Friday, November 18, 2011

More on the interstate/county comparisons and the multiplier

Brad DeLong linked to my earlier comment on the county-level multiplier estimates, arguing that what they provide is a good lower bound estimate. To a certain extent I agree with him. I would actually use this sort of argument a lot in project meetings at the Urban Institute. You are never going to completely rid yourself of endogeneity and bias, but what you really worry about is bias that overstates your conclusions. There were several instances where there were still estimation bias issues we couldn't deal with, but they biased the results towards a null result. I would always note that this wasn't a bad thing. We want to produce some insights, and if the only biases we have make those insights more conservative that can actually help convince people that there's something there.

So a lower bound estimate of 1.5 should help getting the point across that the fiscal multiplier is very real.

I am on board with that... mostly. Bias is still bad though, and you really need to try to think it through. What happens to these multiplier estimates when resources are fully employed? In that case, the results would be biased away from a null finding. Fiscal infusions draw resources away from comparison counties giving the impression of a larger multiplier when actually the fiscal infusion made no real contribution - and likely made the economy less efficient.

That worries me that we have underestimates during recessions and overestimates during normal times. Brad's point is a good one, it just makes me nervous.


Andrew Bossie also has comments on the last post that are worth sharing. First, he writes:

"There is a level I agree with you to this. In which case, metropolitan areas are a good robustness check."

I'm not entirely sure of what he has in mind here - hopefully he can elaborate in the comments here. One robustness check (well, really just a different specification) that I was thinking of was using some sort of propensity score matching approach to match counties in completely different regions of the country to each other on the basis of pre-recessionary characteristics. That way you're looking at two areas (really, a big collection of pairs of areas) which are comparable, but which are not going to lose demand to each other. I'm not sure if this would really work, though. After all, a "low fiscal stimulus" area is likely to be near another high fiscal stimulus area, and its ex-post performance is going to be impacted by the fiscal infusion in that neighboring area - and that fiscal infusion is definitely going to be correlated with your treatment area. Ultimately, I have no idea how big these biases are - and that's part of the problem. And I wouldn't know how to go about estimating how big the bias is. I suppose one way to estimate it would simply be to compare it to something like Barro and Redlick's work: how do national multiplier estimates from military procurement compare to state and county-level multiplier estimates from military procurement? Does that sound like a good approach to people?

Andrew goes on:

"In the united states, particularly "historically speaking" loanable funds markets have been HIGHLY localized becuase of unit banking. Even in the new banking environment there are still something like 8000 local banks."

This, I think, is the wrong way of looking at it. Certainly there's wide variability in loanable funds markets, and I think I said this in my original post. The point is, the main impact that fiscal policy has on the loanable funds market would not vary across localities. That's quite different from saying there is no variation across localities. It's not that there's no observed variation - it's that there's no reason to expect the variation in these county-level markets to be correlated with the impact of national fiscal policy in the cross section.

These sorts of models difference out all the variation common to a single area over time, as well as the variation common to both areas. Indeed - that's why we like these sorts of models so much. They get rid of a lot of the heavy lifting when it comes to holding things constant. You'll still have variation between local loanable funds markets to work with, but that will not include the most important variation for macroeconomic considerations: variations at the national and international level.


  1. I just want to say that I don't really disagree with your critique all that much and in fact I think it is quite good and important to keep in mind when doing this kind of econometric. I think the big difference is a difference in "attitude" in the sense that I'm not overly worried about bias becuase I don't really view econometrics as producing anything terribly solid anyway. I mean, you do want your estimates to be as honest as you can make them, but the whole exercise is imperfect and I'm willing to let a lot of stuff slide.

    I've also brought up before--though in fairness you seemed to have soften somewhat on your position--that you are throwing the baby out with the bathwater by giving county level regressions a hard time.

    As I've said before, time series estimates of fiscal policy are as problematic as these cross sections. I view the fact that for the most part time series can only give an average effect of recessions and expansions as unhelpful and distorting as spillover effects are. On top of which--I forget who had pointed this out--but you cant really even call VAR impulse responses "multipliers" since a multiplier by definition an endogenous response and impulse response functions are constructed to be wholly exogenous.

    Anyway, back to the two quotes you pulled out. I was being abusive with the term "robustness check" what I meant was metropolitan areas could serve as a way of demonstrating "robustness across specifications". But actually, I didn't really mean that. I actually think that metropolitan areas are a more legitimate level of observation because of the reduced spillover effect. I actually have plans to go back and re-do a couple county level papers at the metropolitan level when/if I ever finish my dissertation precisely becuase I feel like the county level is inadequate and (living in the New York City metro area) I feel like the state level is too arbitrary, which I guess you can appreciate living outside DC. Also, add to it the population size issues and I think metropolitan areas are a reasonable compromise between inclusiveness and the number of observations.

    I like your idea though and I may steal it someday. Though I'm not sure what leads you to assume low fiscal stimulus counties with similar characteristics would be next to next to high fiscal stimulus areas.

    As for the second point. In a sense I think there is no such thing as variation at the national level. I mean, to speak to Delong's point monetary policy is a purely national phenomenon but I don't think you can really (except as a necessary short hand) talk about a national economy. We live in a country of Buffalo and Detroit on one hand and Phoenix and Los Vegas on the other and San Francisco and new York on yet another. Sometimes, in fact, I think the focus on national statistics obscures more than it reveals.

    Again, I'm not convinced your critique is wrong, but I do think we are in really fuzzy territory.

  2. Something I'm interested in is the difference between:
    * "Part 1" Keynesians.
    * "Chapter 12" Keynesians.
    * and "liquidity trap" Keynesians.

    I've seen Daniel make all three sorts of argument. In the past I always saw these positions as closely related but not the same.

    Is it the case there are no "pure" part 1 Keynesians, only people who use the part 1 stuff as part of their overall argument?

  3. DK wrote:

    What happens to these multiplier estimates when resources are fully employed? In that case, the results would be biased away from a null finding.

    Daniel, I'm not sure why you and DeLong are saying this is a lower bound. Since you're admitting that in the case of full employment, getting this same data would clearly overstate the estimated multiplier, then we are obviously relying on an underlying theory to interpret the empirical results.

    So it seems you and DeLong are saying, "We know that when the economy's not at full employment, government spending clearly raises employment; we're just not sure how much. Let's look at the numbers."

    But what if you're arguing with somebody who thinks that's wrong? Why couldn't a full-scale Treasury View guy say the same thing you yourself would say, if we initially were at full employment?

    As Scott Sumner said in DeLong's comments, these results are perfectly consistent with the multiplier being zero.

  4. Right - you're always working off some theory when talking about bias.

    When we talk about bias in the educational returns literature, we're usually talking about the omission of a natural intelligence/ability variable. We say that educational returns estimates are biased towards zero because our most plausible theory is that ability/intelligence is positively related to both educational achievement and labor market performance. Certainly if someone disagreed with that they could say that the estimate is in fact biased upward.

    I would say they've made a bad argument.

    This is part of the reason why I hesitated at even calling it a "lower bound". I think Brad is right, and that's how I read the numbers - but bias is bias. Better to rely on unbiased estimators than to describe the nature of the bias you can't avoid. But if you truly can't avoid the bias, you do what you can do.

    You should know from past experience that I'm open to the point that data is open to more interpretations than most people are willing to admit. That's why I criticized Krugman in the first place on this, and why I hestitated/qualified Brad's support for a "lower bound" interpretation (even though I think that's reasonable, it's not guaranteed to convince everyone).

    btw, Bob, what did you think of your Mises compatriots posting that Phoenix Center multiplier study?

  5. Current -
    I, for one, got to know Keynes before I ever heard about these various camps. I've never found any reason to chop up the GT and oppose different parts of it.

    I think probably what you have is a historical accident. Without a real liquidity trap for a while, that part gets somewhat superfluous (Boianovsky has a great history of the liquidity trap and its telling that there was a ton of discussion of it early on, and then very little until Krugman came along and started talking about it again in the 90s).

    Chapter 12 is a more interesting point. I think this has always been of interest to people in finance, behavioral economics, economic psychology, etc. - it's just so outside the normal neoclassical paradigm it seemed safe to ignore for theorizing purposes - but relevant to drag out and quote every time the stock market tumbled a little.

    I suppose this makes me lucky - to have my formative years at a time when I can take the whole things seriously and comprehensively.

  6. re: "We say that educational returns estimates are biased towards zero because our most plausible theory is that ability/intelligence is positively related to both educational achievement and labor market performance."

    Holy moley - I mean biased away from zero.

    Education is positively related to ability, ability is unobserved, so the measured relationship between education and labor market performance absorbs the positive impact of ability and is biased upward. IF you believe the theory behind it, that is. Whenever we're talking data, we're always talking theory too. That's inevitable. That's fine - we just need to be cognizant of it.

  7. Which of course means the conversation between me and DeLong and the conversation between me and you on multiplier studies might be entirely different conversations.

    This is to be expected - truth claims are weighed with respect to standards of evidence and underlying assumptions of relevant communities of discourse.

    My banner talks about changing our mind when the facts change. We could also say that we change the facts when we change who we're talking with.

  8. Oh that sumner comment over on 'Grasping' reminded me of another point I wanted to make that I think you would agree with. I do think there is a problem extrapolating from the part to the whole. One thing I would like to see--which may not make much sense with metropolitan are studies--is an estimate of the total cumulative effect. Right, for instance most county level studies look at per capita data so they are compared in a "dimensionless" way. Either by estimating some kind of weighted average or my preference is to do a quantile regression to whatever level of detail and estimate the total effect across quantiles.

  9. Sorry missing some words there. The last sentence should start with "Total effect should be estimated either..."

  10. Daniel, I don't have time to open that Phoenix study (or this one). I read your critique of it, and last I checked the comments, Thornton hadn't answered you. So I'd like to hear his side of things before offering even a snap judgment.


All anonymous comments will be deleted. Consistent pseudonyms are fine.