Friday, July 1, 2011

I really don't understand how John Taylor's brain works

This is a draft of a forthcoming JEL paper. I've only skimmed it, but it is entirely consistent with everything he's written about the stimulus in the past. The PCE regressions on page 9 truly astound me.

How are these regressions identified?

I have no idea. ARRA spending is not exogenous, Prof. Taylor. That variable deserves to be on the left hand side of the equation as much as PCE. That's the whole f%#&ing empirical exercise. If we could just do what you've done here we could wrap up this whole economics thing in a matter of weeks and get on to figuring out how many angels can dance on the head of a pin.

On another note entirely - I worked with Ned Gramlich, who he cites in the introduction, before he passed away a couple of years ago.

UPDATE: Another question. Even if Taylor did convincingly identify any of his models, why is he so focused on PCE? You would want to see an impact on everything, but I would have looked at investment first. Isn't that what Keynesian policy is trying to affect directly, with the impact on consumption only indirect?


  1. Daniel...I too just skimmed the paper and find it incredibly suspicious he does not provide any regression results, methods, or tests of robustness. But one thought about your particular objections. I don't think treating ARRA as exogeneous is totally off base here from an empirical standpoint (only). Taylor specifies his lag-length as 1 period (quarterly I assume) and the endogeneity of the ARRA certainly originates further in the past than 1 quarter and is influenced by drops in other components of GDP than PCE. Also, if you endogenize ARRA you will lessen the impact empirically especially in a VAR environment which it appears he is using.

    With respect to PCE, I don't know but let me proffer a hypothesis. The Greenspan Fed was quite focused on PCE as a measure of the economy rather than GDP. Maybe Taylor is part of this camp?

  2. Assuming a drop in PCE causes an increase in stimulus, ignoring this endogeneity should bias the impact of stimulus downward. Why do you say that it would lessen the impact empirically?

    Is that what you're trying to say - that it would lessen the measured impact? That I would agree with. That's why the model needs to be identified. Romer does that. Barro does that. Lots of people do that. I'm not sure why Taylor pretends it's not even an issue.

    re: "With respect to PCE, I don't know but let me proffer a hypothesis. The Greenspan Fed was quite focused on PCE as a measure of the economy rather than GDP. Maybe Taylor is part of this camp?"

    That, and I'm guessing that he's just thinking of "hydraulic Keynesianism".

  3. It's been years since I've done VARs or time series, btw - I'm coming at this with the sensibilities of a microeconometrician that tries to plausibly identify program impacts for a living. Perhaps there is something I'm missing.

    But if that's the case, why do people like Barro and Romer spend so much time working on convincing identification strategies for this? Why don't they just run a VAR?

    For that matter, why do I spend so much of my time figuring out plausible quasi-experimental specifications to get these impact estimates? Why do I worry about constructing comparison groups?

  4. I guess what I am saying is that the relationship from PCE to ARRA is not a continuous function, there are threshold effects and what would drive stimulus would not be levels as much as changes. Also, from a political economy standpoint, what is the rationale for stimulus? I'm sure PCE is a factor but I would hardly say it was the dominant factor in the decision to implement ARRA. Peak to trough, PCE dropped 1.7% while fixed investment fell 24.7% in real terms which is your point about why measure PCE. But it also underscores my point that it is valid to treat ARRA as exogeneous to PCE. As far as how this effects the regression equation, given the VAR setup then it is not a left and/right hand side issue it is a question of in which time period they correlate (or over all). If it is the case, then in a VAR, the presence of both the ARRA and the PCE at a lag on the RHS will overlap in the explanation of the contemporaneous PCE variable and therefore the estimate of the ARRA coefficient will have a higher standard error/standard deviation and will be biased down, not up.

  5. That's ok Daniel, a lot of us don't understand how your brain works. ;)


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