Monday, March 5, 2012

DeLong does not understand what Cochrane is claiming

Brad DeLong has a criticism of John Cochrane's post, which I linked to earlier on here. I think he's missing Cochrane's point. DeLong writes:

"I had thought that we had gotten John Cochrane to admit that the national income identity was not a behavioral relationship--that when the government spent money there was no necessity (or even likelihood) that the people who would eventually pay the taxes to ultimately redeem the debt issued by the government to fund its spending program immediately cut back on their spending by as much as the government increased its. But here we have Cochrane again expressing what sounds a lot like the totally-wrong British Treasury view of the 1920s, again expounding what I call Eugene Fama's fallacy."

Brad is right that the British Treasury view of the 1920s is wrong, but that's not quite what Cochrane is getting at. Cochrane is addressing much the same point that Krugman does in this post: proper specification of the counter-factual in state-level multiplier studies. Take the Nakamura and Steinsson (2011) paper that Cochrane is discussing in the post. They look at military (i.e. - federal) expenditures at the state level to estimate a multiplier. That's not a bad thing to start with. Barro, Ramey, and many others work with military expenditures because they have a better chance than most kinds of spending to be uncorrelated with macroeconomic conditions. But there's a crucial identification problem associated with doing this at the state-level.

Let's start with the theory of the multiplier. What is the biggest reason why we Keynesians don't advocate spending trillions and trillions of dollars in stimulus every single year - depression or no depression? A simple Keynesian cross alone would make that look pretty appealing, would it not? But we don't. Why? Crowding out of private investment and consumption. Brad DeLong and I agree that crowding out is a red herring right now, in this liquidity trap, but we also both agree that it's a real concern in normal circumstances. That's why we're demanding a trillion dollar stimulus in 2008 and 2009 but not 2004 and 2005.

Nakamura and Steinsson use data from 1966 to 2006 - a period when I think Brad would agree with me there was a real issue with crowding out. Looking at this at the national level, like Barro and others, would allow the crowding out to be captured in the multiplier estimate produced. And - no surprise to me - Barro finds relatively low multiplier estimates. It makes sense. We weren't in a liquidity trap, and some of the biggest instances of military spending (Vietnam) came at a time of full employment. I would be surprised if Barro didn't produce relatively small multipliers.

Any crowding out in the Barro analysis is going to reduce the multiplier estimate. That's exactly what we want as empiricists - we want to know how bad of a problem crowding out is. Nakamura and Steinsson's analysis is different. Since taxes and crowding out of private investment to finance military expenditures are going to be spread out across all states, when Nakamura and Steinsson compare low military spending states to high military spending states their counter-factual group (the low military spending states) are going to bear a portion of the costs of the military spending in the treatment group (the high military spending state). Unlike Barro and Ramey and that work, this biases the multiplier estimates upward.

I don't know if Cochrane holds the "Treasury view" for the current downturn or not. I do know that his critique of the Nakamura and Steinsson paper on empirical grounds is solid, and you can agree with him on that without holding the "Treasury view" (I don't hold the Treasury view after all - I don't think massive fiscal stimulus right now would result in much crowding out).

Regular commenter Andrew Bossie thinks these state level analyses still have value. I can sympathize with that. I don't want to impugn the people that worked hard on these studies, and it's true we don't have a ton of options in macroeconomics. But I do think a big, dark, nasty shadow of dubiousness should be cast over state-level analyses. And personally, I really, really hate arguing for a policy based on an upper bound estimate. I'd much rather argue for a policy based on a lower bound estimate, with lots of good reasons for thinking the multiplier could actually be higher.

Not all state-level analyses are upper bound estimates. There are many that are biased downward. Think about the studies that use any sort of spending that originates in the state itself. I think Andrew himself (or maybe someone else?) once showed me a doctoral dissertation that used windfall increases in pension fund investments as an instrument - that would be one could example and a contrast to the military spending that originates at the federal level. These studies are probably going to produce multipliers that are biased downwards because commerce across state lines is going to dissipate any aggregate demand boost (while international trade penetration is not particularly high in the U.S., we exchange a lot across state lines). In Nakamura and Steinsson, it was the cost of crowding out that leaked across state lines. In studies of spending that originates within a state, it is the benefit of the demand boost that leaks across state lines. I'd much rather have a multiplier from these studies that's modest but that I can stand firmly on than a multiplier from the other studies that's huge and that I have doubts about.

This stuff matters. If we're not going to do empirical economics right, what the hell is the point of doing empirical economics?


BTW - similar issues abound in the literature on the impact of various labor market policies. If you don't specify your comparison group carefully, you may be counting displacement of comparison group workers in the counterfactual, thus biasing your estimates upward. I think I have a neat work-around to this problem in an evaluation of job creation tax credit that I'm working on - I'm hoping to get a lot done on it over spring break, so maybe I'll have more details on that soon.


  1. I've made this argument before but the Barro Redlick/Ramey Shapiro analysis suffers from the opposite problem of the state analysis, and it's not more valid because it produces lower bound estimates. Again, I think the problem of pinning down a real estimate of fiscal policy shocks is essentially intractable. It makes more sense to me to look at some kind of average effect across studies that seem legitimate than to just take whatever estimates are lowest.

    The B/R paper is driven by WWII, which indeed was a massive crowding out and also completely inapplicable to any analysis of aggregate demand management. The R&S paper isn't much better in that regard, their results are driven by the Korean War shock. That war was almost entirely financed by a tax increase. Again, not applicable to any properly "Keynesian" discussion of the efficacy of fiscal policy because for fiscal policy interventions to have a net effect they necessarily must be deficit driven.

    In some ways the state analysis actually captures a properly Keynesian effect better for the very reason you are criticizing it. These state level shocks are more "exogenous" shocks to the states than they are at the national level and so are more like deficit financed shocks.

    1. And I've agreed with you before on those criticisms, but there's a big difference between the bias of an estimate and the generalizability of an estimate.

      Generally speaking, I would call Barro/Redlick and Ramey/Shapiro unbiased but hard to generalize.

      I would call the state studies biased, but perhaps easier to generalize.

      But that just raises the question - why in the world would you want to generalize a biased estimate?

      Anyway, as a matter of actually how to move forward I think you and I are probably closer. I don't want to toss out this information. We have a good sense of in what ways these are biased. I think a better way to put it than "averaging" the effects is to say that we need to triangulate the multiplier we're interested in, given what we know about the various biases and issues with each study.

    2. Yeah I think we agree more or less. I just think you give the VAR/time series lit too much credit. I think the narrative approach is prone to bias. Also, as the health economist Mike Grossman has said "Id rater have an interesting biased estimate than an uninteresting unbiased estimate".

      Also, would simply including a variable that captured the relative tax burden of each state solve much of the problem above? I'm not familiar with the N&S paper though I obviously should be!

    3. Never heard the Grossman quote before, but a shudder went down my spine when I read it.

      I would have to think about the solution - and I'm not closely familiar with the N&S paper either.

      The other thing that came to mind is using military spending to produce a predicted net federal spending value, and using that. Treat it like the instrument it is.

      I'm still not sure about this though - I'd have to think about it. What's really going on here is that high military spending states are getting transfers from low military spending states. You're estimating the impact by differencing the two. That seems problematic to me.

      So let's do it your way and let's say taxes are evenly distributed across all states - only the exogenous spending isn't evenly distribute. If the tax burdens are evenly distributed, they're going to be unrelated to GDP, and they're not going to change the multiplier estimate. How does this help? I would think we would want a solution where we could imagine evenly distributed taxes and unevenly distributed exogenous spending where doing something to account for the taxes reduces the estimated multiplier. Simply controlling for that doesn't seem like it would do it.

    4. If taxes are distributed evenly (I don't know if I buy that but it seems like its a key assumption of N&S) don't you have a treatment and control group that both have the negative tax effect built in? Like, and I'm not sure I'm thinking about this correctly, but aren't you estimating (Positive Spending Multiplier) - (Negative Tax Multiplier) across all states where in many states PSM = 0 in which case what you would see is the negative effect of the increase in taxes as the "multiplier" estimate?

      Where I see the issue is that net tax burdens are not evenly distributed, there are big net tax transfers from the north to the south. This is probably as true for tax transfers for military spending. But I'm not sure how you measure the increased tax burden in, say, New York to finance a new military base in Arizona.

      I dunno I feel like I'm wrong about this. I also feel like this is meaningless without having really read the N&S paper.

  2. Wow maybe I just have a psychological need to always quibble with you, Daniel, but I actually totally get why DeLong is complaining in this fashion. I too thought that Cochrane was reverting back to the view that "if the government spends money, it must come from somewhere, so all it can do is rearrange output, not increase it on net." You're right, he made good points about problems with looking at state-level differences, but then he sort of motivated the whole thing by an apparent appeal to the Treasury view.

    1. I know you said we need better Keynesian commentators in that last comment thread, Bob, but I didn't realize you were volunteering for the job!!!


    2. Anyway - we all know what's REALLY going on here. I'm just trying to get Brad to link to me again :)

  3. Ummm... Nakamura and Steinsson are estimating a constant-monetary-conditions multiplier—that is, they are comparing different levels of government purchasing while holding real interest rates constant across their identifying variance. That is not the policy-relevant multiplier when monetary policy conditions are not in fact constant in your counterfactual. But it is a fine estimate for today.

    There are, I think, the places to question the applicability of Nakamura and Steinsson:

    1. Supply-side regional factor mobility biases their estimate up.
    2. Demand-side regional factor mobility biases their estimate down.
    3. In the current conjuncture, the counterfactual is not one of higher government purchases with constant monetary and financial conditions--a constant real interest rate and constant risk premia. A stronger economy means faster inflation and fewer bankruptcies, and thus lower real interest rates charged to firms. There is not crowding-out but crowding-in, and N&S's constant-monetary-and-financial-conditions multiplier misses this effect.

    No, I do not know how these three effects add up. But I have to have a view by 1 PM or else...

  4. Crowding out of private investment and consumption.

    In the history of the World, this has never happened in a way that ought to concern us, for political and not economic reasons. Recall Keynes letter to France cited elsewhere, about the level of taxation being set by politics.

    You are in the territory Roosevelt talked about when he said nothing to fear but fear, itself.

    The Chinese are proving this truth, right now. They are lending and spending more solely because they, justifiably or not, have more confidence about the future than do we.

    Last, take the long view. The Pyramids had to, in the short run, been a NPV (and who knows about crowding out) but in the long run they were priceless.


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