Saturday, November 19, 2011

Wag the dog

I'm not much of an econometrician, but at least I'm not afraid of trying to code things from scratch.  This déformation professionnelle can probably be traced back to Frank Fisher who, in the first problem set in the PhD econometrics sequence at MIT, had us do OLS using a pocket calculator: yes, we had moved beyond Blaise Pascal's mechanical calculator of 1642 in those days.  I guess you really respect PCs after you've been subjected to that sort of cruel and unusual punishment.

In this day and age, people are unfortunately used to pushing a button, or at least invoking some pre-packaged command, which will usually spew things out directly as a table into a Latex file, in an aesthetically pleasing format, wherever you want it to appear in the earth-shattering paper that you are writing.  This is wonderful.  It's simple.  But do people, especially newly minted économètres en herbe, actually know what they are doing?

I know I sound antediluvian, but sometimes I really wonder...   Let's take two concrete examples, which span the entire spectrum of the profession.

First, let's take someone on the highest (Great White Shark) rung of the economic food chain: Robert Barro.  In a memorable 2005 Journal of Monetary Economics paper, Barro and Lee, apart from managing to run one of the crispest forbidden regressions in history (with a half dozen covariates included in the structural equation not appearing in the first-stage reduced forms) and, apparently without knowing it, multiplying their exclusion restrictions by a factor of three (by including a whole bunch of covariates in the first-stage reduced form that do not, in turn, appear in their structural equations), make the following statement (footnote 16 on p. 1253):

"Fixed-effects Tobit or probit models cannot be implemented because there does not exist a sufficient statistic allowing the fixed effects to be conditioned out of the likelihood.  See Wooldridge (2002)."

Apart from demonstrating that Wooldridge's (2002) textbook was never opened (in section 16.8.2, on p. 541, Wooldridge shows how to use a garden-variety Mundlak procedure to correct for time-invariant heterogeneity in a panel data tobit regression),  this would be news to Bo Honoré, one of the smartest econometricians out there. In a 1992 Econometrica paper, Honoré developed trimmed LAD, which is Tobit with fixed effects.  He even furnished the corresponding GAUSS code, called PANTOB --PANelTOBit, get it?-- which I used with two co-authors in a 1998 paper.  Come on, Honoré's work is 13 years before the Barro and Lee paper, and Econometrica is not exactly the sort of journal that goes unnoticed! And Jim Powell was already dealing with this issue in an Econometrica article back in 1986.  That's 25 years ago!  But I guess the specific command didn't exist in Stata.  And that adding all of the covariates in their time-invariant country-specific means incarnation to a random effects tobit, which is all that the miracle of the Mundlak procedure calls for, was just not worth the trouble.

Second, for the plankton end of the food chain, consider the following conversation that I have had with economics graduate students, on several continents, a depressing number of times: 

Me: "Why did you include a lagged dependent variable in your econometric specification given that your theoretical model didn't call for one?"
Student: "I had to."
Me: "What do you mean by 'you had to'?"
Student: "I wanted to use GMM to estimate the equation."
Me: "So? GMM is just efficient IV, so do the Nike thing and 'just do it'!  I still don't see why you included the lagged dependent variable... it doesn't appear in the theoretical model."
Student: "Actually, it's because xtabond2 doesn't let me estimate if I don't include a lagged dependent variable."
Me: "So, estimate with pure GMM without a lagged dependent variable!"
Student: "I can't find a command that allows me to do that."

Here is another prima facie case (see my post on the Talibans of randomization) of the tail wagging the dog.  The prosecution rests.

What is to be done about this?  For Robert Barro and his merry band of co-authors, it is unfortunately too late.  Arguments d'autorité rule in the rarefied atmosphere of top level academic publishing, and anyway, he has some very interesting stuff to say.  So why gripe about the minor point that he would have been failed by Jerry Hausman in his PhD econometrics class.  Frank Fisher, one of the sharpest intelligences it has been my priviledge to witness in action, would have simply given him a negative grade.  I kid you not, Frank, quite reasonably in my humble opinion (now, with the benefit of hindsight, not back then...), preferred no answer to a stupid answer, and used to regularly hand out negative grades  ---hummm, maybe I should consider this at the Graduate Institute.... I wonder how that translates into the strange Swiss 6 point grading scale...?

But for graduate students in Economics, there is still hope, although the manner in which graduate econometrics is usually taught leaves little room for optimism.  I try to teach the stuff by getting people to code various procedures à l'ancienne in R, using basic matrix algebra.  It's sort of like making your Hollandaise sauce from scratch, the old fashioned way, instead of tearing open a plastic bag of the chemically synthesized ersatz stuff.  But this is not exactly a recipe for popularity, by dint of the simple fact that most graduate students these days are not really used to having to actually understand what they estimate.

Indeed, by roughly mid-November (i.e. now) I, alongside my partner in econometric crime Ugo Panizza, are without doubt the two most hated men in Rigot, the building resembling a refugee camp near the entrance to the UN in Geneva, where we teach.  Sometimes I wonder whether the daily demonstrations on the Place des Nations aren't going to turn into an uggly free-for-all with angry protesters brandishing "no more R coding in Econometrics II !" placards breaking down my office door (which wouldn't be hard, given the state of our beloved building).  I should probably keep a tear gas cannister handy next to my desk. 

Now I have no problem with being seriously unpopular in the name of a noble cause.  As Stalin is purported to have quipped, it is far better to be feared than to be loved.  I positively revel in the opprobium.  After all, if people are made significantly unhappy by having to struggle with code, it means that they had no idea how to do it in the first place, and probably did not really understand the procedures that they were running. The push-button antibodies are therefore likely to be taking hold in their organism.  And that, to be fair, is painful --had a tetanus shot recently?

It's probably a losing battle, but what the heck, there is something glorious about charging the machine gun nest on horseback with sabre drawn.  And that poor dog really has to regain control of his tail.

1 comment:

  1. "Upon this a question arises: whether it be better to be loved than feared or feared than loved? It may be answered that one should wish to be both, but, because it is difficult to unite them in one person, it is much safer to be feared than loved, when, of the two, either must be dispensed with."
    (Niccolò Machiavelli: The Prince)

    ReplyDelete