I'm going to be chairing a panel on generalized maximum entropy (GME) methods in economics at the Eastern Economic Association conference in Boston this March. We'll be presenting at 8 am on March 9th, so if you are attending please come by and see it! All of the papers come from research that we did in an information theoretic econometrics seminar this past fall.

Paul Corral and Mungo Terbish will start by presenting their paper on a new STATA command they developed for discrete choice GME models. I believe the ado files are available now, and the paper itself is under review at

*The Stata Journal*. Since they are basically explaining how the program works, this presentation will be a nice introduction to GME methods for audience members that are not familiar with it.

After that Ermengarde Jabir will present work that uses GME methods to look at household asset allocation using Survey of Consumer Finance data. This rounds out the panel nicely because it's an application of GME methods to an actual problem in economics (unlike my paper and Paul and Mungo's paper).

I'll conclude with a discussion of a GME version of propensity score matching (with estimation of the propensity score using a discrete choice GME model and estimation of the average treatment effect in the second stage as usual). I'm specifically interested in whether GME outperforms standard discrete choice estimation of the propensity score in cases of low common support. Eventually I'd like to run this method on LaLonde's (1986) data from his classic test of propensity score matching, but that probably won't be ready by the conference.

I've heard of GME but don't really know what it is, but I find this area really interesting. The Dehajia & Waba (2002) and Smith & Todd (2005) back and forth was great. Is there a good go-to introduction for GME?

ReplyDeleteDW and ST are going over PSM, not GME.

DeletePSM is a prominent quasi-experimental design, and I agree it's intriguing because it's more accessible than most alternatives - but it comes with a lot of cost in terms of biased estimates. So it really requires a lot of careful justification and introspection... which makes for interesting discussion.

Think of GME as an alternative estimation strategy. So whereas PSM is an alternative to random assignment, RDD, DID, etc. you can think of GME as an alternative to least squares or maximum likelihood. It's a way of getting an answer mathematically that can be applied to all kinds of models. Why use it? It has certain properties that are desirable, particularly in cases with small sample size or if you don't want to make firm distributional assumptions (which you have to for maximum likelihood or least squares).

Perhaps I'll do a post on the basics of GME soon.

Basically with ML you maximize a likelihood function that you specify based on your model and your assumptions about distributions of the error terms. With GME you are maximizing the entropy of your outcome variable and an error term: -(SUM(plog(p)).

Why would you do this? Where is the model?

Well that takes a little more explaining. You do it because it's a more conservative estimate that requires less assumptions about the distribution of the errors. The model is specified in the constraints of the maximization problem.

The best introduction I have is my professor's slides! The second best introduction is one of many very expensive books by Amos Golan.

DeleteA great summary of the GME model I'm using in the PSM context is in:

A Maximum Entropy Approach to Recovering Information From Multinomial Response Data

Author(s): Amos Golan, George Judge and Jeffrey M. PerloffSource: Journal of the American Statistical Association, Vol. 91, No. 434 (Jun., 1996), pp. 841-853

That is the discrete choice version, but it should give you a sense of how these models are set up - and I believe they give an overview of the benefits of GME (otherwise there should be cites).

I found it extremely readable for an article in a statistical journal, which I usually find hard to follow - but that may be because I was immersing myself in it all semester for the seminar.

Rereading your comment I'm wondering if you actually were on the same page as my first comment already. Either way hopefully it provides a better summary of the lay of the land of what I'm actually doing for other readers.

DeleteYeah my wording was poorly arranged but what I meant by "area" was not GME but rather the debate over how well PSM could recover the true parameters of an underlying model. I mentioned DW and ST bc their analyses built off of Lalonde's work, which you mentioned in your post.

ReplyDeleteThanks for the comments, though, they're very helpful. I've come across GME a couple of times and it always seemed forbidding so I just glossed over it.

Out of curiosity Daniel...has there been any intellectual history articles written on GEM before?

ReplyDelete