use of work for home among tele occupation who are defined and classified as tele workable ones. So once I validate this, these, let's say this, this, these, these, this measure, I use it in order to see what is the effect of telework ability on the workers quitting behavior. And so on the left side here I have dummy, which marks whether an individualized going to voluntarily leave his employer at time t and then on the right side have again this telework ability measure, interact by ears for each dummy and then a control for individual occupation, region, sector, and times fixed effect And so the coefficient that is going to be estimate here is going to tell me how much those telework workable occupation are, are likely to voluntarily leave their employee, their employer. What I found is that after COVID-19, it seems that those who are inte workable occupation overall are less likely to voluntarily quit and so are less likely to voluntarily leave their employer. And these results, I mean, is lying with the idea that now people are happy with this new kind of flexibility. And so when employer is providing you this work for home option, the probability of staying with your current employer desire. So that's why we observe that those who are in those kind of telework occupation actually are more, are less likely to voluntarily leave their, their, the, their, the, their employer. And so in a way, or for home is a tool that can increase the attachment that you have now with your employer. But of course this diagnose this does not mean that when change, when the wage are changing, then these results is the same. So in a way this means that, I mean, people in tele workable occupation now can also find a new job in other labor market and so can become more le more sensitive to wage differential. So in order to see whether, you know, tele those telework occupation actually are becoming more sensitive to wage differential. I estimate these, this quick elasticity and they estimate discre elasticity model modeling by the, the probability of having a separation as an instantaneous separation rate that can potentially happen in a given interval of time. And is it is a function of some controls, observable controls. And the, the feature of this kind of model is that when you use the, when implemented through a complementary log log model, basically the coefficient beta that is going to be estimate can give you the direct estimation of the quit elasticity. In general, higher quits are associated with lower, with lower wage. But what we observe is that after COVID-19 seems that people, their workers are becoming less sensitive to wage differential. But in general, this decline is less pronounced for those who are in tele workable occupation. So suggesting that actually those inte workable occupation are remain more wage sensitive after COVID-19 relative to those in unworkable ones. So overall, what I found is that those unworkable occupation are less likely to voluntarily leave their employer. However, after COVID-19, there is a decline in the quit elasticity, but this decline is less pronounced for those in tele workable occupation. So in a way, work from home actually allow, allows increase, let's say the attachment with your, with your current employer. But however, when the wage change, it seems that those in tele workable occupation are more likely actually to, to are more sensitive to these wage changes. And that's because basically also an employer can offer you this kind of nor wage amenity. And so you are basically becoming more sensitive to wage differential in a way work for home may have protected workers for these, let's say decrease in the elasticity after COVID-19. So with the study so far is that work from home as, I mean the impact to work from home on voluntary quitting and then on liquid elasticity. And what I found is that those in workable occupation are also those who are less likely to voluntary leave their employer. But those in workable occupation remains also relatively more rustic after COVID-19 with respect to those in non tele workable ones. So in general, the main takeaway is that of course, you know, work from home has this kind of potential or reshape in general the employee, employer, employee relationship. So that's okay, thank you.

- So in the models that rationalize the quit elasticity with respect to wages, it's not obvious to me that those models would imply a linear empirical specification.

- Yes. - And so I would think that if you're getting above, if you're getting a high wage relative to your job and person type, you stick with it. And if you get a low wage and the lower the wage is relative to that benchmark, the more likely you are to quit.

- Yeah. - And that, that logic suggests that instead of the kinda linear in the wage specification you have and you have, you have, you have fixed effects to control for like the mean wage and the occupation and so on, that you'd actually want to construct measures that are the wage deviated about the, about the say the mean for that person's occupation, experience and so on. And look for non-linearities in particular, a differential response to wages above or below.

- Okay. Okay.

- So, but I don't know this literature very well. When, when people explore this, how do they typically work with linear specifications

- In general? I mean, in order to, to have like a direct estimation of the, the liquid elasticity, you should have an exogenous variation of wage, which is quite difficult to have in reality. So one way is to model, let's say this kind of, you know, the, the, the, the, the elasticity through this kind of of model. Okay? So it's like if you have, you know, you, you basically are, are modeling the probability of having a, a separation condition in a given interval on time conditionally this separation has not, is not happened yet. And so it's a function of this kind of control. So it's, you, you are modeling basically this probability as an model. So when you take basically the derivative of that, of that, but not that of that function with respect to wage, you get basically an insight on how much should be the, the, the elasticity when wage as is changing. So,

- So I I think one meta question from your paper is how, how much is what you're capturing a transition effect from the few years after the pandemic and how much of it is going to stick around for the longer term? And, and so I think maybe labor markets in your setting are, are, were less messed up in the wake of the pandemic than, than say in the US And, and so that can alleviate some of those concerns. But, but, but I, I, I think, yeah, more data from later periods would be very helpful.

- Okay, thank you. But I guess it's something that, I mean for sure, maybe I'm capturing this kind of transition, but I guess one way to, to, to deal with this issue is to take the tele telework ability of the individual before you know, COVID-19. And maybe in this way you can deal with this problem of transition because it's now the, the main problem I guess is that now you are, maybe you are leaving your job because now your occupation is tele workable.

- I am gonna start by thanking the organizers for putting such a wonderful conference together and for having our paper on the program. This is joint Work with Tom Kirkmeyer at LSC and we're gonna be talking about working from home in the public sector. Organizations around the world are grappling with a new status quo and they're trying to understand whether they should allow their workers to work from home or not. And when they're trying to make that decision, they like to know about the benefits and the costs of such an action. So one very clear benefit is that the workers no longer have to commute. And previous research shows that when workers are allowed to work from home, they take fewer sick days, they take shorter breaks, they also report higher workers satisfaction, better wellbeing. And these type of arrangements may be particularly helpful for the workers who have caretaking responsibilities. At the same time, it might be harder to monitor the workers when they work from home. Kyle showed us something about that a couple of days ago and the work by Emma and Natalia show that workers miss out some important interactions at the office especially so the less senior ones. And there is evidence that communication and coordination costs might also be higher when workers work from home. So what we're gonna do in this sec, in this paper is that we are gonna study the impact of working from home on the performance of the workers in public sector organizations, sorry, organization it it's one single organization. Specifically we are gonna rely on administrative and survey data from the crime recording and resolution unit. This is a division of the Greater Manchester Police in the United Kingdom. The empirical analysis is structured in three parts. In the first part, we are gonna try to estimate the productivity effects of working from home. And there we are gonna leverage a time period in which work location was plausibly randomly assigned to the workers and the tasks were also plausibly randomly assigned to the workers. In this setting, we find that on days when workers work from home there are 12% more productive than the days in which they work in the office. However, these local average treatment effects masks a lot of heterogeneity and we're gonna be able to map up the full distribution of individual level treatment effects. We find that reduced destructions at home are an important driver of the productivity gains we estimate. In the second part of the paper, we try to understand whether managers and supervisors in this setting more specifically can play an important role in managing working from home. And here we are leveraging an earlier time period when the managers would assign the tasks to the workers. And when we, we find that when managers are allowed to sign to assign the tasks to the workers, the productivity gains of working from home are much larger than in a scenario where the tasks are plausibly randomly assigned. In the third part of the paper, we would like to understand whether increasing the share of time that workers spend working from home has any impact on the performance of the workers. And I'm gonna mention this briefly now. We will not have time to to discuss it in detail today. I am happy to talk about it offline, but here what we are doing is that we are running another CT where we compare the status quo, which is a setting where the workers work from home 70% of the time with one treatment arm where workers are called into the office one day per month, similarly to what it was presented yesterday. And we find that working almost exclusively from home does not offer additional productivity gains over the hybrid environment in this setting. We also don't find additional costs. So that's perhaps a reassuring. So our work speaks to two broad sets to to broad sense of literature. First and foremost, literature on working from home. And what we think we can offer to this literature is that we are studying a public sector organization. And this is relevant because it's a setting where the incentives are particularly weak. What we mean specifically is that these workers cannot be fired. There is not much of a career progression within the organization. So the typical career incentives you might have in the back of your mind are really muted in this setting and there is no pay for performance as well. And the reason why weak incentives are important is because that's the setting where you think working from home might be costly in terms of perhaps reduced effort at home or perhaps more, more shirking. Another important contribution of our work is that we leverage a within worker design. So that means that we are gonna be able to estimate individual level treatment effects and we can go beyond the average treatment effects or the local average treatment effects that have been previously estimated in the literature. Our work also speaks to the studies that examine the social determinants of workers' productivity and especially so to the subset of of studies that look at the impact of managers, supervisors, and middle managers on the performance of the workers they receive. And what we find is that supervisors are play an important role in the setting. We study, and this is not an aspect that has been explored in the context of the working from home literature and I was very excited to see that some of you are working on this and we think this is a very exciting avenue going forward. So let me tell you a tiny bit about the background. We are gonna be working with a crime recording and resolution unit. This is a division of the Greater Manchester Police. Their job is to record crimes in a computer system. This is an individual job, there is no team aspect to it. And importantly they can do one of three tasks. They're called work streams in this, in this setting. So they can be recording cases from incoming phone, phone calls. They can be recording cases from outgoing phone calls, follow up phone calls, or they can be triaging cases. That means that some workers evaluate whether the the cases are they follow under the, the purview of this specific division or if they have to be reassigned to a different one. And but triaging represents like 8% of the workload. So that's not the bulk of the work they do within each work stream. The cases are assigned on a first in first out basis. So that means that they're plausibly randomly assigned to the workers. However, the question is how are the workers assigned to these work streams then? And we have some interesting variation. So prior to September 22nd, 2023, the supervisors would assign the workers to the work streams after that day. There is a computer algorithm that makes the assignment. Our argument is that the assignment made by the algorithm is plausibly random and we're gonna show you some evidence that corroborates that claim. All workers rotate through a rotation schedule. This is what the rotation schedule looks like. It determines whether the workers are assigned to work from home or not. The orange Andres are when the workers are assigned to work from home. The blue one blue ones are when the workers are assigned to work from the office is a five week cycle. At the end of the five weeks, the cycle repeats and there are five teams. And at each point in time, each team is in a different week of the rotation cycle. So that ensures that at each point in time there are some workers who are working from home and some that are working from the office. We are gonna rely primarily on the daily performance data from November, 2022 to October, 2024. This data includes the number of cases that each worker records in the system. That's our preferred measure of productivity. We also have some measures of the time spent by the workers to input these cases into the computer system for a subset of months. We have a measure of quality. This is important because you wanna be able to assess the trade off between quality and quantity. And we are gonna combine this data with the personnel files of the workers and data on medical absences. And we ran a brief survey, sorry, anonymous survey in October, 2024. So we are moving into the first part where we exploit a plausibly random variation in work location introduced by this rotation schedule starting from September 22nd, 2023, which is when the computer starts making the assignment of the workers to the tasks. And so we're gonna be comparing the productivity of the workers when they're assigned to work from home versus assigned to work from the office. So this is what the raw data looks like. The orange markers, they know the average daily log number of of cases recorded four days in which workers work from home. The blue markers are the corresponding estimates when when the workers are assigned to work from the office. And what you see is that on average workers are more productive when assigned to working from home. So next we are gonna put that variation into a regression framework. We are gonna regress outcome Y for worker I in data on an overall constant dummy for whether the worker is assigned to work from home or not. And our preferred specification includes worker and day fixed effects. We don't rely on them for identification. If you do not like them, we can exclude them. All the results are unchanged. Most of the results I'm gonna show you today will come from the two stages list squares procedure where we asim, sorry, we instrument actual work location with assigned work location. Before looking at the results, we need to think a little bit carefully about identification here. And so our argument is that all the workers rotate through the same deterministic schedule. So if that is the case over a long enough time period, you would expect our working from home dummy not to be able, sorry, assigned to work from home. I mean not to be able to predict the observable characteristics of the workers. And that's what we, that's something we can test empirically and we find that similarly our argument is that the computer algorithm assigns the test to the workers in a way that is plausibly run. Now if that is the case, you would expect the dummy of being assigned to work from home not to be able to predict the case characteristics. Again, that's something we can test in the data. And so we see that workers work on observationally similar cases when they're assigned to work from home and from the office. So let's look at some of the main findings. Column one reports, the first stage columns two, three and four. The two stages list squares estimates. So what we see from column one is that being assigned to work from home increases the probability that the worker actually works from home by 73.4 percentage points. Column two reports one of the headline numbers of the paper. This is our preferred measure of productivity and what we find is that when worker workers work from home, their productivity increases by 12%. Then you might ask, is it that they work longer hours or is it that they're faster? And column, column three and four tells you it's really coming from the fact that the workers are faster when working from home. So while you might have seen a version of these in other papers, other settings, what I think you have not yet seen before is the next set of graphs which we think are very interesting. So our within worker design allows us to recover the individual productivity of the worker when they're assigned to work from home and when they're assigned to work from the office. So here we are gonna be plotting the individual fixed effects when they're assigned to work from home on those sent to work from the office together with a 45 degree line. And these fixed effects are shrunk to account for measurement error. These picture shows you two important patterns. These two sets of fixed effects are very highly correlated. So what that means is that workers who are quite productive in the office are also quite productive at home and most of these markers are above the 45 degree linings in line with the positive treatment effects we just looked at. We can visualize the same variation in a different way so we can construct individual level treatment effects as the difference in the estimated treatment effect when the worker is assigned to work from home and assigned to work from the office. And so what this picture shows you is the distribution of this treatment effect together with the average ITT, which is the dash line you see in the picture. So what this figure shows you is that there is a lot of heterogeneity in treatment effects. Some workers are quite a bit more productive at home than they are in the office. That's not true for everybody. A little bit less than a fourth of our worker are actually less productive at home than than in the office. So I think the next step is trying to understand can we predict this variation using workers' characteristics? And what we have in the data is gender, age, and tenure. Using these three variables and the interactions between these three variables, we can explain 12% of the variation you see in the graph. So that means that if you're gonna try to predict if a worker is gonna be substantially more productive at home relative to the how productive the worker is at home using the observable characteristics, you're not gonna be doing a great job. So we try to understand the mechanisms. Now in the interest of time, I'm gonna give you a glimpse of what we have in the paper. So the way the, we start with a simple open-end set of open-ended questions. We ask the workers, what are the benefits of working from home for you? Now almost 90% of them say, I don't have to commute a little bit more than 50% of them say it saves money. They are talking primarily about transportation costs. They report better work-life balance. This is not surprising to anybody in this room, but what we think is particularly relevant for this setting is that 35% of the workers report fewer distractions at home than in the office. And approximately 22% of them report that they are more productive at home than they're in the office. And these can be explained by the fact that the vast majority of them also report having a dedicated workspace at home. So that might be helpful in explaining the productivity effects we find. But another thing that's quite helpful is the office layout. So this is a call center type of situation. So imagine large open room where people are making and taking calls. So this is a inherently very noisy environment in the paper we, we do much more to corroborate this with with the data we have. So in the interest of time, let me just give you a brief overview of the mechanisms. We think that these reduced restrictions at home are an important drivers of the effects we see. We wanna acknowledge that our treatment effects are also consistent with the fact that the, the workers report better mental health and less stress at home. This is not something we can validate directly, but we are not acknowledge that's a possibility. We, we know this is not in line with the work by online co-authors. Importantly, these results are not driven by differences in absentism. So we can measure that directly and we find that workers are less likely to be absent when assigned to to work from home. So if we wanted to incorporate that into our estimation, that would make our treatment effects larger and not smaller. Now an important dimension of of the productivity effects of working from home are young children, the presence of young children at home. This is a setting where 60% of workers do not have any children and the median age for of the children, for of course the children that are there is 17 years old. So for these group we have very, very few of these workers that have young childrens children at home. And so that is also something you should keep in the back of your mind when you think about the generalizability of our results to other settings in the paper we try to rule out alternative explanations. I'm gonna mention only one, which is we have a measure of quality stemming from internal audits. And so that allows us to evaluate whether workers who are assigned to work from home who record more cases and are faster do that at the expense of basically a worse job. And we don't find evidence of that. So in the second part of the paper, we try to figure out whether supervisors can help harnessing some of these benefits of working from home or can they play any role in this setting. And so again here we are gonna exploit these earlier time period when the supervisors used to assign the workers to the work streams and hence implicitly to to the tasks. And what we find is that when supervisors are allowed that are allowed to do that, they assign different tasks to the workers when the workers work from home and when the workers workers work from the office. And what we are gonna do in this picture is that we are comparing the baseline two stages, list squares estimates, these are the blue bars in the picture and those are the estimates you in the table a few minutes ago with the corresponding estimates that are computed over these earlier time periods when the supervisors made the assignments. So the difference between the orange and the blue bar blue bars is that in the orange bars the supervisors assign the tasks to the workers in the blue bars, they're plausibly randomly assigned to them. And if you focus on the two bars on the left where, which is our preferred measure of productivity, you see that the productivity gains of working from home are larger when the supervisors make the assignment relative to a setting where the tasks are plausibly randomly assigned to the workers. We try to dig into this a little bit more and what we find is that supervisors seem to know what tasks are more available to be done at home and use that information in the assignment. So I'm gonna conclude rather than going through the results again, I just wanna highlight a a few key feature of the, of the study. So we're studying a public sector organization that engages in the semi routine task. This is important when you think about external validity and you know, to what other settings this might or might not speak to. We find larger heterogeneity in the individual level treatment effects, but we find that, again, that's not something I was able to show you today. When we increase the share of time spent working from home, we don't see additional productivity gains from that. So overall this is a rosy picture of working from home, but we are only looking at productivity. If we were able to account for the time the, the, the savings generated by the fact that they no longer have to commute and they need less office space, the picture would look rosier. So I'm, I'm gonna stop now to to hear your feedback and question and I'm also happy to talk offline so many people isn't it? Thank you.

- Great. So there there's, there's two other numbers that would actually I think push the productivity effect up. One question is you normally see quit rates full and productivity increases with tenure. So 12% is like holding the person constant. But for the firm, I've seen this in other settings and no one's ever really estimated it. People stick around for longer and they become more experienced. You could see the number be, you know, 15, 16%. It'd be interesting if you can estimate that. The other thing would be this is productivity condition on hours worked. It would just be interesting though if you include commute, which coming back to some of the discussions yesterday, people treat as working time and it's not clear that, I mean that's kind of a, it's not obvious that's the right number, but then you're gonna see numbers like 20. I mean you could see really big numbers put it this way because also they're saving on commute and that's right now seen as we're not including that in working time. But you could easily include that in working time

- Your points are well taken. Unfortunately we don't have data on commuting. But the other two suggestions are things

- The 10 year, we've seen it a lot actually. I often think in the studies for a firm what you care about is the productivity of your workers, not the productivity of Nick. You care about your average worker, you know, they quit less and they get better over time. That boosts it

- Point well taken. Thank you so much Chris.

- Yeah, that 45 degree line is super, super interesting. One of the things that I noticed is that the rate of non-compliance with the assignment is pretty high. Yeah, you have like 13% of people who are assigned to work from home or sorry to work from the office who are working from home. Can you identify them or any persistent characteristics about who opts out of their assignment to go into the office? Because that might be very interesting to think about whether their characteristics are different.

- I, I think this is something we can, we have not yet done. It's something, it's very feasible. We have non-compliance in both directions and it would be interesting to see those two dimensions and examine them separately potentially. Thank you for suggesting it. Lily,

- This was so interesting. I, I was just interested when you were talking about the dedicated workspace and you're talking about external validity, like did you ask more if the dedicated workspace is like their own room? Like I'm, I think it's interesting what goes into a dedicated workspace that causes people to be able to focus more and like how much does the company have to invest to get to that threshold of the dedicated workspace being sufficient?

- This is a very good question. We were allowed only to run a very small survey so we couldn't ask many questions and so we, that's the only question we asked them. So we asked them what we meant by dedicated space is that they needed to have a chair, a desk and, and the company provides like the soft, like the laptop and the phone. I wish I could tell you more about it and we can ask, ask if we can survey them again. I dunno if that's gonna be possible, but that's something we could potentially ask in another round, another round of survey. Oh yes please.

- Thank you. Very nice experiment I got it looks like there's coming a consensus where if you have individual tasks in big shared offices, it tends to be better to work from home. I wondered because you showed that there are some people who have kids at home or family at home and then you have the survey response, fewer distractions at home. Is there an interaction if you have kids at home, maybe there's actually not fewer distractions at home, but that's a counteracting thing. And the second question, very nice of course the productivity improvement, do you have any data on whether that comes at the cost of quality? Can you, can you test that?

- Yes, so I can test the ladder and it does not importantly for, so most of our data is administrative. The administrative data does not contain information about the children. The information about the children comes from the survey. The survey was anonymous. So what I can do, so I cannot match my individual level treatment effect with the information. Do you have a child or not? But what I can do is within the survey I can see that if you, the age of the child is a determinant of whether you report fewer distractions at home and the results go in the direction you would expect. So you're more likely to report fewer distractions when your children are older, not, not likely to do that at all if your children are very young, I think. Okay. I I think there's a tiny bit more. Yes, please.

- Yeah, no, so going back to the topic of external validity, it seems to me that that you probably have high external validity for this sort of job that is requires fewer interactions, maybe something like IT support and so on, not necessarily in in the public sector. And, and it's not obvious to me that most public sector jobs are gonna that, so so your external ability is probably better with those sorts of like it support non-interactive jobs than with more generally public sector jobs more broadly. So yeah, how, how do, how do you think about that? I mean yeah, you're, you're kind of, or from the title, I thought this was gonna be more broadly applicable to the public sector and, and, and

- So that is a challenge of our work. So the challenge is measuring output in general, but especially in the public sector is really hard. So once you have an agency where you can measure that in a credible way, it's of course a very specific one. So I do take your point that externality is an issue. So the way we think about it is we think well incentives are likely to matter whether the organizations are public or private. So we think our results are, might not be likely to generalize to settings where the incentives are much stronger. Our routine, our task is pretty routine, right? Many public sector jobs don't have that. And so again, you might not be willing to generalize these two schools or hospitals just to make a concrete example. And again, here there's no team interaction. So in some, in some way you can think of this as the best case scenario for working from home because you don't have these creative tasks that, and you don't have these interactions. So I, I do take the point that we should be careful generalizing this to, to other public sector organizations, but this is the best we can do in terms of outcomes. If you have a way to measure better outcomes in a more broadly applicable organizations, I would be really, really happy to, to work with you.

- No, so, so I, no, I'm I'm kind of spit balling here. I, but I mean, I think not, you don't necessarily have measures of productivity, but you might have measures of job performance as reflected in performance reviews or in promotions or, or something like that.

- Yes.

- I mean, so, so I, yeah, you'd have to step a little bit away from, from kind of hard productivity metrics, but

- Yeah, there are other measures that could be looked at in other and more broadly applicable.

- Very nice presentation. Thank you. Along the same lines, I think it'd be interesting if you can, if there's any val, you know, variation within this setting in the complexity or the value of tasks. I'm thinking about this in the, you know, like in the garo like rossi, hanberg type knowledge hierarchy framework where they're, you know, when you have to solve problems sometimes you have to either kick problems to folks, you know, talk over problems within your team or kick 'em up to the next level of the hierarchy and all that. So, and that would be, you know, maybe you would speak a little bit to this, like generalize the idea if either way, if, if this were, you know, applicable to more complex tasks and you still find a pro two bump or it goes the opposite way. I dunno if there's any measure of like

- Unfortunately

- If you can get more granular on the task, basically.

- No. So unfortunately we don't have any way to determine how complex each of the these cases are. So I understand why that would be very interesting to do. I could not find a way to measure it in a way that you would think credible. We can think more about it, but I'm not so hopeful on this point.

- Really interesting study. I was wondering if there's any significant differences in treatment effect across the different teams.

- No, they look very much alike.

- Okay, interesting. Thank you.

- Thank you so much. I think we're basically out of time and if you wanna talk to me offline, of course, I'd be delighted to.

- Thank you. Thank you. It's a honor to be here, be included on the agenda. I just wanna also say thank you to the organizers. It's really remarkable how you've created a field that just didn't exist a number of years ago, and it's not the first time that you've created a field outta nowhere with management and all the turnover work. Thank you to the Hoover team, obviously for the incredible mics and all that. I remember just a couple years ago when it was like on the laptop, right in the other room. So this is beautiful. So the message of this paper is actually pretty simple. It's that when the pandemic hit and a lot of people went remote, time spent at work, declined time when leisure went up. But then as we started adjusting, figuring out what remote work was really like, how to do it, how to optimize with hybrid, that there seems to be some convergence in that the differences between remote and non remote have shrunk. Now the obvious question is what happened during to productivity during this time? Because we didn't see a big decline in productivity during the pandemic. In fact, we actually kind of saw a little bit of a, I mean, immediate decline, but then it eventually started, continue going up. And so there's gonna be a story of selection here and I'm gonna hopefully be able to unpack some of those ideas. So I, I started with a original motivation slide and I thought we already have enough motivation in a conference and it's towards the end, so no more motivation. But I did wanna put a Simpsons plot up here because doing justice to, to all the stuff that Nick does and bringing in good Simpsons quotes, this I, I literally plugged in my introduction to Chachi Bt and asked him to make a picture of Simpsons. I thought this was a pretty funny one because it was like 2019. It's everybody in suits except for Homer and, well, I guess only, only one guy had suits, but it was like typical office. And then right hand side is like multitasking, doing a bunch of things, eating a donut. So let me fast forward past the kind of jokes aside part and get to the main main course. Let's see, I think the, the digital,

- That's it.

- Yeah, that's exactly, that's the, that's the story. The digital gods wanted me to keep, keep just on the jokes. But, so yeah, we've had enough kind of introduction to all this. So I, I don't want to kind of beat a beat a dead horse, but yeah, we've seen a rise of hybrid. There's obvious questions about productivity. Unfortunately I don't have nearly as good of empirical setting as Alessandra does, but hopefully, we'll, we'll have some useful takeaways. The big question that the American time you survey obviously allows us to do, and I'm gonna be using the American time you survey to cut to the chase and I'll show you a lot on the data side is this granularity around the way that people are spending their time, how we spend our time on a day-to-day basis, how we break up activities between home production, leisure, et cetera. And we've already had a number of papers on the American time use survey and then looking at differences across demographic brackets. So you can look male, female, married, unmarried children, non children, et cetera. And then the big question of, well how did this carry through to productivity? A critical mechanism in this paper, and I'm gonna build a ROY model where people are able to self-select based on preferences and capabilities is that jobs are different. So job demands are different. I grew up in a restaurant, restaurants can't really be done remotely, but other things, a lot of the work that we do, a lot of the research that we do can be done remotely. So there's differences in job requirements and then there's differences in preferences. Some people like remote work more than others, some people do better in environments that allow for a lot of autonomy. And so I'm gonna build a model with that. This is, this is a little bit of just some raw data. It doesn't look super pretty, but it will tell you kind of the, some of the main results in the blue. I'm gonna be using the dingle dingle neiman index and I'm gonna get more into measurement, but take it for now that I'm using dingle Neiman to classify remote non-REM remote. So in the blue, you see time spent at work among high remote is is higher in 20 18 20 19 than the low low dingle neiman. But then during the pandemic you see a reversal of this in particular in 2022, I mean it couldn't kind of be any more reversed. But then by 2024 you see this convergence. And so what I'm gonna do is take this to a multi-variate setting and then build a model around it. So American time use survey, we already know that, well, I'm not gonna repeat it. BEA, we also saw Octavia talk about the BEA. And so it's basically same sort of industry level data. Dingle leman, we've seen this a billion times. I'm gonna focus on employed workers. So there is an open question about flow into the labor force. I'm going to fix sample on employed full-time workers. I'm gonna also focus on non self-employed workers. I'm gonna deflate earnings, all that sort of normal stuff. I will make, well actually one more point, I just brought in some data from Gallup and I don't have a slide on this, but Gallup has a workforce panel that's about 20,000 respondents a quarter and it's panels. So we can basically trace out the evolution from the same individual from 2019 onward. Now the downside with the Gallup data is that we don't see these granular measures of time use, but we do see hours worked per week. So these results are going to be robust to focusing on that panel level data and basically comparing the same person 2019 versus 2023 and up to 2024 and has great measures of remote work and intense stuff that we're Nick and I are doing. We've, we've, we've validated the dingle neiman index, I guess a number of papers by now, but just to again, make the point that in this paper it seems to be a good proxy. What I'm doing here is plotting for high and low dingle neman occupations. I'm looking at the share of time working from home versus the office. So again, a beautiful thing about the American time use survey is that you see where somebody is doing the activity. So you see if they're having coffee at home, you see if they're having coffee at a coffee shop, you see if they're working at home versus working in the office. And so you do see this really flat line among the low dingle neiman, and if you saw more kind of volatility, then you would say, well maybe this proxy over a five year span of time is not that, or actually six year span of time isn't that good, but it actually seems to be sort of reasonable. Another way to validate it is just to collapse this at the occupational level. Look at the dingle neiman intensity at a five digit sock level and then compare it with the share of time working from home. There's a positive coefficient, obviously not perfect, but again, some, some evidence why it's, why it's useful. Okay, so the kind of interesting part begins. So we're gonna have outcome variable is gonna be time spent per per minutes spent per day in a particular activity. K for an individual I in year T and I'm gonna be looking from 27 20 18 to 2024. And I actually should have put a plot on this around pretre and going back to 2016, but we'll, we'll get to some of this. So minutes per day, right hand side is index for remote ability. And in particular, I'm gonna be looking above the median, below the median, I'm gonna be interacting that with your fixed effects and, and then normalizing. So I'm gonna be doing basically 2020 to 2024, normalized to 2018 and 2019 control for budget demographics, control for the hourly wage or, or or weekly earnings put in industry and occupation fixed effects in. And so in the most parsimonious specification we control for kind of the income channel, the fact that your, your earnings might be changing and therefore changing the return to labor market labor supply. But in addition, we're gonna be comparing observationally equivalent people in the same occupation before versus after the pandemic. There can still be a lot of unobserved shocks that are not being picked up by these fix effects in particular selection. And so I'm gonna try to do some work to, to show that it's not just driven by some of those selection effects, main results. Remote workers are allocating less time to labor supply over time. This doesn't seem to be driven by just commuting. The fact, and this came up about how do people classify work and do you consider time driving to the office as work leisure also goes up? There isn't a big change that I find in, in this data on home production and shopping. Not to say that necessarily there aren't changes. Obviously we saw some really nice evidence from even better data. And for example, Stephanie's data on shopping was just, I mean, I'm not gonna try to horse race ATIs shopping data to the quality that that you have in your paper, but, but in this, it doesn't seem to be driven by as much there. So I, I try, I took the motivation of trying to make beautiful plots. It maybe doesn't match the beauty of some of the plots that we've seen over the past two days, but this will, this is one of the main results. So on the far left, you're just seeing a baseline specification that puts in basic demographics and you're, you're looking at the treatment effects for different years. So 2018 to 2019 is the kind of the omitted category. And so remote workers do spend more time, I should say individuals and occupations are rank above the median in remote ability. They do spend more time at work, but then you see a reversal during 2020 onward and it really declines in 2022. In 2024, however, which is the, the last one, the kind of smallest with the biggest airband. That's where you see this convergence. And so one of the themes, and this is gonna be part of the interpretation of the paper, is that during the pandemic, when we were all thrust into environments that we were not expecting that the adjustment was very difficult and maybe shirking went up, maybe just time at work went, went down and it was, it was more difficult. But then as we started getting adjusted, as organizations started making more clarity about what does hybrid work look like, then it started to thin then what the middle panel does. And then the far right panel does it layers on additional control. So in the middle you're showing it controlling for earnings, that really doesn't change the, the results. So whether or not you put log earnings in really doesn't make a difference on the far right. We put occupation fixed effects. Not surprisingly, when you put occupation fixed effects, the orange one is gonna disappear because it's coline. So you're just seeing the interactions with year and results are, are pretty much the same there. And so I kind of tend to think that yes, there's probably other omitted factors that I'm not controlling for, but it's a fairly insensitive to the inclusion of a fair amount of controls. And, and that gives me a little bit of confidence that these are genuine. You can look at the converse. So you look at decline in labor supply, IE decline in time spent at work, time spent working, and you look at the changes in leisure as well. So leisure is basically the opposite of that. Leisure went up during these years and it kind of continued into 2024, although it's not as big of an increase as it was in 2022. And again, results are fairly insensitive to the inclusion of additional controls and to occupational fixed effects. This is to make you work. Now we're at the half point of the presentation and make you look at a more difficult table, actually not that much to see here other than there's not as many stars as there were in the, in the previous plots, first three columns, home production, shopping, other, and childcare. So other is really interesting. So I'm gonna start with column seven through nine and it's just saturating the, the specification with additional controls. Column seven being kind of basic demographics, column eight, adding in earnings, and then column nine, putting in the occupation and fixed effects. And what you can see is actually that time spent in this other category goes up and it's statistically significant, especially in 2023 and 2024. And so actually this is a question and maybe, maybe somebody has an answer for me in the q and a period is, is what does this other category in ATA represent? Maybe it represents like social media and doom scrolling. I, it's possible maybe there's better categories in ATA that precisely measure social media time, but we kind of have anecdotal evidence that people go on or on TikTok and they're there for, they think they're there for 10 minutes and then it's like there for a whole hour. So the other category could be representing remote workers going on social media a little bit more often. Again, that's a possible interpretation. I'm, I'm not a hundred percent sure you do see some changes obviously in in home production, but not as much statistical significance here. An obvious point is that, well hey, maybe you wanna look at the individuals that have kids and so you restrict the sample. And so I do some, some work on that. So in columns one through three, a big part of home production is caring for children. And so I've also restricted columns one through three just to focus on the people with kids. But actually to cut to the chase a little bit, there's a, one of the interesting results, and this will come up in maybe one or two more slides, is that the people where labor supply declines the most are actually single males without so without kids. And so this is sort of consistent with like when you give young guys a lot of time and not as much responsibility, the best don't, don't expect the best things to happen. And so I'll show you that plot on another one. So it's not like, hey, remote workers and then the normal person that's at their job is shi it, it actually seems to be a little bit more the youth. This just basically presents the plots that I showed you in a tabular format and with two different definitions from Aguiar and Hurst about how you define leisure. Leisure definition two is a little bit more general. It includes sleeping and some other activities and find similar effects. Nothing too interesting. I did a placebo where I just looked at before. So I, the better way to do this and I, I just need to put a plot in. I should have put a plot in the, in the slides where you just do a normal pretre and you make a nice nice plot instead of a table. But I did a placebo where basically I looked at 2017 to 2019 and whether or not remote workers have any, any differential trends here, there's no stars. So it's not as if that the high dingle neman versus low dingle neman were on separate trajectories. I'm gonna get back to that. Okay, so then a natural question that comes up is about commuting time. Obviously commuting changed a lot during this, this span of time. HS has a nice way to look at pure work activities and so I call pure work in column two, which is excluding commuting time. And you do see this, this decline in that that happens there you again, you don't see as statistically significant effect. I mean you see really no effect in in 2024, but it's really concentrated in 20 22, 20 23. And then I also break it out by work related and other income generating activities. So individuals that maybe hold secondary jobs or maybe have a real estate business on the side or or whatnot. A couple alternative explanations. One could be theorizing our workers reallocating time to other productive activities. I, I don't see this big change in household chores or shopping, which is consistent with the heterogeneous treatment effects. That'll show you in one or two plots, which is basically that it's younger males and so younger males without kids don't have that much chores around the house and maybe a studio to take care of. So that also will make sense. No significant change in job search activities as well. It just asks people about, are you spending your time to look for other jobs? So let me keep speeding up because actually I do wanna get to the model and I don't want to just like basically show you an equation and be like, and there's a conclusion. So the only thing I'm going to leave off on kind of already previewed the heterogeneity results, squinting at that table is just not right. So let me show you this thing with the Gallup data. So Gallup tracks these individuals over a period of time and what I'm going to do is put in person fixed effects and compare individuals in 2019 relative to 2022 to 2024. And we have a measure of of hours worked, hours worked per week basically. So I look at the log of that, you find small effects that could be because there's measurement error. And one thing that we know, when you have a variable with measurement error, there's some attenuation and, but the key result is just to see 20, 22, 23, 24 negative coefficients, column two that adds to person fixed effects. It's also negative of 1.1% decline, but again only significant at the 10% level. So what I'm not sure about is whether or not this data is particularly suited to to study time use, but it's a, it's a helpful robustness exercise. The one thing that I can't reconcile though is that because there is a significant effect in 2024 and in AIS there's not as much of a significant effect. It does raise a little bit of a question. So a little bit more work I'll do on sample selection and scene. Okay, implications for productivity, this is the first desk rejection I got in this paper was you should talk about productivity. So we're talking about it now and the, I'm gonna show you some reduced form evidence and then I'm gonna show you the model. If we take the BEA data that Octavio already explained and basically put in a regression format and you do regressions of real value added growth or real output growth on your interactions with dingle Neiman. So now instead of working at an occupational level, you're working at an industry level. And so you basically use the OEWS data to crosswalk that into industry. And then you have shares of remote intensity and look above the median, below the median you see positive effects. And so that's kind of interesting because you say, oh, well if we're finding these declines in labor supply, why are we finding this increase in productivity? And so I'm going to build a bit of a model to explain why, why we might see that. So in a, in a ROY model, there's a, the, the critical idea is that there's self-selection into jobs based on people's ability. And there's a nice paper by and Andreas OSA and co-authors in a EJ macro in 2024 where they lay out a nice version of a model that I'm going to adapt for these purposes. I always remember, I mean as a first year doctoral student, you read papers, you're like, how did they think of this model? And then you're like, oh, okay, somebody's building on another model. So I'm gonna be piggybacking on aosa all. So there's gonna be an intensive margin of labor supply, there's gonna be these preferences for leisure and non-linearities and effort, non-linearities and effort is gonna be a major concept. I'm gonna get to that in with an equation in just a second. There's gonna be this heterogeneous sorting structural changes in the workforce might lead to new sorting patterns. So when companies decide to shift more to remote work, then there's gonna be this, this new sorting pattern that opens up workers might. And so this is a key insight. Workers might transition into roles that they're actually better suited for in particular now that you're able to apply for a job in a completely different geography that opens up new matching patterns. Preferences are, are pretty simple. CRA, we've got log consumption plus a disutility of labor supply and labor supply. Critically PHI is going to be indexed by eye. So there's gonna be heterogeneity by individual and that heterogeneity is going to allow for some of this self-selection. You could also represent that heterogeneity on the production side. But this, and there is gonna be, I will show you something on the production side. So we're gonna have two different types of labor. One is gonna be high remote HR and not, not human resources, hr high remote and then low remote. LR Zeta is gonna be governing the elasticity of substitution and K is gonna index the sector. So sectors will have different degrees of substitutability due to coordination costs and all that sort of good stuff. And then on the individual side, we're gonna have individuals putting in time, but theta K is going to govern the non-linearity of labor supply. So some individuals can put in an hour and get a lot more mileage than somebody else that maybe puts in an hour. And then a is going to be indexing the individual productivity term. I'm going to basically match this data, calibrate it with method of moments match, match everything. Exactly. And let me just kind of land the plane now that there's a minute left to show some of these results. Actually let me so match. So I matched the data that works well and then the two counterfactuals. So you always build a model to do counterfactual. So one of them is say, is gonna say let's hold employment shares constant in 2019 rates. Let's pretend like the pandemic didn't happen and that there wasn't this resorting, what is output gonna look like? The second is gonna say, assume that individuals have homogeneous preferences. Assume that individuals, actually some people, I mean I I like, I like working, I I my, my PHI is like, I don't know, put a zero there or something like that. So there's homogeneous preferences. So heterogeneity preference is actually great for the labor market because people are able to sort into jobs that maybe somebody else doesn't wanna do, but I wanna do. And so what I show in the bottom panels here is these two counterfactual simulations and maybe putting a plot is actually, maybe it's be be better just to put numbers, but on the bottom left in the orange you see in the present. So basically what ended up happening versus 2019. And so we're comparing the two shares and just showing that again the counterfactual hold employment shares consonant 2019 rates. So if we had held the shares constant, you would have lower output. The difference between the blue and the orange is about 12%. So the fact that we had this sorting into new industries, new occupations led to a increase in output on the far right is the second simulation that basically says hold fixed everybody's preferences and imagine that they're all the same. And you see that there's lower output in the case where people have homogeneous preferences because there isn't this optimal sorting. So lemme just stop here by saying main results is that labor supply goes down time and leisure tends to go up, but the reason why we didn't see this big productivity decline in the aggregate statistics seems to be because of some of this reallocation on the individual sorting side. So really looking forward to comments, this is the optimal place to be for feedback. So really grateful for it.

- Ah, it's a very interesting presentation. One thing I was wondering, I think you can only see one person in the household, is that right? So maybe you know, you could see like different responses for a household with two adults where some of this could be a reallocation of tasks within the household. So if you split it out by single adult households and body adult households, you might get a better sense for how much is reallocation and how much is a change in the like household level time allocation. I guess this is specific to what we're working on, maybe too specific, but I was thinking about how online shopping is classified in the time use survey, whether this shows up in the grocery category or whether if you're using an app, say it shows up in the organization category or some other type of category.

- Yeah, great questions. There is, individuals are asked about spousal labor supply spousal earnings. And so I can look at that because the effects and I, I will I guess bring it a table up, although it's a bit of a pain to look at, is the heterogeneous treatment stable? So in the column three, column four, so you see the biggest decline in 2022 among single, that's the 71.4. So 71.4 minutes per day, less in in labor supply. And so that's part of the reason why I I, I mean reallocation is probably playing a role, but because the biggest declines are concentrated among single and those without kids tells me that there's something, something going on. And this connects with the big debate that people were having about social media and that individuals might be doing it out of their own volition, but there's this, the literature on dynamic and consistency. And sometimes you say yes to something that maybe in retrospect you wouldn't have said yes to do. Thank you. But, and I also look into the online shopping, definitely the work with Eric Nissen that we have with the remote life survey. We do ask about online shopping and so we need to go back, that was one of the papers we presented like two years ago, but never actually utilized that online shopping data. So this is maybe a perfect opportunity to do that. Thank you.

- Thank you. This is super interesting in the literature actually, there's a way to capture multitasking and the presence of other people when you're doing activity. I think now you could use that kinda marriage to tap into the productivity, right? So for example, one hypothesis is that why the married people with children are now decreasing their work time so much as the single people is because they're doing other things while working, right? Yeah. So then maybe alternative explanation to your, your study.

- Yeah, that's a good mechanism or a kind of explanation for, for why we see also people, I mean if it's, it's sort of like the, the quote in dodge ball, if you can dodge a wrench, you can dodge a ball. So if you can have kids, you can multitask with kids, you can probably do a lot of other type of multitasking and do some hard things. So cheers to all the parents in the room. Yeah. Next,

- I think it would be very interesting to u use the Gallup data to say something about who changes their time allocation, because obviously that's gonna be a heterogeneous thing in the context of your model. I mean, I had big debates with people about what the labor supply effects of COVID were because I worked all the time and then my friends who were software engineers were just like, yeah, we sat on the beach and didn't do anything. And so as a result, like I think one of the things that you can do to better contextualize some of the ROY results, is to use the Gallup data to think about who increases their hours and who might cut back on their hours based on pre-PA characteristics. And then if you can tell us something about how that sorting maps into the model, that would be I think, very useful. Because as of now, I think the model is just this GMM type of thing where everything happens through an error term and heterogeneous preferences and whatever. But you can actually probably pin some of that down with the data.

- Yeah. Is there, on the pre pandemic characteristics, what do you think would be the most salient? Is it like industry occupation or something at the individual level that you think particularly jumps out? Can you see performance pay or career concerns or something like that? Like Yeah. Yeah. There's some information on that. Yeah, actually, and yeah, appreciate it. Yeah, the Gallup data has these really great measures of perceptions about the workplace and attitudes and engagement, and that's probably the best is that one thing Gallup is known for is something called the Q 12. Basically 12 questions they asked a proxy for employee engagement. And so that would probably be the best proxy. I can definitely do that. Thank you.

- I'm not sure what your referee comments were with regards to productivity, but I, I, I think something big that you're showing that, that that, so we certainly see and we think is a big effect, is the increase in leisure, which is going to be consistent with a big increase in welfare when people work from home because they, they're, they're saving all this commuting time and, and so having, having hard facts about how much that is, is, is, I think very valuable. And, and that's one way in which you could pitch the paper. I think the productivity effects, I don't know, it's, it's not so clear to me that, that you have the data and the setting that, that would really help you make a contribution on productivity. Whereas with leisure and, and the potential welfare effects of that, you, you de you're definitely better positioned.

- Yeah. The value of leisure and putting a price tag to that. Yep. Is and is there, you think there's a, a good way to, okay. How much people value leisure? Is there, is there a

- So in, in the new version of working from home of why working from home will stick, that we keep promising to put out and keep delaying, we, we, we are trying to do this, okay. Basically by, by, by calibrating to people's labor supply, their reported commute time and, and, and so on. So, but I, so I think you could do something similar and, and, and we can talk more.

- Well, we have a commitment device now that the paper will come out because it's on the record. So, Nick, you had a question?

- Sorry. On the, on the, just on follow up, there's a, there's an old science paper with Alan Kruger, Danny Cameraman, geez, some others in 2002 that values activities per hour.

- Yeah.

- And so it has in there, I just, I know that paperwork because it has the value of the, the hate of working is the same as the hate of commuting, so people dislike them. So I, I had a question on, I guess for people that use time, use diaries and a bit following 'em from some of Stephanie's comments, which is I assume if I'm in Attis and I'm going to work nine till five, I report, I'm working for eight hours. If I'm at home and I'm working nine till five, but say I take half an hour for lunch, watch, you know, 10 minutes of TikTok or whatever and you know, chat to my spouse or something, what do I report? 'cause in the office I'm probably scrolling around and I'm having lunch with a friend and chatting about football. I just, I dunno, I dunno if anyone's looked at it, but I, I would think that maybe people report they're working more in the office when they're goofing off than, but I does, does anyone know, has anyone compared actually self-reported to other hard data like keystrokes or, I dunno if Liam knows this. Yes. Do you know, do you have any idea from time, doctor? Well I guess you'd need to match up tis to like actual, 'cause you have a, it's

- Completely different data sets here. It would be very interesting to see how compare to more of workforce analytics tool to be able to kind of,

- That it be interesting to take your, I dunno if you could, and just record in time doctor, because I don that, that seems to push against people. That would be one reason why people are home appear to be working less. Yeah. There may be

- Just upwards bias towards reporting that you're working.

- Yeah, because if I, if I imagine I'm work from home and I spend 45 minutes having lunch with, you know, my spouse or something or just watching tv. Yeah. I probably don't record that as work. Whereas in the office, I probably just think my working time starts when I get in and stops when I leave.

- Yeah. I mean, the short answer is that they ha with the diary is basically people are recording the particular activity every, it's either every 30 minutes or every 15, it must be every 30 minutes. And so there is, there is an endeavor to have precision, but nonetheless there could still be, oh, we'll call it a working lunch because I'm, I'm meeting a colleague and so like

- The equivalent is law firms. Law firms, you bill every six minutes, but apparently in their day there's no going to the toilet. But I mean, lawyers must go to the toilet. I mean they must do. So, you know, they don't just, but you see their full days. So they're clearly

- Including Yeah. Okay. They're thinking. That's funny. Okay, great. I think we're outta time, but thank you so much for the engagement and for the great questions and feedback.

- Okay, thank you for sticking around toward the end and, you know, looking forward to getting your comments. This is joint work with an amazing team of co-authors, Smith Guardian, Bianca Sarda, who are at Good Business Lab and my frequent coauthor, not nice of them, who's at Michigan. And this is actually one of these rare projects that's, I'm presenting results. It's like probably 85, 90% completed, but we're still in the field on this. But it happens to be that we've done things in, in waves and surprisingly, although this never happens to me in experiments, you know, just looking at the data we have yields a lot of precision. So I can show you some results that I'm fairly confident about and hopefully they'll continue to, you know, dial in on precision and answers some other questions with our latest round of data. So this is basically gonna be very different from a lot of what we've seen until now in this, in this conference. This is, you know, thinking about work from home and productivity, but in the, in a, in a context of a very low income context. And thinking in particular about how, you know, women can be connected to global value chains via home-based work. Let me show you, tell you a little bit more about this, this context. The general policy question that a lot of countries in, in the kind of developing context that are, are, are wrestling with is, can job flexibility help enable economic empowerment, especially for low income for women? So I'm gonna be looking at an experiment in, in rural India. So lemme tell you a little bit about the, the context in India, you know, India has very low labor force participation of women relative to other countries in its like sort of per capita income category. So this is like this sort of u in, sorry, U-shaped labor force participation curve for women. That's been, you know, talked about in Claudia Golden's work and, and lots of other subsequent work thinking about how do we move, you know, women's labor force participation up in some of these countries. India is sort of way off the curve. If you look at the, the graph right here. A big part of that is that norms around who does is responsible for household work are highly variable across context. And in India actually there close to the most skewed in the world, meaning that women are largely responsible for household work across India. And that is kind of potentially linked to a lack of labor force participation in this context. So women are, you know, working about six hours a day according to this survey from the OECD in India per day on kind of unpaid household labor. Men are working about one hour a day and that, you know, gap exists across countries, but it's sort of much bigger in India than it is in in in many other countries. So what are we gonna look at today in this, in this project, the question we're asking is, can access to flexible job opportunities increase the participation of women in the labor force in, in rural India? And if so, kind of how exactly does time use adjust for these women to make room for that, for for work, right? So if you're working a lot on household labor and that's the norm, how does your kind of time use adjust to accommodate the opportunity for work? We're gonna ask, you know, the question that's been asked in previous studies here today and yesterday, the day before, looking at productivity as a key outcome does work, you know, work from home generate greater or less productivity for these women than working in an office context. And then what are the income gains for women from being connected to global supply chains, which is kind of the, the actual intervention that we use. Okay, so this part, this is gonna be a study in rural raan in India. We have about 450 women in this enrolled in the study across two villages. We are working with a, a partner called the Kala Trust, which is basically an organization that provides handcraft skills. So it trains women in kind of creating local handcrafts and then helps to connect them to sort of global markets or domestic markets that, you know, mostly in urban areas by selling those handcrafts to sort of intermediaries or retailers. So we worked with this organization to, you know, randomize the enrollment of women into their program, gave these women the skills they needed to create handcrafts, and then looked at different work arrangements and saw how, you know, what happened to their production and their productivity and time use as a result. And we also tracked a, a randomly allocated control group to look at kind of what kinds of, you know, time use patterns and income gains are achieved in, you know, from either of these kind of work arrangements working from home or from a sort of centralized, you know, factory setting versus, you know, outside option. Okay, this is kind of a, a a little bit on the experimental design. This is not exactly accurate in the sense that we did things in batches. And so we actually, the first couple batches of women we enrolled, we only had two treatment arms work from workshop and work from home. And we randomized women into each of those arms and then we, you know, learned that we actually might gain a lot of information from having a pure control group. So subsequent batches actually had a pure control group as well. We further, you know, sort of had a sub experiment here, you know, thinking about whether participation of women from home, working from home versus in a centralized factory setting might differ for various reasons. So the kind of sub experiment was, was aimed at kind of sussing out the, the, you know, mechanism for greater potential participation of women working from home. And in particular, there's a lot of talk in India and you know, some reasonable evidence that there's often stigma for women working in, in households. And that kind of is what really prevents women from joining the labor force that basically joining the labor force has, you know, comes with it certain assumptions about women's families kind of income status and or you know, their purity norms that prevent women from kind of taking up work because they'll be sort of viewed as impure if they're sort of spending too much time outside of the home and that might inhibit women. You know, that sort of perception or stigma might inhibit women from joining. Another explanation, another potential mechanism is that norms around household labor are such that as we, you know, saw earlier in the summary statistics, we, you know, there's women are just responsible for the majority of household chores. And so those norms are fairly sticky. They don't change very quickly. And so you can't, you know, if you're doing a substantial number of hours of household, you know, responsibilities, doing full-time work outside of the home is just, you know, impossible given the hours in a day or very difficult. So, so, you know, another hypothesis is that, you know, it's the time use, you know, structure and the kind of sticky norms around responsibilities at home that prevents women from really joining. So we're gonna sort of evaluate, you know, this, try to knock out this the, the relevance of this stigma question through a sub experiment. The sort of, you know, I'm gonna, I'm gonna skim over the details of that a little bit because we basically find absolutely nothing there. So, you know, indicating that that in, at least in this context, stigma is not a major factor in women's decisions. The way we assessed this was essentially by further randomizing each of these group, the participants in each of these groups into two arms. One in which they basically, their participation in our program was announced via a poster at the village center. They basically made public that this woman was working as part of a, as part of this experiment. And so, you know, that idea was that, you know, if that revelation and the publicness of the, the signal was what mattered a lot, we should see a lot of women dropping out even if they were working from home. So we actually don't find that that's true at all. Okay. We recruited women, you know, from, from from their homes and villages. They were, you know, given a two week training in, in crochet, which basically made these crochet products, which we are, are sort of our primary product that we assess. We after training women were randomly assigned to either produce at home or in a village workshop. So we essentially set up these village factories you might call them, which are very, very close to women's homes and you know, they were sort of within a 15 minute walk of, of women's homes, which were open 10 to six daily and women could go there at any time and, you know, they could also bring their children. So we kind of tried to make these jobs fairly flexible. So really we're only varying whether you can actually do this work in the, in the workshop or you can do it actually in your home production and pay for these women were identical and the kind of support and materials were also identical across the various arms. Okay. Some pictures, this is the product that we had the women make. Essentially it was a nine inch by nine inch square that was made via crochet. And then this kind of product was integrated into kind of further products that were then eventually sold in by retailers. So this might go on a handbag or a backpack or a sweater and you know, so there was, anyway we, we, we have a set separate experiment that's assessing kind of creativity and innovation in the creation of these crochets as well. So there is some leeway in, you know, what women could do there. These are two kind of pictures that describe, you know, women in, in the workshop setting. This is basically just a room where, you know, women could come and, and work on their products and the materials were available there and the, the, you know, as you can see, you know, children are, are allowed in the room. They're hanging out in the background there. And you know, the, the picture on the right is a woman making the crochet at home. Okay. These were the posters as I mentioned, so I'm gonna skip over this because we find basically no, no differences at all between, you know, women who are randomized into the poster treatment versus the other treatment. I'm happy to chat more about that afterward if you have questions. Some summary statistics on, on what the kind of population here. This was a, you know, women in their mid or early thirties, most of them were married, only about a fifth of them had ever worked in any income generating activity outside of the home. About half of them had childcare responsibilities and they spent a very large amount of time on home chores and childcare. So at baseline women were spending about seven and a half hours a day doing various activities related to unpaid labor within the home, including childcare. And when asked whether they preferred to work from home or from kind of a centralized workplace, 85% of them said they preferred to work from home at baseline. Okay, let's skip through randomization. Alright, so what do we find? Women who were assigned to work from home were much more likely to take up the job offer. So about three quarters of women who were assigned to work from the workshop took up the work opportunity, meaning they at made at least one unit of the crochet during the intervention period. That was nearly a hundred percent, it was 97% in the, in the work from home group. So sort of on the extensive margin, women were sort of more likely to participate if they had the opportunity to work from home. They also participated more regularly if they had the opportunity to work from home. So two in five women assigned to work from the workshop showed up to work on any given day, whereas four and five women were working from home on any given day and we sort of, you know, collect hours as, as part of our time use diaries, we collect hours that they were actually working on the crochet products. So we we're, we can, we can identify, you know, what they were doing in that time. Okay, here's a protein. So I'm gonna show you a bunch of graphs by the way. And as I mentioned, you know, this is, I'm very happy to say that this is one of these rare experiments where everything is precisely estimated. So these are all significant at the 1% level, these differences. So this is, you know, quality. So we had, we had a, a few different ways of assessing output and, and productivity. So women assigned to work from home produce more output and work more hours than women who are assigned to work from the factory. So this is, these are all intent to treat by the way. So, so women working from home are producing 15 more quality adjusted units per week, which is nearly 84% higher output relative to the women in the workshops. And that sort of commensurate with a percentage increase in hours worked per day. So women who are working in the workshop, were working about three hours a day. They're in the working from home, they're working 1.7 more hours per day than, than working from home, I'm sorry, than working in the workshop. They're also more productive when working from home. So we measured productivity in, in a variety of ways. The first, and I think the most convincing was sort of like pure productivity, meaning the time that it takes for women at the end of the intervention to produce one acceptable unit of crochet. And so we did this via a lab in the field experiment where we basically had each part or a sub sample of each group come into this lab and create, you know, one crochet product for us in sort of identical conditions and saw how long it took them to do that. So the average time that it took for a woman in who had been in the workshop treatment was about 42 minutes. The average time it took from work from home participant was about 39 minutes was a basically a 7% improvement in productivity.

- Can I, can I just quick, so you're saying when they came into the lab, the ones that were normally working from home

- Yes.

- Passed in the lab.

- Yes, exactly. Exactly. Yeah.

- So there has to be some training effect or experience.

- Yeah, exactly. So, so I think what we find is that, you know, and so actually sorry just to con conclude that, you know, thought basically, you know, and there's a, there's an improvement we see in the actual realized, you know, if we just look at quality adjusted output per hour from the, from the actual data during the experimental period when they were in these conditions, they were also about 9% more productive when working from home. So, so why do we see this effect that's supposed to say understanding the treatment effects on productivity, it's cut off. But you know, basically what we find is that women, because they're producing so much more of these crochet products at home, they're actually moving further out of on the learning curve. And so, you know, we can actually map out these differences and by the end of the intervention period, you know the difference between the end of the red line, the end of the blue line is about 7%. So it actually kind of maps up really exactly with this kind of differences in learning curves. The extent of learning by doing that's achieved is just greater because women are producing much more when they're working from home. Okay. We had, you know, en numerator assisted time diary surveys that were quite extensive and they measured a lot of things including multitasking across different types of tasks. So what do we find here? We find a substantial kind of time reallocation. So, you know, work from home women spent 83% more time on crochet than work from workshop as I mentioned earlier. And I'll, you know, that increase in work hours is kind of accounted for by about one hour, less of kind of personal wellbeing time. So this is basically sleep, leisure and personal care time, which dropped when you add those things up from 12.5 hours to about 11.4 hours a day, a 9% decline, as well as a drop in kind of home chores and care activities. But that's a smaller drop, which is not statistically significant. So they weren't actually doing less of their chores, they were still doing all of those effectively, but spending a basically a little less time either doing, you know, leisure or, or sleeping. This is just a, like a pre-post fact that I thought was interesting that basically attachment to the labor force kind of increases after the intervention. So using just the first two, you know, batches of data because we haven't followed out the, the subsequent batches yet. We're still, you know, about a couple months away from that before the intervention, only 31% of the women had ever worked in any income generating opportunity post intervention that went to 61% said that they had worked in some opportunity other than, you know, the, the one that we gave them in the past six months. So, you know, this potentially kind of expanded other kind of the willingness to try other opportunities as well. And interestingly, you know, the organization that we worked with actually ended up offering intermittent work to these women and about half the participants to up opportunities in, in the following six months if they were part of the program. Okay. So I'm gonna show you a few results on, on our sort of pure control group and then I'll, I'll end there. The, you know, for kind of our latest batches of implementation, we included a pure control arm, we just measured time use and things like that, employment outcomes, et cetera. And what do we find are early results suggest that the program substantially raised women's labor force participation and earnings. So treatment women who were, were, you know, part of the work from home group had four and a half times more income than women in the control group and from the workshop group they had about three and a half times more. So both opportunities substantially increased women's income in this setting, but work from home increased it the most, nearly all treated women engaged in income generating activities during the intervention compared to only about a third in the, in the control group. Despite that kind of large gain in earnings, there is limited evidence that, you know, financial control changes for these women as a result of the income gains. So, you know, again, I think norms may be quite sticky in this setting. And so, you know, providing these opportunities and increasing income doesn't translate very readily into, you know, sort of a, a shift in financial control over resources in the household and other measures of women's economic empowerment. And let me end with a sort of interesting finding that I, I hope we'll get to dive into as we get more data on this, which is that, you know, interest in taking work, taking up work in the future is kind of nearly universal. So across, you know, both the treated and the control group, everybody's open to paid work in the future, but taking up work outside of the home, that percentage actually drops for the treatment group. So 75% of the control women said that they were open to jobs outside the home, but if they were part of our treatments, either working from home or working from the factory that drops, that willingness drops by 18 percentage points. Huge, you know, difference and drops most for women who are working from home. So if these women were working from home, they're much less likely to say, I would be willing to take up any kind of work outside the home. So they really liked what they were doing at home, they did not, they were less likely to say that they were, you know, interested in something outside of the home that might suggest that, you know, there there's a, you know, Sohani otta and, and Lisa Ho have a really nice paper that kind of related to this work that shows that actually in their context working from home or working in a flexible job actually increases your willingness to take up, you know, even work outside of the home later. In this context actually we find the, the opposite effect that, you know, if anything women who have worked from home actually tend to prefer that so strongly that they are less likely than control to want a job outside of the home. Okay. So let me, let me stop there and I will, and I will, you know, happy to follow up with questions and things like that. Thank you for the listening.

- Just real quick, were they paid by the unit or paid by the hour?

- By unit. By unit, yeah. So like this piece rate work? Yes, yes.

- Also on compensation, how does it compare to the compensation available to them in the local, local market? Yeah. One question and second question is, are the costs being covered by a commercial entity that is using these or is there some subsidy necessary to make this work?

- Yeah, great questions. So the first one, what we did was, you know, our original goal was to estimate a labor supply curve by varying the price of this. We found that to be extremely operationally difficult because within village there's lots of informational spillovers and people were very upset if we varied the price too much. And also what we found was that, you know, actually the, the elasticities, you know, fairly, fairly small that women just really wanted to work. And so, you know, we decided to benchmark the per hour compensation from some piloting piloting that we did to, to the compensation that they would get in a government kind of work fair scheme, which is, you know, much evaluated, it's called the nrea scheme. It's much evaluated in India, it's available for a hundred days a year of paid work in every rural area. And so we benchmarked earnings to that scheme 'cause women were eligible for that scheme. And scheme is a Indian way of saying program. So, so that's, that's on compensation. It was meant to be sort of competitive with local rates for women. The on on the sort of like val, I think, you know, your question gets at a little bit is basically the value to, to firms like is this a valuable enterprise for firms, you know, to, to expand, you know, create these market linkages and expand their production into rural areas? That's a complicated question I think because there's obviously other, you know, there there's this sort of economic value, there's also like regulatory infrastructure and things like that that, but I think just, just looking at the economics, right, most buyers in this context really want, you know, stuff from factories because there's economies of scale, there's standardization, et cetera. But I actually think that what these results suggest, this is like fully unsubsidized that these, these things were bought by, you know, just this was, this was standard practice for this organization. The pieces were then bought by retailers, they're mostly domestic retailers that were, you know, putting them on handbags and have a very cool backpack that has a, you know, one of these swatches on it and and being sold at a profit. So, you know, I think this is possible and perhaps there's just like, you know, other frictions that are, you know, that are preventing large buyers from coming in because I think this is not very difficult to teach to women and it's kind of economically valuable if you can link it up to demand outside of their village context. So that's, you know, I think further exploration's needed there in terms of like the scale up potential of that. Yeah,

- For apologies, you talked about it, but I was wondering, so usually you think about that liver supply the other way around that how, how, how much characteristic and circumstances of the woman affect their willingness and ability to work. But about the provocative best finding that you discussed, could it be the other way around economies of skill and scope in providing independent care and expanding the size of the family unit fertility? Could it be that they work from home actually encourages women to engage also more in home activities of this kind? That's why they're less willing to work

- Outside. Yeah, that's interesting that, I mean if anything we've

- And you tell from the data Yeah. What happens to the household size dependent care, elderly and young.

- So in terms of their time use, they're, they're actually sort of there, there's kind of no impact, zero impact there on, you know, actually spending more time it, this is pretty short term so I'm not sure we would see things like fertility or you know, household size, things like that. But we, we can see how much time they're spending in various types of care activities and home activities and generally we find basically no changes. So again, that may be like they're already at the, at the sort of upper bound there or, or that you know, the norms are so sticky that

- You don't see in Mexico what s and I find is that like cousins or elderly uncles move in

- Yeah,

- The household. So interesting the size changes, we thought it was a mistake changes very quickly because it's not newborns or

- That, that's interesting. Yeah, yeah, I can certainly look at it. I'm not sure like we've actually tracked that as an outcome, but we probably have that in the data. So that's a good question that we can easily assess. Yeah.

- Okay. So my question was you don't see that the extra income affects intra household bargaining, but it might be that this wom this money is treated as women's money separate, so it's not going into the household pool, right? So it's their own separate money. So you could see if for example, they're saving it or what they're spending it on, so they're giving up sleep and leisure but they might be buying something or they might be having savings accumulate, which would be interesting.

- Yeah, that's a good question. I, I think I'll see what we can find out on finances. I actually don't know how deeply we have measured, you know, people's finances in this setting, but, but certainly yeah, we don't find that like sort of total control over the pooled household income doesn't really change. But you may be right that they're sort of putting this into or keeping their, you know, own account. Yeah, yeah, yeah. Other

- Just one. Do you think it's the breaking up of the time versus the, versus a block of time when you go to the place, you got a block of time when you're at home you can go in and out,

- Right?

- And this looks like an activity that makes it easy to go in and out of. So is it the, is it the breaking up of the one big giant workday into mini workday that you say is the big benefit?

- Yeah, I mean that's a good question. I think something that we, I think need to assess more in a more detailed way. We do actually have, you know, for these products we actually, for each one we have this little QR code on it and we're actually able to, we, we asked women to document when exactly they made each one so we could actually assess that. I think our sort of anecdotal finding from the field was, it's not necessarily the breakup of the full workday because they're only going in to the workshop for a couple hours a day. So it's already a small chunk, but it's the ability to do things, you know, at odd hours, especially at, in the evenings. So like, you know, the, the thing closes at six, most women kind of stay out inside their homes after that period, you know, of time and women are able to continue working, you know, after bed after a kids' bedtime and things like that. Which is, you know, the classic work from home mechanisms. I know that's when all my productive hours are. Yeah, good to make. It'd be nice to see the share that's off the dock actually. Yeah, that's nice. We can definitely do that. Right, right. I like that. Okay, cool. Thank you very much. I appreciate all the comments.

- Thanks so much to the organizers. This has been a wonderful conference. For those of you who don't know me, my name is Aruna Ranganathan, I am at uc, Berkeley's High School of Business and I'm very excited to have this opportunity to present and get your feedback because this is the first time that I'm presenting this project. And the project is currently titled The Deep Work Divide Gendered Interruption in Remote Work. An important concept in sociological research is time control. Where time control refers to the ability to decide when, how long and under what condition one allocates time across work, family, and personal domains. So the key element is the ability to decide kind of how you're spending your time. And sociological research has found that women often face much less time control than men. Why? Because of deeply entrenched gender norms, but also the expectation that women might always need to be available for family responsibilities. And these gendered patterns of time control have been shown to be especially prevalent in the global south where gender roles are even more deeply entrenched and defined. So for this reason in the nineties some influential sociological research actually showed that the office was a sanctuary for women. Hostile documented that the office offered women protected time away from domestic demands because we know that your physical location or where you work affects interruption patterns and your ability to concentrate in your work. Now in contrast, remote work increases the bleeding of non-work demands into your work hours. And scholars have found, for example, that mothers working from home increase their parenting activities during work hours as compared to fathers. Now a gap here is that existing research documents gender disparities in time allocation between work and family domains when working from home. But scholars have not yet systematically studied how this variation in time use might affect your work related outcomes and your performance. And this might be because of a lack of data. So existing research has relied primarily on time use surveys or ethnographic approaches that provide rich descriptions of time use, but limited data on work itself and work outcomes. So as a result, most studies focus on the quantity of the time you spend on work versus family activities rather than the quality of your work time and the ability to engage in sustained concentration. And so that's the focus of this project, quality of work time. And we think of quality of work time specifically with respect to deep work where deep work can be thought of as sustained distraction-free focus, concentration in one's work and deep work is an important outcome to pay attention to because deep work can predict your experience of flow states, which has been shown to be crucial for psychological wellbeing. Deep work can also predict your job performance and might also affect how meaningful you see your work as being. Now management scholars who have studied deep work have highlighted how interruptions disrupt deep work 'cause unplanned requests and interruptions might force work into fragments. Unexpected interruptions might also disrupt the momentum you have going when you're doing some work. And when breaks are imposed rather than self-directed, they can also become cognitive burdens. A gap in this deep work scholarship is that they haven't really focused on deep work in remote work settings and they also haven't paid attention to the gendered experience of deep work. So in this project we hope to fill these gaps in these different literature is by asking how does remote work shape the experience of deep work for men and women? And if there is a gender gap in deep work, what might underlying mechanisms be? Our context for this project is entry level knowledge work in India and we follow a full cycle research design which combines qualitative theory building with experimental testing. So in the first phase we engage in qualitative ethnographies where we observe real workers as they're working from home supplemented with interview data to generate our hypotheses. In the second phase, we then design a field experiment to causally test our hypotheses. And the third phase we interview our experimental subjects to debrief about their experience and help interpret our findings. So before I move on, let me give you a quick preview of our findings so that you have a sense of where we're going. We find that remote work creates a gender gap in deep work and we measure deep work with respect to the number of focused hours. So what we find is that when women are working in the office, they have 3.5 hours of focused time and this is in a 6.5 hour workday. So that's pretty good, but also the highest among all our four conditions. But when women work on home, they lose about 30 minutes of focus time and more importantly, we don't see this pattern for men where if anything men experience marginally maybe more focused time when they're working from home. I'll also be showing you some evidence that women are interrupted more by family members when working from home affecting their ability to engage in deep work. Now this finding has really important implications for workplace gender inequality in the context of remote work because it highlights this new dimension of inequality, namely deep work. So for the rest of this talk, I'll describe my research methods, then dive into my qualitative findings and hypotheses, then describe my experimental setup, then the experimental findings and end with my conclusion. So as I mentioned earlier, we started this project with qualitative data collection and what motivated us at this at this stage was to really try to understand what does working from home actually look like. And so to answer this question, we conducted ethnographic research where we entered the homes of 21 participants, 10 men and 1110 women, and 11 men working in entry level knowledge work jobs such as IT support admin, QA, work across a variety of industries. And for each of these 21 participants, we observe them from one for one full work from home day where we tried to be a fly on the wall and blend into the background and take copious notes. And what we were trying to observe was how they work, their environmental triggers and their behavioral responses. And actually it was this observation that motivated our focus on gendered interruptions and deep work because we noticed that working from home looked very different for men and women. Now we were silent throughout the working day, but we chatted with the participants over lunch and interviewed them formally at the end of the day. Now we supplemented this ethnographic data with eight semi-structured interviews with men and women where we asked them about their experiences working from home and explored their perceptions of productivity focused flexibility and division of labor at home. Now all of the qualitative data allowed us to generate two hypotheses, which we then tested in a field experiment. And in this field experiment we worked with 205 QA professionals across seven Upwork agencies in India and we randomized their work location. So we randomized whether the QA worker would be working from home or in person for one day of work. And we measured our outcomes firstly through digital tracking to measure focused work time and also through live zoom observation with the help of RAs to track interruptions. And in the final phase we interviewed our experimental participants for one hour to debrief with them but also to contextualize the interruptions and get more data about who the people were who were interrupting them and what was the reason. So with that, let me dive into my qualitative findings. Our biggest takeaway from the qualitative data is that women have very fragmented work experiences when working from home where they're constantly oscillating between domestic and professional spheres, unable to fully separate work from household responsibilities. So in our notes observing one woman we wrote, she gets up at 1104 and walks to the kitchen taking her laptop along with her in the kitchen in observing another woman we wrote, she starts working after a break of over 2.5 hours, but her focus is on the maid in the kitchen doing dishes. When we interviewed one woman, she said during the day my time was split into five minute slots. I couldn't pay, I couldn't plan anything. This was very different from our experience observing men where men seemed to be able to carve out uninterrupted focus time. One woman we interviewed talking about her husband, said My husband has his own room and a door closed most of the day. He didn't have to worry about food or the baby, he would just come out for lunch or tea, otherwise he had full peace. And men also talked about this in their interviews with us, but they had higher perceptions of flow states when they were working from home. One man we interviewed said A mat at home. I'm a nine in the office. I would say five. Definitely more productive here and we have a lot more data that I'm not able to show you in the interest of time. But this leads to our first hypothesis that women experience less deep work than men when working from home. And our next question was, why might women experience less deep work when working from home? And what our data reveal is gender disparities in family interruptions when working from home where women seem to still remain the primary point of contact for household coordination and caregiving needs even during designated work hours. And then family members treat women's work time as more permeable or interruptable than men's work time. And we saw this especially exacerbated when women were living with in-laws. 'cause remember this is India where people are often living in extended family structures and women might also move into a house with their husband and the husband's parents after marriage. So when we observed one woman we wrote, she was on a zoom call with her manager when her daughter walked in asking for a school login without hesitation, she paused, resolved it and rejoined apologizing with a smile. Another woman we observed, we saw her husband came into the room to ask where the detergent was. She stopped her spreadsheet, got up and pointed to the laundry shelf. Now in contrast, men were much less likely to be interrupted during work hours. They seemed to have boundaries that their family members would respect. So in one instance we observed his wife came to the door but waited to knock until his call ended. She whispered a question and left quietly. Now in some instances we were ober able to observe households where both the spouses were working remotely. And we noticed in the same household the man, the woman had very contrasting work experiences. So we noted he worked undisturbed from the study while she fielded three calls from her mother-in-law, two requests from their son and a question from the housekeeper about groceries. In our data we noted that it's not that women were not setting symbolic or Strat spatial boundaries kind of de marketing their workspace. But it's more so that these boundaries were not being respected by the family members. So in one instance we noticed for a woman, despite the headphones and closed door, her son burst in asking if she had seen his project paper. She sied took off her headset and helped him find it. In contrast, those same boundaries when set by the man were more likely to be respected. We observed he, he put a post-it saying Do not disturb on the door and no one entered that room for four hours. So all of this data lead to our second hypothesis that women face more interruptions from family members affecting their ability to experience deep work when working from home. So we then designed a field experiment to test these hypotheses. So we worked with seven Upwork agencies and we selected these agencies based on a couple of criteria. They needed to have at least 50 employees operate under a hybrid work model and express willingness to participate in our research study. Now the first thing we did is we had all the employees at all these agencies complete an online screening test assessing their QA knowledge and skills and only the employees who scored above the 50th percentile were deemed eligible for the study. This resulted in 205 eligible participants who were com compensated by their employer at their regular rate. Now we randomized participants to either an in-person or a remote condition while maintaining gender balance across conditions. So each participant participate for one day of work working as usual with regular supervision structures. So we had a two by two design where participants were in one of four groups by gender and their work location. The in in-person participants work from the agency's office, remote workers work from home. This was a realistic task, manual functional testing, which is a typical individual QA task involving verification of software against predefined requirements. And we developed the software project as well as user stories for QA testing in collaboration with an agency, allowing us to know what the errors in the code were so we could measure performance accurately. Now our first source of outcome data is digital activity tracking. So we use monitoring software pre-installed on the participant's computers to track their work patterns and most importantly to track deep work. And we learned from our conversations with the Upwork agencies that it was really quite common for clients to request installation of specific stuff, software and computers, while engaging for particular client's projects. So the workers didn't really make much of this and this digital tracking data gave us data on focus, which the software measures as working time in which an employee is engaged and working on a single task without multitasking or attention shifts or collaboration activities. And in our pilots, we verified that this focus time was indeed capturing the deep work that we wanted to be measuring. So in particular, our outcome variable will be focused hours in a given day. Our second source of outcome data is ethnographic observation via zoom. So participants were observed via zoom for the entire workday with their video and audio turned on. Each participant was paired with an ra. Now we had the RAs seem as inconspicuous as possible so the RAs videos and audio were turned off and communication was minimal and restricted to chat only. Importantly, the RAs timestamped and classified every interruption that the worker experienced giving us detailed data on interruptions. And at the end of the workday the work the RAs conducted these interviews with the participants to be able to contextualize the interruptions. Now, who are the RAs? They were undergraduate students whom we compensated for this task and we trained them over an eight week period to teach them kind of observation skills. We practiced with them several times and also taught them how to do the debrief interviews. Now the RAs were physically located in one big room during data collection. They were observed by research managers. They could also ask questions if they had anything to research managers. And the managers filled in if the RA ever needed a break and post experiment. The interruption data that we collected was crosscheck by randomized, by randomly assigning the zoom videos to other RAs. Okay, so on the given experimental day we started with 30 minutes of pre-work prep where the participants reviewed protocols and familiar familiarized themselves with the software. They also received their respective zoom links, then work started at 10:00 AM So participants logged into their zoom sessions where the RAs were present and it was a 6.5 hour workday from 10 to four 30 where they computed their QA testing tasks and RAs were recording interruptions. And at the end of the day we had the one hour debrief. So in terms of the measures for our study, our most important outcome variable is focused hours. As I described earlier. We're also able to measure work performance as percentage of accurate QA responses. In terms of independent variables, we have a binary for women, a binary for remote and the interaction term remote times women. In terms of heterogeneity analysis, we look at whether the the particular participant was married, we can also look at total household ME members, which we measure as a number of other people also living in the participant's home. And in terms of our mechanism, we measure family interruptions as count of interruptions by family members. So lemme tell you a little bit about our participants. By design, 49% of them were women. On average they were 27 years old, 26% of them were married and on average they had three other household participants and they on average lived in two bedroom houses in India. Now we then looked at how did these descriptive statistics vary by men and women. And the most salient thing that stood out was that men were more likely to have children. And this obviously affected the number of household members for them. And this actually makes our findings even more stark in terms of our outcome variable. The mean for focused hours is 3.2. In terms of performance, the mean performance was 75%. And in terms of family initiated interruptions, the mean was 1.6 family interruptions. Now since this is an experiment, we wanted to see whether our descriptors vary by our experimental conditions and we were relieved, see that the descriptives were very similar across the in-person and remote conditions giving us rea reassuring us that the randomization occurred successfully. So with that we get to our main effect, which we, where we find that remote work creates a gender gap in deep work. In particular the remote times women coefficient is minus 0.5, which suggests that approximately 30 minutes of focus time are lost by women working from home. Now this represents about a 15% drop in focused work time. And so this supports our first hypothesis that women experience less deep work than men when working from home. And this is a pretty substantial disadvantage 'cause you can imagine that this compounds over weeks and months of working from home. We then explored hydrogen 80 by family composition and marital status and we found that women who are married and women with more household members were especially impacted. And so the additional household members presumably creates more potential sources of interruptions and domestic coordination demands, which affects women's ability to engage in deep work. And then we looked at our second source of data, which was our interruption data. And what we found is that women who are, oh is that back? Okay, perfect. Women who are in the office experience on average, one family initiated interruption, presumably on a phone call. Now this goes up to over just over three if they were working from home. Now men also experience an increase in family interruptions when they're working from home, but just at a much lower scale as compared to women. And when we put this in a regression, we find that the remote times women coefficient is 1.975 suggesting that when women work from home, they experience on average about two additional interruptions from family members. And we validated that this affects our focus time results through causal mediation analysis. And so we find overall support for our second hypothesis that women are more likely to be interrupted by family members impeding their deep work. Now we find downstream implications for productivity. In our case, women working from home are less productive. It's about a 17 percentage point drop in productivity. And we also find negative downstream effects on meaningfulness of work where our, the women working from home described an erosion of professional identity. One woman said in the interview, some days I wonder if I'm really working or just pretending to work between chores. We also find that some women working from home felt that their work was reduced to survival. One woman said there were weeks where I didn't open a book or read anything serious, I was just trying to survive the day. And there seems to be some effect on women feeling like they had to sacrifice career aspirations. One woman said in resignation, only one of us can plan to be ambitious. So just to summarize and wrap up, we find that remote work creates a gender gap in deep work where women lose about 30 minutes of focus time. We find that family initiated interruptions drive the gap where women face about two additional family interruptions, impeding their deep work. And we think this has important implications for workplace gender inequality in the context of remote work. We also wanted to highlight that these findings are probably particularly relevant in context where it's common for workers to live with extended family and possibly live in smaller living spaces. Thank you very much and I look forward to your questions.

- Yeah, go ahead. Thank you Erna. That was an incredible amount of work.

- Thank you.

- So my question was on firm's perceptions about women. So given that firms can, maybe your manager can see that your performing less or your not able to work as much, how do firms respond or, or do firms respond in, you know, how, how what happens from the firm side?

- So from the firm side, we didn't vary anything. So that's why we just said work as usual and that's why we picked Upwork agencies that already had a hybrid working arrangement so that they kind of had routine ways of how to work manage workers when they're in the office or when they're at home. So they just continued as normal. Like they didn't, we didn't vary anything around that. But are you asking maybe if we can collect data on,

- I meant about perceptions, so ah, perceptions and then how this feeds back into, you know, discrimination at

- Work. I don't think we have data on perceptions from the managers at the agencies on whether they thought workers were more productive. We do have a lot of qualitative data, so we can see whether we asked about that in interviews, but I don't think we would be able to systematically get at that.

- Yeah, yeah. Really interesting and fascinating. But I, on the interpretation of the results, what I thought was also interesting is that you actually, the woman dummy

- Yeah.

- Is positive.

- Yeah.

- And so there's something about the in-office work. Yeah. That is also a gender gap. And if I understand actually there is not much of a gender gap in the home deep work, the gender gap is actually in the office. Yeah. Which is, I don't know how you think about that, but that seems like interesting.

- Thanks for that. Yeah, we did notice that. And the way we interpret that is that women, when in the office, this goes back to some of the sociological work that I was highlighting, but hostile has written about how women in the office are just so grateful to be in the office that they can be sometimes very productive because they can put away all of their, you know, domestic demands and personal demands. And so maybe it's highlighting that women in the office like this is a baseline effect that women in the office are more productive than men. But that's kind of being taken away when they are working from home. But yeah, no thanks for bringing bringing that up. Yeah.

- That was so personal. Yeah. Thank you. This is incredible. So I think just to clarify, you, you didn't actually track individuals over time, right? So it's just one time.

- One time.

- I, I think if you could have multiple observations with individual fixed effects, you could balance the men and women a little bit better. And then another comment is more the social, I mean the on the framing. So when the way you frame frame the gap in the literature, the sociological literature didn't look at the quality of work, which actually there's quite a bit of studies looking at the quality of work using time, use time, use studies. But I think for example, in the American time use studies, they can actually look at multitasking, presence of others and episodes of work. We have some research about this, but I think the real contribution along these lines, the measurement precision in your study. So all those measures, they're like self-reported by management, but you actually actively track them over time. So I think in terms of framing, that may be a more accurate framing. Another thing is that I, I just think it's fascinating that you actually can separate the hypothesis from whether disruption is mainly from disruption by family members. Or another competing hypothesis in psychology is women are more disruptive by space. So women need like protective space more compared to men. So, you know, sample sounds like some women, they do live alone. Maybe you could actually test these two hypothesis using the sub sample. Thank you.

- Thank you. And I'd love to get some citations from you afterwards, but thank you. Yeah.

- Yeah. I was curious, I mean this is sort of beyond the bounds of your paper, but I mean something from my own work is like, it looks kind of like when women are in the office in more collaborative types of tasks, they get maybe more demands on their time from others to do mentorship and other things like that. And so not within this paper, but maybe like, I'd be curious about your thoughts on that.

- Unfortunately we can't speak to that at all. 'cause we picked an individual task, but, but I could see that happening and I remember you had some of that in your data from yesterday morning, right? So, but yeah, no, I think where ours is very focused on individual task and even in all the ethnographic work we focused on kind of individual work. So unfortunately I, I don't have much to say on that, but thank you for the question. Yeah,

- Yeah. Your, your basic regression with focus time is the outcome had a very low R squared, which suggests there's a lot of heterogeneity and some of your other results show if I, they went quickly, but if I caught it correctly, it's women with married women and not single women actually went the other direction slightly. So what it would be very useful to see is just a distribution of men versus women on that outcome variable, either ized or not. But it, it looks, it looks like your data may say that most women in your data set look like men and there's a few women who are really driving these outcomes on the folks. And, and that, that's a very different interpretation from everybody. Looks like the conditional mean response in the difference between men and women.

- No, we can definitely split that out. But it is, I would say that I agree with you that it's married women or women who live in large houses who, who are experiencing this rather than we can actually look at someone else's as. But the single women, I'm sure we have at least a handful, but we can look at all of that separately. But thank you for your suggestion. Am I out of time? Okay. Sorry. But I'd love to take questions afterwards.