Do They Add Up? Using Macro Counterfactuals to Assess Micro Estimates and Macro Models

>> Speaker 1: Well, we're very honored to have Valerie speak to us today. The title is intriguing, by the way. Do they add up question mark? Using macro counterfactuals to assess micro estimates and macro models. Thank you for being here. It's so nice you're around.

>> Valerie Ramey: I'm absolutely delighted to be here.

I was getting my slides done at the last minute and I said, what is this? And then I remember the first seminar I gave at Stanford as a graduate student in the middle of the night. I was running off handouts at Hoover because my office was here. So I said, nothing much has changed.

All right, so this is actually an outgrowth of a keynote speech I gave in April for the Royal Economic Society. And the Scottish Society's joint meeting for the 500th birthday of Adam Smith at University of Glasgow. And they also wanted a paper for the economics journal. So my co authors and I have been working on that paper.

It's almost done, not quite. I also gave this as a keynote at the Society for Economic Dynamics in Cartagena. But anyway, it's much expanded now. And so then that's what I'll present. So let me just take a step back and sort of think of the big picture. So, I don't know if you recall, Angrist and Pischke had their journal of Economic Perspectives piece in 2010.

Where they extolled the virtues of the credibility revolution and empirical economics. Talked about all of the great new techniques that applied microeconomists were using and were producing much more credible estimates. However, they scolded the macroeconomists and the industrial organization economists for not adopting these techniques when these were clearly the right way to go.

So they claimed that we were slow to adopt them. And indeed, at that time, the dominant macro methods were either time series, often structural VARs, or quantitative DSGE models. Those were the main things. There were a few macro economists for many decades that use sort of natural experiments, which was an applied micro one of these techniques to estimate parameters or causal effects.

So of course, people have used natural experiments such as wars, also hyperinflations, timing of Social Security checks, a variety of things like that.

>> Speaker 1: Why do you call wars natural experiments?

>> Valerie Ramey: Well, because it will lead to changes in government spending that are typically not associated with what's going on in the economy at the moment.

And as you know, that's a big part of my research is making that.

>> Speaker 1: Just from knowing history, it strikes me as contentable.

>> Valerie Ramey: Well, we could talk about that later, okay. Cuz this is not actually using that identification in this particular talk. Now, the recent renaissance in macro fiscal research, as well as other macro research such as the effects of housing prices on consumption has included the widespread use of applied micro methods.

These methods did diffuse in macro in the year since that 2010 paper. So there are many new natural experiments to estimate household NPC's Bartik instruments to estimate regional fiscal multipliers and panel data and many other examples.

>> Speaker 3: You qualify them as quasi natural.

>> Valerie Ramey: Quasi, that's fine, if you wanna call them.

 

>> Speaker 1: What does that mean?

>> Valerie Ramey: Or unnatural experiment, any of this.

>> Speaker 3: At tsunami all the time.

>> Valerie Ramey: That's fine.

>> Speaker 1: Putatively, I think would cover both what everybody's trying to say at a minimum.

>> Valerie Ramey: Okay.

>> Speaker 4: Can I make one comment?

>> Valerie Ramey: Sure, yes, please.

>> Speaker 4: There's a little bit of a counter revolution in econometrics led by Angus Deaton.

That the claims being made for these methods are exaggerated and the assumptions necessary for these things to hold are quite substantial.

>> Valerie Ramey: You will see in the three illustrations in this talk that I'm gonna make the same point.

>> Speaker 4: Okay, yeah, just to get my.

>> Valerie Ramey: Yeah, I just.

 

>> Speaker 4: Get my views on the table. I think they're an important quiver. But just every paper that says we've estimated the causal effect of this on that.

>> Valerie Ramey: Exactly.

>> Speaker 4: By the way, often in applied micro, when you wouldn't care what the effect of this on that is.

But it's well, they claim it's well identified, but just as a general issue, this isn't settled, doctor and econometrics.

>> Valerie Ramey: No, no, I know, there are a number who do think that this just gives you the credit the estimate and.

>> Speaker 4: For sure

>> Valerie Ramey: Yes, well, depending on what you wanna estimate, sometimes the OLS might actually give you something better.

So in my 2019 Journal of Economic Perspectives paper, I talked about some of the limitations. Which is that one of the things we found out even back in 2011. Which is a lot of these micro estimates, such as using cross state or cross province variation, only gives you relative effects.

All right, this is also the same case with otter, Dorn and Hansen. They could only give you relative effects. And in order to talk about the macro implications, turns out you actually need a macro model to do that, all right? So that's why in that piece I said there's no applied micro.

There's no applied micro, free lunch for macroeconomists. If you wanna answer macro questions, you either have to use macro data as your principal estimate. Or if you use these applied micro techniques, you've got to pair it with a macro model in order to get the macro answer. All right, now, the other caveat is often, and I certainly found this in surveying the estimates.

You can find very different answers from the answers you get, say, using just a DSGE model that's calibrated or estimated or aggregate time series. Versus what you get when you take some applied micro estimate and use some model to figure out what the macro answer is. So what's the theme of this talk?

What I wanna argue here is that we could exploit this micro macro confluence to take the credibility revolution one step further. And in particular, what I wanna argue here is that we should use what I'm gonna call macro counterfactuals that are implied by the micro estimates to assess the plausibility.

I wanna get the new idea, plausibility of the micro estimates and macro models. So why is this tool useful? As I said before, micro estimates and macro aggregates don't always agree. Often researchers who just need an estimate to put in their model or policymakers are faced with a bewildering array of estimates.

And some might be tempted to use the one that serves their purposes best, rather than because one seems to be on a stronger basis.

>> Speaker 1: May they just be tempted to do that?

>> Valerie Ramey: Well, I'm being nice.

>> Speaker 1: You might delete mine.

>> Speaker 3: So there's also an issue of exogenarian magnitudes of shocks.

 

>> Valerie Ramey: Yes.

>> Speaker 3: Part of the credibility revolution. And when I think about a macro model, I don't know what's really endogenous, exogenous. So is this part of your-

>> Valerie Ramey: Yes.

>> Speaker 3: Arguing?

>> Valerie Ramey: Exactly, so the other nice thing about this tool is suppose, and I'll show you how you create a counterfactual suppose that it just doesn't look plausible.

Well, then that helps you look at maybe which estimates, or which models need more scrutiny. Okay, so in all of the reviews I've done of various types of fiscal multipliers, infrastructure, on that, I've gone through a lot of replication files. It takes a long time to delve into the details, and often the is in the details.

All right, we all thought that instruments were so important, but it turns out other things, like just the way you take your estimates and calculate a multiplier, really affects what the multiplier is. So what's nice here is that it says, we should look more at this estimate because there's something that's not adding up.

Or we should look more at this model that translates that estimate because something's not adding up. And then what we found so far, cuz we have our previous paper and then this one, is that the search for reconciliation is often illuminating.

>> Speaker 4: How often are we able to replicate based on public available documents?

 

>> Valerie Ramey: Well, the fiscal people are very careful with the replication files, so I've had good luck with that.

>> Speaker 3: When you say search for reconciliation, you mean that the micro data may be wrong, but also the macro model.

>> Valerie Ramey: Macro model, exactly, it could be and that's why I'm gonna give you three illustrations, because we get a different answer in each one.

All right, so let's go back just to give you some motivation about this macro tension concerning the 2008 us rebates. And I had to see it in here, even when you're not present. Marty Feldstein and John Taylor looked at the aggregate data in 2008 after those rebates were given out over the course of spring and summer.

And they did simple analysis, and they looked at numbers like this, this is just the revised data. So the blue, sorry, the green line is real disposable income, and you can just see that rebate. And the red line is when 50% of the rebate was given out in May 2008.

And then you look at the consumption expenditure line and it has a little bump in it, but there is no spike there. And they concluded that the marginal propensity to consume MPC for short, for the two, 2008 rebate was low. Okay, so that seemed really clear, it was just obvious in the data.

Now, a few years later, Jonathan Parker and his co-authors produced some micro MPC estimates out of that rebate. It was just wonderfully entrepreneurial and foresight on their part to add rebate questions to the consumer expenditure survey, the CEX and the Nielsen Household survey. All right, it was a great natural experiment because the rebates were distributed over time based on the last two digits, the Social Security number, it was basically randomized.

They used the then state of the art applied micro methods, difference in difference, which is basically OLS. You have your fixed effects, they estimated very high NPC's out of the 2008 rebate, 0.5 to 0.9 on total consumption. The majority of the spending was on motor vehicles. When I saw the first working paper version, I said, wait a minute, how can this be?

I said, these guys do great empirical work, but how can this be consistent with what John Taylor and Marty Feldstein had found? Now, because of the credibility revolution, or because they wanted to pick estimates that were convenient, policymakers and most researchers believed the micro estimates and ignored the simple macro analysis.

So what are the aggregate implications of Parker et al's estimates? So a few years after the working paper versions came out, Matthew Shapiro showed me a calculation that he and his co-authors had done. And it ended up in the last part of a 50-page paper in a little table kind of downplayed table 14.

Okay, they looked at what the implied induced spending was by the rebate if you took the Parker et al. Estimates and multiply it by the rebate, and then it was in this table and they kind of averaged it over the quarter, just the induced spending. Well, what I actually did in a discussion years ago, and then what we do in our 2008 paper, we say, okay, let's create a counterfactual, let's take that induced spending and subtract it from the actual.

And see what it implies would have happened had there been no rebate. Now this is a simple micro accounting exercise because it doesn't take into account general equilibrium effects or any effects on the auto industry. It's just a simple accounting exercise. So when we do that, the counterfactual implies that spending on new motor vehicles, remember that was their big MPC, would have dropped 87% in the summer of 2008.

So I have a number of papers on the auto industry. This has never happened, did not happen after layman, I don't even think it happened during the Great Depression. All right, so just clearly something's not adding up there. So what are the key ingredients to a macro counterfactual plausibility analysis?

First, you need micro or subregional estimates of key parameters relevant for macro effects. So for example, the household NPC's, I just talked about fresh labor supply elasticities, firm level supply elasticities. Or say, local city multipliers or state multipliers, any of those sorts of things are using those applied micro types of methods.

You need a policy or event that is big enough to be visible in the aggregate data, okay? Now, those estimates typically are from a study, the micro estimates that happened during that event. But that's not necessary. You could take micro estimates from another period, and as long as there's enough similarities, say it's temporary.

You could also apply it to a policy event in a different period. You also need a macro model that translates the micro or subregional estimates to dynamic general equilibrium effects, or we focus on macro. But you could also do this at the industry level. So for example, you could see if something's big relative to the industry, given some, say, estimate of firm supply.

And then you need a narrative analysis of the time period surrounding the policy or event, because you need to make the case for whether the macro counterfactual is plausible or not. All right, this often requires auxiliary evidence, forecasting equations, lots of reading about what was going on in the time period.

Okay, so I'm gonna illustrate this method in three cases. First, I'm gonna look at two on the micro MPC case. I'll focus more on the 2001 case in this paper, in this talk. The 2008, I'll summarize some of the key findings because it's useful to compare the different sorts of reconciliations.

But that's mostly dealt with in our separate paper, and then also some macro implications of state level multiplier estimates.

>> Speaker 1: Now, in your discussion of literature, you didn't mention a paper by a couple of CBO economists, Michelangelo, and some Somebody. That was about the US, the 2008 case.

I think that was.

>> Valerie Ramey: Okay.

>> Speaker 1: The 2008 case. And what they showed is that there was no such thing as a macro MPC, basically. Because they had data broken out by income group, by how many assets they had. So they had highly disaggregated kinda data where they could do this.

And some groups had high MPCs, some.

>> Valerie Ramey: Yeah.

>> Speaker 1: Had low MPCs.

>> Valerie Ramey: That's okay.

>> Speaker 1: So that was informative. And from policy point of view, maybe I care about the people who have high MPCs. And then just the money that I waste on the low MPC people.

Well, okay, that's just a waste. But it's a cost of getting money to the high MPC people.

>> Valerie Ramey: Right.

>> Speaker 1: But now you didn't mention that study. Doesn't that fit into your.

>> Valerie Ramey: Okay, I have not seen that particular one. But I think you'll see that what we're doing is.

 

>> Speaker 1: The AEJ.

>> Valerie Ramey: AEJ.

>> Speaker 1: Yeah, AEJ Macro. So unless those journals are not.

>> Valerie Ramey: No, no, no, that just had not seen that one.

>> Speaker 1: Okay.

>> Valerie Ramey: Okay, what is their conclusion about the aggregate effects?

>> Speaker 1: Well, they kinda get the. They probably got the average effect to be what the macro say.

But the thing is that the key issues really aren't summarized well by the aggregate.

>> Valerie Ramey: Well, wait till I get to the model and see what you think. Because we are gonna have some heterogeneity there. And then we're gonna talk about.

>> Speaker 1: That was.

>> Valerie Ramey: That heterogeneity. But then how that turns into what we call a GE MPC.

But I'll get to that, so.

>> Speaker 1: I think you have a hang more on some cases.

>> Valerie Ramey: Yes, you could do it again.

>> Valerie Ramey: All right, so let me just give you a little bit of background. There's the Johnson, Parker and Souleles (JPS) for sure, which studied the 2001 rebates.

There's the Parker et al (PSJM) which the studied of the 2008 rebates. Also Broda and Parker, and I already told you this. Each study relied on all these great natural experiment methods and new data creation. Okay, their estimating framework is straight out of the traditional permanent income lifecycle test.

Okay, where people take the change in consumer expenditure. So this is at the household level i. And they have month fixed effects. Fixed effects for each month. They have some control variables such as age and change of household size. And then they have this rebate variable, which can take the form of the dollar amount, an indicator for receipt, or the dollar amount instrumented with the indicator.

So they run it several ways. And the idea is that is for the most part exogenous. Because the timing is based on the last two digits of your Social Security number, although it does cut out at several income levels. But I won't get into that detail here. All right, the 2001 rebate.

So this was the Bush 10-year tax cuts that were passed in early June 2001. The checks were either $300 or $600, depending on if it was joint or married filing. Sorry, single or married filing. Jointly, it went to 92 million households. The total rebates were 38 billion, which was about 6% of monthly disposable income.

So here's a graph of the 2001 case. Disposable income is the green line, the blue is the rebate. The rebate was big. It happened over several months. And you can definitely see it in the aggregate data if you look at that disposable income line.

>> Valerie Ramey: Now, when they do their estimates.

So first let me talk about their baseline estimates, which just looks at the contemporary effects within basically a quarter. So people are only interviewed every three months. And so they often take just the average of the three months over the quarter. They construct something they call a nondurable spending category based on Anna Maria Lousardi's sort of strict kinds of things and broader classifications.

And we spent a lot of time looking through these categories. It's actually a mix of some nondurable goods, services, and even some durable goods. So jewelry and medical equipment happen to be in there. Now, their instrumental variables estimates produce marginal propensity, consume a 0.375 standard error. That's reasonably low.

And that's based on a quarter-to-quarter difference in spending. Now, their preliminary analysis finds a statistically insignificant MPC on total consumption that is less than the MPC's on their subcategories. They attribute that anomaly to noise induced by durable expenditures. Thus, they say they're gonna ignore durables and total consumption.

And just focus on the very strict nondurables in this category, right? So remember that, cuz that's gonna be key. All right, so they also look at dynamic estimates allowing lag rebate to have an effect on the current three months of consumption. And when they estimate this, they get a contemporaneous effect of 0.386, an additional lagged effect of 0.273 for a cumulative 6 month effect of 0.66.

All right, very, very high, all right? So how do we create the counterfactual? So first let me do the micro, just using the micro stuff with no general equilibrium effects, all right? We calculate the induced spending using their MPC estimate of 0.375. We assume it spread evenly across the three months of the quarter.

So that's what the spending is. For the six month estimates, we're just doing the same thing. But then allowing that additional effect from the rebate from three to five months ago. So then our counterfactual is the actual aggregate nondurable consumption using their category minus induced spending. So here's what we get.

So this is for 2001 through early 2002, just so that you can kinda see the whole picture. The black is the data for their definition of nondurables. The blue line is when we just take into account the contemporaneous three month effect. That's with the micro MPC of 0.38.

The purple line is 0.66, all right? So this is based on the micro estimates, no general equilibrium effects, all right? A couple of comments on that. First of all, you get what we also saw in our 2008 paper, what you saw morphico these sharp V shapes. And of course even more so when you allow that other effect that takes longer to get back.

Now, 911, that's there in that September one. And you can see even the actual data go down there. But it goes down quite a bit here. Likely, had some effect on this. We would have expected that decline, particularly since airline spending is one of the categories. But there were already declines in August.

We're gonna, God. Tim knows this cuz we read a whole bunch of the things. So there was something else going on. And most of the analyses were saying there was a quick recovery rebound of spending from 9/11 member. President Bush said, go out and spend. And then there were all these automobile incentives that was actually durables.

But we're gonna analyze this more carefully after I show you the general equilibrium counterfactual, because, as I say, we're not taking into account general equilibrium forces. So, let's think about the macro counterfactual.

>> Speaker 1: Well, you're also ignoring the fact that tax rates, were reduced in. So, basically by the experiment where you only focus on the rebates, ignores the, is an income effect, but ignores the price effects that were also.

 

>> Valerie Ramey: Yes, but in place in a standard new Keynesian model. Yeah, you're not gonna get a lot cuz there's so much sticky.

>> Speaker 1: New Keynesian.

>> Valerie Ramey: Well, because.

>> Speaker 1: For investment decisions and savings.

>> Valerie Ramey: Okay, yes, so, you've led on to something else that we're gonna talk about, which is people, a lot of the literature has treated this as a temporary rebate, because.

But in fact, it was a ten year tax cut. And if households can be forward looking but not completely Ricardian in that they don't take into account the government's budget constraint, and you would expect a higher MPC. But let me leave that till a little bit later.

>> Speaker 3: I can ask.

 

>> Valerie Ramey: Yes.

>> Speaker 3: There are two other issues I see here. So, one is that depending on your expectations of the persistence of the shock, you will behave differently. Yes. But even econometrically speaking, these are linear distributive lag models. If I take even the simple growth model, business cycle version, and I impulse shock, well, depending on the horizons at which the shock itself manifests, I will have a different impulse response.

I mean, locally speaking, it will be differently nonlinear in a way that the typical regressions we run, cannot possibly. So, if I even extrapolate one period ahead, cuz that's what they do, I think they have one diagonally.

>> Valerie Ramey: Yes.

>> Speaker 3: It's not exactly interpretable with the VNMO. Yeah, because first, it's all relative sorts of things, and we actually talk about those issues.

All our impulse responses actually reabsorb, super nonlinearly.

>> Valerie Ramey: They're not, yeah, yeah. Except we actually don't find that in our model. We do things both using shooting algorithms and log linearization. And the results are really similar given the particular rebate check was more temporary, and that's where the estimates are coming from.

You only got it once.

>> Speaker 3: Standard first year model, that if I were to look at the elasticity of response, it's certainly not time constant.

>> Valerie Ramey: Yeah, but it's not that different we found, okay? So, what we do is construct a medium scale, two good two agent new Keynesian model, all right?

So, these are the kinds of models that a lot of policy institutes are using, and these are the ones that they say, give us big multipliers, all right? So we're gonna use that model, but we're gonna show.

>> Speaker 1: Public finance guys have been using much less aggregated models for decades.

I mean, for example, Dale Jorgensen has a multi sector model.

>> Valerie Ramey: You'll see, so, it's my fault that we couldn't get the paper out, because we're really careful. What we do in our model, as you'll see, is we are gonna calibrate the parameters, such that if you take our model and generate data from it.

Jonathan Parker will get his same estimates, and then you can see what's going on at the aggregate. So, what matters is that we're capturing those micro estimates, and then what that says for the aggregate response.

>> Speaker 5: If I could raise a question kind of at a conceptual mathematical level, we all confront this issue of aggregation, exact aggregation.

There's theorems about that going back to Debreu and Sonnenschein, that are unlovely, and there are temporary Hank models or a step forward soiler equations with aggregate, etcetera. But it seems to me that this is a valuable exercise. But kind of the first question you should be asking is, do the macro estimates or micro estimates even bound the true sample mean, for example, things like that?

And I'm wondering, are you.

>> Valerie Ramey: Well, so.

>> Speaker 5: You're making a bridge from the micro, the macro. But is the bridge, is the bridge sufficient, I guess. If you wanna know what the truth is on NPC's or something, some abstract thing called the truth of the. We have some probabilistic thing going on.

We're trying to get an unbiased or consistent estimate of the sample mean are, and we get these vastly different things. Could it be larger than Parker? Could it be smaller than Marty and John? And do we know anything about that? Or we should take these two numbers and try to go back and forth.

 

>> Valerie Ramey: That's actually the point of just looking at the aggregate data, and thinking about the plausibility of any estimate that you put in there. For how big that effect could have been of that noticeable policy during that period.

>> Speaker 5: Okay, I get the plausibility idea, but going back to different groups, etcetera, there's a set of potential NPC's.

And if we disaggregate by income, education, etcetera, etcetera, and maybe we get some estimates of each, but those are just, they're one draw of a sample, and. Yeah. So, the basic question I'm trying to grapple with, is what you're doing. I think what you're doing is very valuable and useful.

I should start there. I don't mean to suggest, but is it informing about anything, about aggregation from micro to macro in general or not. And maybe it's just not possible. I mean, there's a JEL survey by Glendale and Stoker, for example. But, it's kind of a permanently vexing problem that people have been chipping away at.

And you're adding an important part to this. Should I just focus on this, or should I be thinking more broadly about it from what you're doing? Should I be.

>> Speaker 4: What Michael is talking about is called uncertainty quantification, in the computational engineering literature, and it's growing in its use.

And that is, because even people who build nuclear bombs, aren't sure about the parameters, the bombs. And so, they developed these uncertainty quantification techniques. Now, in economics, doing that is forbidden. So, that's the problem.

>> Speaker 5: Maybe the answer is, this isn't, that's a broader, more complex goal that you're tackling right here.

So, that's fine.

>> Valerie Ramey: I mean I.

>> Speaker 5: Try to understand, what I'm supposed to make of this.

>> Valerie Ramey: So we are not focused on the kinds of aggregation theorems that you're talking about. So, for example, one of the things, because we wanted our model to be transparent, so we just do the two agent model.

We've looked at some heterogeneous agent sorts of things, and we find that the counterfactuals become even more implausible. But that's just for one particular variation on a heterogeneous agent model, since there can be so many. But you're right, I mean, you could. But we're not, again, we're macroeconomists and even though we now do heterogeneous agent models, we don't do things at the level of detail that public finance people do.

So I think that's a really interesting question that I would wanna think about more. But I wonder to what extent this technique that we're trying to use here could be extended to the richer public finance sorts of aggregation questions.

>> Speaker 5: There are lots of applications for this, but anybody works with microdata, survey data, for example, census data, whatever it happens to be, is just immediately struck by the immense heterogeneity.

 

>> Valerie Ramey: Yeah.

>> Speaker 5: Yeah, including the heterogeneity, observationally equivalent when we condition on all of our stuff.

>> Valerie Ramey: Right.

>> Speaker 5: And it's not just sample variation, it's clear there's something else going on.

>> Valerie Ramey: Right.

>> Speaker 3: Can I ask you, you're asking also about model identifiability, sample coverage and model uncertainty, which brings to the questions which I think is super fundamental, am I right?

Are we identifying anything that is interpretable by a macro model from these micro exercises?

>> Valerie Ramey: That's also a quip. I mean, I think it probably is, but it's a question of how you interpret it and how it feeds into the model and exactly what your parameter means. Yeah, I mean, and so that's why with this simple system that we're using here, we're trying to get at that to some extent.

 

>> Speaker 3: Cuz consumption responses are not parameters alone.

>> Valerie Ramey: Right, right, absolutely. So we just take basically an off-the-shelf kind of model. In my infrastructure paper, I had extended Jordi Gali's et al's model to update a few things, and I had infrastructure spending at that time. But then we just take that it has the bells and whistles of a medium scale New Keynesian model.

It has sticky wages and prices, variable utilization of capital, investment, adjustment costs. We need two goods, though. Usually those models have only one good. We need two goods in this case, because the JPS categories assume that all spending is focused on their non-durable category, which is only 53% of total personal consumption expenditures.

So we add that second good. We use a Taylor rule, lump-sum taxes. And we assume they respond to debt with a 12-month lag just so that we don't have any taxes changing. And then-

>> Speaker 5: So-

>> Valerie Ramey: Yes.

>> Speaker 5: You're assuming.

>> Valerie Ramey: We're assuming, yes, it's this-

>> Speaker 5: Which of course is totally counterfactual there, because the deficit, the debt level grew substantially during the early 2000s.

 

>> Valerie Ramey: Yes, well, the key is that in 2001, they believe that lump-sum taxes will be raised later.

>> Speaker 5: That they believe that lump sum taxes were gonna be raised later. I don't think so. Ask Glenn Hubbard, I don't think he dare say that we believe lump-sum taxes were going to be raised in the future.

 

>> Valerie Ramey: The consumers in our model, that's all that they have to assume.

>> Speaker 1: To have some assumption-

>> Valerie Ramey: Yes. Let's see if we also put in. But your point about tax rates, cuz tax rates definitely-

>> Speaker 5: You're assuming that basically the debt level is stable, and it certainly wasn't, and nobody expected it.

I don't think.

>> Valerie Ramey: Yeah, but it's really hard to solve a model and make projections when you have government insolvency.

>> Speaker 5: No, now that hurt. Yeah, I mean, the thing is that you can't do log linearization because it's a non-stationary thing. Solving non-stationary models is easy nowadays. Not when you're a graduate student, but things have advanced.

 

>> Valerie Ramey: We do use the shooting algorithm, but what matters is what was going on in 2001. How did people respond to this? And do we think it stimulated aggregate consumption, given those MPC estimates? That's the basic thing. So it's much more short term than the usual long term.

 

>> Speaker 5: So you could show the sensitivity of your results.

>> Valerie Ramey: Yeah, yeah-

>> Speaker 5: To whatever these assumptions are about lump-sum taxes and lag.

>> Valerie Ramey: Yeah, it's-

>> Speaker 5: And so, Cheryl, do you prefer people expect an increase in marginal tax rates in the future, or do you want an increase in default probability?

It's marginal tax We have never had a US default, yet, right?

>> Speaker 1: Okay, so.

>> Valerie Ramey: After the revolutionary war.

>> Speaker 5: After the revolutionary war.

>> Valerie Ramey: Cuz they didn't pay everything.

>> Speaker 5: That was before the constitution.

>> Valerie Ramey: Okay, well, you said in the US, it was still-

>> Speaker 5: Well-

 

>> Speaker 1: Basically, if you're gonna postpone the tax hike.

>> Valerie Ramey: Yeah.

>> Speaker 1: Discounted value of anything further in the future will be less, sort of less impact.

>> Speaker 5: Yeah, but you could test that.

>> Speaker 1: Well, wait, though, present discounted value of revenue has to offset the present discounted value of the transfers.

So isn't it the same? Instead, it's ten years from now, there's gonna be lump-sum taxes.

>> Speaker 1: It depends on what consumers are discounting at.

>> Speaker 4: There could be heterogeneity in the discount, right?

>> Speaker 1: Absolutely, and consumers will be looking after tax, okay?

>> Speaker 5: Yeah, by the time those lump sum taxes are paid-

 

>> Speaker 1: The government collects-

>> Speaker 5: I'll be dead, and other people are gonna pay them.

>> Valerie Ramey: Okay, the reason we do this is because it matters for the hand-to-mouth households. And in fact, I didn't have such a lag in my infrastructure paper when I presented it virtually at Stanford, Adrian Eau Claire made a point, but that's gonna have an effect.

And I said, I should have realized that it was too late for that one. So then that's why we have this delay, so you don't start raising the taxes right away during those sorts of things, okay?

>> Speaker 1: Yes, but the second good, is it a durable good or is it?

 

>> Valerie Ramey: No, in our 2008, it is a durable good. In this case, it is another non-durable. And there's gonna be perfect substitutability and production between the two goods. So the relative prices are going to be constant. And so it just makes sure that we're scaling things properly.

>> Speaker 1: Now, macroeconomists usually write down models that are closed economies.

 

>> Valerie Ramey: Yes, I was closed economy, too.

>> Speaker 1: Okay, and I thought after Feldstein's testimony as chairman of the council being so disastrously wrong in the early eighties, this is before you were paying attention to things. I would think that after that, people would learn to, that the closed economy assumption, which is what he used when he said the Reagan deficits would explode interest rates.

I mean, it's wrong.

>> Valerie Ramey: Yes.

>> Speaker 1: And so this business of a closed economy is just false and importantly false.

>> Valerie Ramey: Yeah, so we don't have big changes in interest rates as a result of these that we're looking at, cuz these are-

>> Speaker 1: You probably, do you even have an interest capital market here?

 

>> Valerie Ramey: Yes, we do.

>> Speaker 1: Okay.

>> Valerie Ramey: Yes, we do.

>> Speaker 1: Why don't you continue?

>> Valerie Ramey: Okay, and so then we simulate the response of consumption to the rebate. That's what's so handy, this model. We can calibrate the fraction of hand-to-mouth households to exactly match that micro. Micro MPC that you would get from microdata generated from our model.

Okay, so just a few things. Pretty standard utility function, although I'm sure Ken will have something to argue with. It's two types of consumption, also hours. So we have non-durable consumption as defined by JPS. And then C2 is all other consumption, S1 is the share of the JPS, non-durable expenditure, and total consumption.

We have elasticity of substitution across the consumption goods, it turns out not to matter, and a fresh elasticity. And then, as is usual with the medium-scale new Keynesian model, labor supply is not chosen by the household, by instead by this fiction of a union. And that's just to get people off their labor supply curves.

All right, optimizing household constraints. Very standard households save in the form of a nominal bond. There's a gross nominal interest rate, R, there's a gross inflation rate, W is the real wage. Their expenditures are on the two consumption goods, and then there are actually, T0 is not transfers, T0 is taxes.

But if it's negative, then it's transfers, and then profits is really income. So there are no distortionary taxes. There are no distortionary taxes, we could, that's an interesting point given that this was not just a tax, I don't think it will matter, but it's very easy for us to look at it.

Yeah.

>> Speaker 1: Back in the 70s and 80s, people were writing papers about how inflation, the tax system is nominal.

>> Valerie Ramey: Yes.

>> Speaker 1: Inflation has a real impact, that's for sure, even if you have the standard money neutrality. So now you're, that's one easy thing. And basically, by the way, the thing is that inflation is also going to affect the local eigenvalues when you log linearizes, I've shown this in a paper years ago.

So, but the thing is that the dynamics are gonna be affected.

>> Valerie Ramey: I mean, this is a small thing, so simplicity, right now, let's just think of this as a temporary tax rebate, which is the way people were doing. You don't get much inflation. These effects are pretty second-order, okay?

 

>> Speaker 3: There's no constraint on asset holdings?

>> Valerie Ramey: On asset holdings, no. Okay, so the hand-to-mouth households, in contrast to the optimizers, follow this hand-to-mouth rule. In steady state, we assume that they had the same after-tax income as optimizing households. What we then do, in order to exactly match know Jonathan Parker's estimates, we assume instead of making them consume everything in the first month, we're gonna spread it over the three months.

And so we have these dynamic marginal propensities to consume and then whatever they don't consume initially, we assume that they just get the same rate of return of what's sitting in their bank account. And JPS assumed that the MPC on other consumption is zero, so we assume that for, for these hand-to-mouth consumers.

All right, so we could exactly match what kind of estimates they're getting. We calibrate to a monthly frequency. As I said already, we spread equally over the three months. The baseline for the contemporaneous is that 0.375 or 0.66 for that six-month specification. And as I mentioned before, the elasticity of substitution doesn't matter because relative prices don't change.

All right, so what do we get? So in the left panel, panel a is the micro that I already showed you. What happens when we start taking into account general equilibrium effects in this new Keynesian model? Well, those v shapes become even more pronounced. All right, so those declines are more, which shouldn't be a surprise because new Keynesian models are rigged to get you amplification.

All right, so any of those sort of micro MPCs are being amplified by the standard Tank model, and here we just have the TG-TANK model. Now some comments on this, all right, as I said, the v shapes are more pronounced. Now, in our 2008 paper, we defined micro MPC, which is the kinda thing people estimate at households.

And then we talk about what we call a GE MPC, which is the general equilibrium response of consumption to a rebate, all right? So that's gonna be equal to the micro plus any amplification or dampening, all right? Now if you have a closed economy and if you have a temporary stimulus, then in general, the GE-MPC will be very close to the output multiplier because investments not going to be changing, there are no net exports.

So it's the same as a multiplier. Now, if the micro MPC is that 0.375, here we just rounded to 0.38, the GE is gonna be 0.5. But as that micro MPC climbs because of nonlinearities, that GE multiplier becomes even bigger. In this case, the 0.66 turns into 1.33.

So is that counterfactual decline plausible? Remember, this is a really important step in our method. So we do comparisons to other declines historically in this category, RGE counterfactual implies a three-month decline of 1.5%. The only bigger declines that you can see in the data that go back to 1959, this is the monthly data are during COVID.

And then this decline after this weird spike in 1960, we spent so much time trying to figure out why there was a spike, we still can't figure it out. But anyway, so it's just sort of something with the data. That's the only other time it was that big.

Now, 9/11 certainly accounts for some of the spending dip in September 2001. But as I mentioned, observers such as Blue Chip and others made numerous statements before 9/11. In fact, some things published on the 10th of September, they were calling the rebate a non-event because it wasn't showing up and spending in August.

That rebate was paid out from late July through late September. Okay, so other things that we do, we'd also done this in our 2008 paper. We say let's do our own forecasting model, behaving as though we're doing sort of real-time stuff. So we have a monthly forecasting model that uses the JPS category of consumption, disposable income, consumption price deflator, gas prices, and the Gilchrist-Zakrajsek excess bond premium.

We find that those variables we had tried all these other variables in 2008, are just very good at forecasting consumption. Once you put those in Michigan survey doesn't matter, all of those sorts of things. So what we call the contemporary model uses information they had since May 2001, and what we say is it assumes that no one realized that the economy was already in recession because the recession had started a few months before.

The pessimistic model, where we can get a more pessimistic forecast, assumes that the economy was already a recession, and then also takes gas prices as exogenous rather than responding endogenously. This actually significantly reduces the forecast path, particularly since the 2001 recession was much more mild than the average recession.

So putting that recession dummy in there really lowers that path.

>> Speaker 3: Can I ask question for the expectational high false specifications that you find more plausible? What margin of response to transparables would need to be shut down for the macro?

>> Valerie Ramey: What margin of response would need to be shut down for the macro model to get closer to the macro model?

So other thing, I mean, so for example, there's the usual levers. So for example, we could change the parameters on the Taylor rule. We could make it harder to vary utilization, we can do things with labor in all of the usual.

>> Speaker 3: But that's not enough, I see.

 

>> Valerie Ramey: Yeah, but I'll tell you why we didn't pursue that when I get to the estimates.

>> Speaker 4: Okay, this is consumption.

>> Valerie Ramey: So this is consumption, the JPS category of consumption. The black line is still the actual. The, let's remember, the purple line is that micro counterfactual, the green line is the macro counterfactual.

So you can see that amplification there. This is just for the contemporaneous case because otherwise the graph gets too messy. The blue line is a contemporaneous forecast. If you don't know you're in a recession, all right, generally you're going to forecast an increase in consumption because consumption almost always increases.

But the pessimistic forecast that just takes into account that they're in a recession gives you the red line. Now, notice that the red line is not that different from the black line, but it's definitely quite a bit above those two counterfactuals that we have there, the micro and the macro.

So we consider those v shapes implausible. So the question is how to reconcile micro and macro, and this is a case where we say, okay, let's go into more details in the data. There are a lot of details in these data. So the reconciliation with either smaller micro MPCs or GE dampening, general equilibrium dampening rather than amplification.

We examine the micro MPCs based on the CEX data. Again, this just lists what you could do with the tank model, but you'll see, once we've looked at these data, we figure we don't need to mess around with the model. All right, so the JPS non durables categories, remember, it included many non durable services and some durables, but it's only 61% of the Bureau of Economic Analysis sum of non durable goods plus services.

So let's compare the MPC for their categories versus the BEA categories. So we go to the replication files, but of course we have to draw the CEX data ourselves. So we quite close, not a perfect reproduction, but very close to what they get. So first, look at the JPS definitions.

So this is looking at the MPC out of the rebate, strict non durables, which was more like food and those sorts of things, and then the non durables, so look mostly at the non durables category. So they got 0.375, we get 0.32, but that's pretty similar. But, if we instead say, well, let's look at the BEA definitions, non durable goods and services, look at those as separate categories.

We're getting quantitative and statistical zeros, 0.06 on the non durable goods, 0.03 on the services. When we look at total consumption, we find what JPS mentioned in the footnote is, if you look at total consumption in the CEX, you get an MPC of zero. All right, why is that happening?

Because it turns out there are a bunch of categories with negative and sometimes statistically significant MPCs. So, maybe there was some shifting around, but you can't just set it equal to zero for some categories and then claim a significant MPC for the categories that are getting you the nice positive numbers.

So.

>> Speaker 1: Example of assumptions producing as a result.

>> Valerie Ramey: Yeah, well, you can see they had this beautiful natural experiment, new data were created and sometimes no matter how much you torture the data just doesn't want to speak. These data didn't want to speak, I think, was one of the issues.

 

>> Speaker 1: They did speak and they didn't want to listen.

>> Valerie Ramey: Well, yeah, sometimes the data are trying to tell you things, I've had that case where I said, no, don't tell me that, eventually you listen.

>> Speaker 4: I don't understand, were GPS cherry picking here? Were they just trying to find data that showed what they wanted?

 

>> Valerie Ramey: No, they thought that the durables, there was just so much noise in the durables category, that's why they didn't do total consumption. And their definition of non durables did build on work that Anna Maria Lusardi had done for other things because the worry that some of what we're calling non durables really have a durability component, I mean that made sense.

But then, yet in their category, when we started looking at, they had jewelry in there. So it wasn't clear how those categories.

>> Speaker 1: MPC, if the standard air exploded because of the durables, I could see the logic saying, it's just not informative. Because that durable, were they really big, the standard errors for that?

 

>> Valerie Ramey: Yeah, they were pretty big, they didn't explode.

>> Speaker 1: So like if they,

>> Valerie Ramey: So I need to remember.

>> Speaker 1: 0.4 or.

>> Valerie Ramey: It has been a while since I've looked at the total. Well, yeah, but the point estimate was near zero. And so then we said, well, it's not even worth doing the counterfactual because if it's zero, it's equal to the data.

But, but I could easily look at it, we're still doing.

>> Speaker 1: Now, of course, there's what the multiplier is multiplying, that's not very big.

>> Valerie Ramey: Yeah.

>> Speaker 1: Okay, yeah.

>> Valerie Ramey: Right.

>> Speaker 3: So, going back to your initial statement, in some way this is not a robust paper, the Jonathan quote, right.

So it's not that the macro model, I mean, the macro model was the tool to find that is not robust, but in some way, there are other ways you could have done it by redoing the categorization. Is that fair?

>> Valerie Ramey: Well, I mean, that's the whole idea. We use the macro model because if we want to compare things to aggregates, we need to allow for general equilibrium forces, okay.

Because there's a possibility of dampening issues.

>> Speaker 3: So the alternative is to do, the micro was right and the macro was not taking into account.

>> Valerie Ramey: It's not taking into account, so that's why we go, and when I do, well, if I have time on the 2008 rebate, I'll show you that actually the model had an issue too, because durables were important there.

So in each case, we're finding a different interesting answer, this is a little bit more mundane. We have fancy estimates in the 2008, but as I tell you, most the time, it's not the wrong instruments, it's the devil's in the details. And so that's why going through replication files is so important, but one has only so much time, and that's why finding this disconnect is a useful thing to figure out where we should look more.

The other thing, as Ken mentioned, and then in the process of doing this, we figured out, my goodness, the 2001 rebate wasn't a temporary stimulus, it was an initial payment on a ten year tax cut. If even, forward-looking optimizers, if they don't take into account the government budget constraint, you should have an MPC of, say, approximately 0.33, if they're not Ricardian.

Now, if they're full Ricardian consumers, then it should be close to 0. So our conclusion is well, the micro estimates are not robust, all right? These are nothing useful estimates for guiding our models. And they're so imprecise that they're not gonna shed light on the macro model. It's not the kind of experiment people want, so we don't even revisit the model.

Doesn't mean that the model's right it's just, we don't need to do that.

>> Speaker 1: We're kind of following Ken Judd's thought of using what actually happened, yeah. Assuming people had perfect foresight and they were in a couple of papers on that would actually happen. Another way to divide the consumers up into those who were excessively Ricardians, thinking their taxes were gonna go up more, and those who assume that they get no hit because all are gonna be taxes on higher income people and the like.

So that's another way to think about it. Certainly Those taxes in particular are distortionary.

>> Valerie Ramey: Right.

>> Speaker 1: And since interest on government bonds are taxable to households, but the government doesn't pay taxes, there's this big debate about whether households and the government should be using the same discount payment

>> Valerie Ramey: Yeah, no, there are all kinds of things.

I mean, so certainly, if we had found robust micro estimates, then that's when we'd really have to start tearing apart the models and saying, all right, this macro model is too stylized, we need to do something. Okay, so let's see. I think I've already told you everything. Yeah, typically, factoring in general equilibrium forces amplifies the problem.

And re-examining the micro estimates reveals the non-robustness of those results. And it's not even a good experiment for looking at temporary effects, right? Let me talk a little bit about the 2008 rebate that's in this other paper. It was even bigger, and you saw some of the graphs I'd shown you before it was passed in February 2008, distributed May through August.

The average rebate was $1,000, and these rebates were temporary. So it's a really nice, clean experiment, okay? And this is based on our first paper in this ongoing project, which is the case of the 2008 rebates. I already showed you that counterfactual for the Parker et al. So then we do a GE counterfactual analysis of this rebate and we'd learn different things here, okay?

We have a two-good, two-agent new Keynesian model. But now we have nondurable and durable goods. And the reason we need to do that is because Parker et al in that paper found that most of the spending was on durables, particularly motor vehicles. So in our model, we interpret the durable good as motor vehicles.

Now, we argue that the high micro MPCs imply implausible macro counterfactuals. We offer a detailed narrative of the events in spring and summer 2008, and we compare the counterfactual path to the professional forecasts and our own forecasting model. Now, few details of the calibration of this particular two-good model.

We match micro estimates of durable demand elasticity. This turns out to be absolutely primary importance, okay? And I'll talk about it in a second. We set the fraction of hand-to-mouth households again, to match the Parker et al estimates, we set the motor vehicle MPC at 0.4 that's what he had gotten.

In our baseline model, we say, okay, let the relative supply curve of durables in terms of nondurables be infinitely elastic and match the size and timing of the rebate. When we do that, our counterfactuals just look crazy, right, the data is the black line, the micro MPC for all of consumption is the blue, the purple, because they had given a range up to even 0.9.

And the way we happen to have graphed this, it's off the graph, which is effective way of showing how pronounced that V shape is. Okay, how do we reconcile this? Well, several things, first of all, we go back and re-examine the micro estimates in light of the new econometrics, all right?

So, when they were doing their estimation, they were using state-of-the-art methods. And then this well, big econometric literature came out that figured out that there were all these issues with different methods, all right? So, we used some of theirs plus we found other sources of biases and used the latest econometric techniques, including our own.

And when we correct for those biases, the MPC decreases by 40% or more, all right? So here's a case where they were doing everything right given the state of knowledge when they wrote that paper. But now we know that what seemed so obvious wasn't. And I don't have time to go into the econometric details.

The MPC on nondurables is economically and statistically about equal to 0. So, we get those MPC estimates down, but we still have counterfactuals that are somewhat implausible. All right, so this is a case where we actually have to work on both fronts, the estimates and the model. So we modify the macro model to allow for general, or you could almost call it partial equilibrium dampening.

What do we do? We have an upward-sloping relative supply curve for motor vehicles, all right? So we actually get crowding out and this is based on really careful calibration to the leading estimates from elasticities and those sorts of things. What we get is results from our model that are consistent with the increase in the relative price of motor vehicles during the period.

So, the relative price of motor vehicles goes up like 1%. And you see it when that rebates going out, it goes up. And when we look at narratives of what was going on in the auto industry, there were all kinds of interesting things going on. There were shortages in the smaller cars, which is what most of the rebate recipients wanted plus oil prices were going up.

So that was also decreasing demand. Now, if we had to have a more complicated consumer problem in this model, because if you just put durables in, as usual, you get such high elasticities that you would just have this huge decline in the demand for durables. So, we figured out a nice way of putting sort of Calvin adjustment costs that are consistent with the micro estimate.

So, we spent a lot of time on that one, and our model produces that same change in the price. When we do this, with this less elastic durable supply, it still has elasticity of 5, all right, so we just went from Infinite to 5 here. Then our micro counterfactual, using our new micro MPC estimates that remember had been reduced with the thing, looks like the blue line which we think is completely reasonable.

We cannot argue that's implausible, but it means that that GE-MPC, which remember is about equal to an output multiplier for the 2008 rebates was probably less than 0.2, all right? So a very small multiplier and guess what? That adds up to the kinda data that you and Martin Feldstein were looking at.

 

>> Valerie Ramey: So we learned a lot of lessons from this, so heterogeneous agent models are the real hot topic in macro. And what we're arguing here is in some cases, heterogeneous goods can be as important as heterogeneous agents and particularly when it's durables. Both overall MPC and the distribution of spending across durables versus nondurables really matter for the GE outcome.

If we calibrate the MPC to the 0.34 our new estimate, but assume it's a one-good nondurable model, we still get implausible estimates. Because in those cases, you don't have that high intertemporal substitution that you get with durables for the nondurables. So we actually have to make that distinction there with durables.

And again, the reconciliation implies a really small multiplier on temporary rebates.

>> Speaker 1: So speaking of intertemporal substitution and rebates and fiscal policy, so there was a program called Cash for Clunkers.

>> Valerie Ramey: Yes, it's huge.

>> Speaker 1: It's a huge effect of people turning in their cars and then demand collapsed two quarters later.

And I think studies showed it was ten times the EU carbon trading price for the emissions reductions. But that's something that's also, I think very informative.

>> Valerie Ramey: Yes.

>> Speaker 1: It's almost perfect, just shifting.

>> Valerie Ramey: Yeah, we actually had in the 2008 rebate paper, which I've just summarized briefly here.

Yeah, we have a graph that shows what was going on with motor vehicle purchases after the rebate when it was supposedly the versus what was going on with the Cash for Clunkers. No, there's just a huge difference there absolutely.

>> Valerie Ramey: So 3rd illustration, so this is at a different level of aggregation this is state-level ARRA multiplier estimates.

So we're moving to 2009 finally, all right? So Gabe Chodorow-Reich had a wonderful synthesis of the cross-sectional state-level ARRA literature. This is his piece in the AEJ Economic Policy and remember that was passed in 2009. Now, people were getting very different answers, Dan Wilson would get, how many jobs created per 100,000?

Somebody else would get something else. And we all thought it was because they were using different instruments based on different aspects of the formulas. But what Gabe showed was when you put all this together, when you standardize, again, the devil in the details. Exactly how you're estimating things, doing the lags properly, that all the instruments were giving the same answer, all right?

So that same answer with that which I think his good econometric model was about two jobs were created for every $100,000 a federal dollar spent in a state. Now, using theoretical insights from Farhi-Werning, he argued that the state level multipliers, remember, they're only relative multipliers. To go to the national level, we need a macro model.

Using the Farhi-Werning-

>> Valerie Ramey: He was arguing that the state-level multipliers were a lower bound on the aggregate multipliers. Now, he also converted the jobs multiplier to an output multiplier and it's about equal to an output multiplier of 2, all right? So big multipliers, that's very different from the multipliers that I get in aggregate time series data.

 

>> Speaker 1: Or just taking the average effect of the total spending divided by the change in jobs, which be more like-

>> Valerie Ramey: You'll like the graph here.

>> Speaker 1: Expensive here.

>> Valerie Ramey: So here, creating the macro counterfactual. What we did was, we just refined a macro counterfactual I created, actually, a 2017 discussion at the Peterson Institute on the, I forget what the volume was called.

We use Chodorow-Reich's estimated impulse responses of employment, using this cross-state variation and his estimates of how much the ARRA was spent by December 2010. The reason that he stops in 2010 and that we do, is that there was additional stimulus, which then confounds the effects later. So we're gonna look at 2010.

We created a counterfactual unemployment rate by adding the induced employment to the actual number unemployed. We're implicitly assuming that there's no change in labor force participation, which is, I think is a reasonable assumption given how little cyclicality there is in labor force participation. So here's what we get, so the unemployment rate, the actual was the green, it got very high, it got to 10%.

But if you take his numbers and take his model for taking it to the aggregate level, it implies that had there been no ARRA, the unemployment rate would have risen to 15.5%. And then stayed up near that level for quite a while. As I said, we only take it to the end of 2010.

That's quite a striking number, is it plausible? So how to assess plausibility and this is an illustration how you have to be creative for different case studies that you're doing. So what we decided to do was compare the counterfactual rise in unemployment to what happened during the first two years of the Great Depression, all right?

So what we do is we match up what we call the signature financial crises at the start of both periods. There was a stock market crash of October 1929 and the failure of Lehman Brothers in September 2008. So I'm gonna show you a graph of what was going on with the unemployment rate and the counterfactual unemployment rate relative to the month of the crisis, which is Month 0, all right?

So Month 0 is October 1929 for the Great Depression unemployment rate and Month 0 September 2008 for the ARRA. All right, so there's the graph, so the green line is that same actual line and this shows the change in the unemployment rate relative to that month, okay? So relative to that month, the actual unemployment rate in 2009, 2010 rose to I think it was by 4% points, okay?

 

>> Speaker 1: I think the time is 0.

>> Valerie Ramey: What? At 10, yes, but relative to where it was because remember it'd been rising since May 2007. Okay, so this is the additional after Lehman Brothers, how much it had risen. The red orange line is that same counterfactual that I showed you on the previous graph, remember That had gone up to 15%.

The rise relative to Lehman Brothers is almost ten percentage points. The purple line is the actual change in the unemployment from 1929 to 31 using the historical monthly data that Sarah Zubairi and I created for our JPE paper, okay? They're basically going up by the same amount during those first 18 months.

According to this, the counterfactual implied by Chattelroy Reich's estimates versus the actual Great Depression, okay? So it says that the rise of the unemployment rate during those first 18 months would have been as great as the Great Depression had there been no air all right? Okay, well, I did show this at this conference where Ben Bernanke.

 

>> Speaker 1: So modest, drove to GDP that it's kind of ridiculous.

>> Valerie Ramey: Well, so policy comparison to the Great Depression, besides ARRA, the responsive policy was very different across the two periods, okay? We all know during the Great Depression the Fed began raising the discount rate in 1928 cuz they were worried about equity prices.

They lowered them after the 1929 crash, but then started raising it again in fall 1930 because they were because of the whole gold standard pressures. And then the Fed allowed the nominal money supply to fall, we all read Friedman and Schwartz. In contrast, during the Great Recession, the Fed lowered the federal funds rate.

Starting a year before Layman, the Fed drove the funds rate to zero and then implemented new methods to provide liquidity and stimulus, there was TARP, the money supply increased dramatically. So, what we conclude is the belief that without the ARRA, the Great Recession, unemployment rate would have risen by as much as the first.

I should say 18 months of the Great Depression requires that one also believe that monetary policy doesn't matter because we're talking about dramatically bad monetary policy. And then we're talking about a chairman of the Federal Reserve during the Great Recession apologized to Milton Friedman for that bad monetary policy on the part of the Fed.

Ben Bernanke, during that famous quote, who did everything possible to prevent a great Depression. And yet that counterfactual suggests that the unemployment rate would have gone up the same amount.

>> Speaker 1: I mean, related point, I asked Bernanke this once at a conference and he said yes. Would they have been more aggressive with unconventional policy if it hadn't been for the ARRA?

And he said yes.

>> Valerie Ramey: Yeah.

>> Speaker 1: Yeah, so again, that's controversial, whether those unconventional monitor policies have been real effects, but it would reinforce this point.

>> Valerie Ramey: Yes. So how do we reconcile the state estimates with the macro counterfactual? Well, in my Journal of Economic Prospectus paper, I argue that the state level estimates are not nationally representative and therefore shouldn't be directly aggregated up.

So, those studies use per capita variables, so each state is weighted equally. That means that North Dakota, with less than 700,000 people, gets the same weight as California with 36 million, okay? If there are heterogeneous treatment effects, that could be putting too much weight on North Dakota. Most studies also don't take into account the possibility of induced state government spending as a result of those transfers from the federal government.

But then there is a study, somehow it didn't get into these slides. Leduc and Wilson argued that there was induced state spending. So what I did in that paper was, first of all, I just replicated Gabe Chattarroy, Reich, Sesame, with this wonderful replication files. As I said, the fiscal people always have very, almost always have great replication files.

Then all I do is just reweight and do population weighted estimates to see what it looks like, and it drops from two to 1.15, the standard error goes up there.

>> Speaker 1: How about wealth weighted?

>> Valerie Ramey: Yeah, whatever weighting you want. The third column also does the population weighting, but takes into account total spending, right?

Cuz if you're talking about a multiplier, you wanna count any induced spending by state and local as well. So I take federal and state and local, and low and behold, the multiplier drops to 0.9 or the jobs created, which is basically the multiplier. So a rough calculation, we wanna do this more carefully, suggests that the counterfactual from using the 0.9 but using the farhe warning kind of theoretical modelling for aggregating up, suggests that the unemployment rate would have risen to about 12%, with no ar eight.

That's very preliminary, still significant, all right? But not 15.5%. And so we're currently looking at re estimating the full path with this. Yeah.

>> Speaker 1: Useful, so there was an ex-ante and then real time set of forecasts being done by Bernstein and Romer and Romer, etc. When all this went into effect, and whatever the effect was in the first two years, the gigantic difference is how weak the recovery was, the weakest since World War II.

We can argue what the argument thinks about that, but if you look at the difference in job years between their estimates, the actual turned out to be 20 million. It's mostly that they assume the economy gets back to 5% a lot more quickly. So there's kind of a flip side of this, and maybe it was done then, affected the flip side, and you could argue monetary policy change.

We had QE, one, two twist, etc, all this sort of stuff. But it seems that's kind of the whole story should involve something about that as well, it seems to me.

>> Valerie Ramey: Right.

>> Speaker 1: And the whole story in evaluating policy dealing with recessions should deal with the aftermath.

I mean there's this kind of fun paper by Harold Ullig in the ARRA that's looking at the present value of expenditure multipliers. If you merge into a growth model and have to deal with the higher taxes later, if we're not in a negative, it turns out it's negative.

 

>> Valerie Ramey: Yep. No, I know that.

>> Speaker 1: And then they're all, Michael Woodford, it depends on what people expect about taxes. And it could be negative anyway given when they expect taxes to rise and how much, etc. So it seems to me that this is great, but I'd like to hear the other side of this at some point from Valerie.

 

>> Valerie Ramey: On the slow recovery?

>> Speaker 1: Yeah, or what happens after the initial. If you're focused rightly for the time being on dealing with, on the initial stuff, but there's an aftermath and the aftermath is not, probably not independent of what policies were effective to do this.

>> Valerie Ramey: So I can advertise my co author Johannes Wieland's Econometrica paper and AR Insights paper with Alistair Mackay that talk about.

So our durables model comes from their paper because what they find is a lot of those policies just pull demand forward and then you get the dip afterwards. Yeah, so I can highly recommend those papers because they're making exactly the kind of point you're Making and as you say, Harold Uhlig's paper about they need to raise taxes, and no, it's not going to be lump sum, it's going to be distortionary, and then you get those negative effects.

Absolutely. And yeah, so our idea is, we've got these great new tools in our box, they're, I mean, they're not useless tools, they're useful. But it's particularly nice to have macroeconomists using them, because we tend to think big. And looking at what's going on in the aggregate and making sure that what those micro estimates are doing, along with the macro model, is adding up to what we see in the aggregate, or is it producing implausible counterfactuals.

So we can assess that plausibility, I've illustrated it using three examples from our work. And as you see, the search for reconciliation can lead to new insights about things you need to do with the model. Such as durables can be really important, say, in a model on better estimates of the micro parameters.

And you can think about using this for natural experiments in other countries, as well as for other questions, and you can even think of partial equilibrium questions like industry equilibrium, so.

>> Speaker 1: Now, some of these observations are obvious, like it's dumb to talk about ARRA without talking about monetary policy, that's hardly a big insight.

Now, by the way, regarding 2008 and thereafter, have you seen Casey Mulligan's book?

>> Valerie Ramey: Which one?

>> Speaker 1: Redistribution Recession, I think is the name, because he goes through a lot of micro style little facts that talk about the impact of various micro things, unemployment, etc. So that bears something you should look at.

Now, I'm trying to think of why is Ge different than the micro stuff? And see, the thing is, your micro regressions, or like I said, the Michelangelo somebody paper is one I'm more familiar with. Basically, they kind of assume that expectations about the future are unaffected by any policy.

Now, when you put g general equilibrium in your general equilibrium thing, what you're really doing then is replacing this no change in expectations with the general equilibrium change under perfect foresight. You said you're doing shooting, so you're doing perfect foresight. And so now I guess my question is, why would I trust a model that assumes that people have perfect foresight over a model that assumes people basically don't think that there's gonna be much change in anything.

And so the way to go, I would think, is stay with sort of the micro data kind of thing, but then actually try to ferret out what are the expectations. Because this MPC thing is really talking about how individuals react to a change in policy. And that's how they react to the rebate today and how they react to the changes in future things, interest rates, tax rates, etcetera.

And that's what is goes into those individual decisions, and the perfect foresight general equilibrium way to deal with this expectation question that is unrealistic in many ways. And like I doubt many people, particularly people who don't have a whole lot of money, really can solve through, in their mind, ricardian equivalence kinds of options.

 

>> Valerie Ramey: So remember, our hand to mouth consumers are pretty myopic, so they don't hand.

>> Speaker 1: But I don't think the thing is that the number of people that are hand to mouth is not fixed over time, sometimes people are in a hand to mouth state and then they leave the hand to mouth state.

So you don't want the analysis to read up, revolve around a large fraction of people constantly being hand to mouth. So again, that's the expectations. And so that's the difference, right, between the micro and the ge is that you pin down, you assume that the expectations are that determined by a perfect foresight equilibrium.

That seems to be the difference between the micro and the macro.

>> Valerie Ramey: Well that's not the only difference, it's also the standard new Keynesian amplification.

>> Speaker 1: Yeah, yeah. Well okay. By a general collision thing.

>> Speaker 4: I wanna have some other questions before we go to.

>> Speaker 5: Yeah, I have a couple of other observations and if they're useful, and one question and one kind of fun anecdote about people wanting details.

So the observations are that there's a lot of literature about federal spending offsetting state, local spending. Now the argument during the rebates and hand, well, so much the rebates, but the ARRA and all sort of stuff is because of 44 of the 50 states have balanced budget requirements.

They're gonna have to, they're not gonna respond and their revenues are gonna fall and then we're gonna have to make them whole. Or we're gonna have to offset that somehow with greater spending or sending them resources. That of course seems to have not happened this last time during COVID when they basically just saved all the, they didn't need the money.

But that's something that's kind of worth thinking about because state and local government spending and taxes in aggregate are 60 70% of the federal level. The hand to mouth consumers, it's not so much the percentage of consumers, it's the percentage of consumption, right?

>> Valerie Ramey: Yes.

>> Speaker 5: That's what you're interested in.

So I was wondering, maybe you or Pete, is there any recent studies about that? Remember hall in Michigan had this paper that it was 80, 75% Washington nut was longer term and only a quarter was handy. There's more research on that, anybody have any. Is there a current going, conventional wisdom among macroeconomists about what that percentage is or what percentage of consumption is not hand to mouth, or conversely, is hand to mouth?

 

>> Valerie Ramey: Yeah. Of total consent, so Jonathan Parker and several others have recent papers trying to look at who are the hand to mouth, and then there's another one I'm trying to look.

>> Speaker 1: Bills and.

>> Valerie Ramey: Yes, Bills, so there are some reasons, but do you remember what the answer is?

 

>> Speaker 1: The answer is it's not low liquidity people.

>> Valerie Ramey: Yeah.

>> Speaker 1: It's like trying to remember what the category was that they thought was distinguishing it. Yeah, I thought it was something like how permanent their labor supply is or something, it was some unusual thing that was indicating whether they're likely to be hand to mouth.

 

>> Speaker 5: I think the casual commentary was people were shocked, shocked at how many people seemed to have so little in reserve when they got laid off. So that's, I think, some commentary on that.

>> Valerie Ramey: Right.

>> Speaker 5: And then Tom Sargent spent more time than he probably wanted to trying to convince me there was a large difference depending on whether there are really large changes to anything.

And since one of the debates has been, I think, with what I think I understand probably erroneously, is that some of these were just way too small. There was an argument from some economists. Well, ARA was, wait, Paul Krugman, if you wanted to, Former economist, you can take Paul Kruben.

Well, actually, I think that's a fair point in any event.

>> Speaker 5: So the question is, if it had been four times as large, would that have been? And Tom seemed to be big on regime shifts, and he ends up for big inflation, all this sort of stuff. So he had this kind of notion about that.

And then we interviewed somebody for a job market paper, his name is Scapes, and we didn't wind up hiring him that had done some empirical work on this. Was there any recent work on anything like that among empirical macroeconomists?

>> Valerie Ramey: Yeah, so, well, there's nice theoretical work starting with Kaplan and Violante and econometrical, where the size does matter and the MPC is actually bigger for the smaller size.

Because if that rebate is that temporary income is bigger, then you actually pay the fixed cost of adjusting your portfolio, and so your MPC will be smaller. Which counters Paul Krugman's argument, saying, we just need it to be bigger. Well, not necessarily, it could have actually had less impact.

 

>> Speaker 5: I'll take a look and that's what I was kinda looking for, just one cute anecdote. So in Iowa, CA chair, we had a brief, mild recession. And at the time, senior Bush being a Republican, the Democrats had 58 senators and 275 congressmen. So we didn't wanna do anything, say anything to Capitol Hill, because we didn't, first of all, I didn't believe that spending would do a whole bunch of good for a short, brief recession, a mild recession.

But in any event, and John Taylor, I think, believe that, too. So what we designed was the change in withholding. So we didn't need the Congress to go along.

>> Speaker 1: Yeah.

>> Speaker 5: And we announced this cuz people were being over withheld and over time, moreover, withheld. And so we changed the withholding schedule, which moved some disposable income forward.

And I can't tell you how much time I spent responding to inquiries from every economist in Wall street. All the macro, everybody in the blue chip was calling and saying, what's the exact details, etc? Because they were trying to get the details into their models. So there's some people, maybe not the way you think of it, but there are some people who actually obsess over things like that.

I was kind of shocked at that. I thought it was very small, but he could say he did something that was the basic idea was to minimize the damage for anything he did.

>> Valerie Ramey: The fact that we don't do the small details, it doesn't mean we don't think they're important.

It's just we're trying to do a sort of simple, transparent, stylized model to make the point and then it tells us which cases where we need to go into the details.

>> Speaker 5: Yeah, what parameters are important.

>> Valerie Ramey: Exactly.

>> Speaker 6: Valerie, I just wanted to say I learned a lot, it was a really lovely presentation.

But if you were to testify before Congress, Senate, President, privately, CEA chair, whoever would your advice about doing a stimulus change now after this research and what would it be?

>> Valerie Ramey: I think, well, you wanna get better estimates. If you can get them to spend it mostly on non durables, you'll get a bigger aggregate effect because you don't get that relative durables price kind of eating up.

Yeah, but you can't do that. So you could think about.

>> Speaker 6: I'm just wondering, targeting and say don't do a stimulus.

>> Valerie Ramey: Well, you have to be careful, I'm not seeing a lot of bang for the buck. Now, I think that there are a few cases where you might want to give out checks.

So certainly during the lockdown to COVID, those were like life preservers simply because people-

>> Speaker 5: There's a strong humanitarian case, whatever the economics were.

>> Valerie Ramey: Exactly, and politically, the fact that the government was not letting you go to work and do all of these things, it was incumbent on them to try to do that.

But all of those follow up ones, the one in December 2020 and then the one in March 2021, just did not make any sense to me.

>> Speaker 1: Well, it really made no sense, it's about a year ago I got a check from the state of California.

>> Speaker 1: Now that was totally absurd.

 

>> Valerie Ramey: It's because their schools are already so well funded, they have sky high test scores. So what else are they gonna do with their money?

>> Speaker 1: I told Teresa that can't be true.

>> Speaker 1: But no, it was true, we get it, California.

>> Speaker 6: Thanks again.

>> Valerie Ramey: All right.

>> Speaker 4: Thank you very much.