Logit vs xtlogit when data are organized as a panel data.

[Mark Debrigian]

Dear Statalist
First of all, many thanks for the help you give to us. I’m a bit confused about the use of “logit” when data are organized as panel data. Let me explain what I mean. I have a panel data organized as follows: xtset Firms Year (…strongly balanced). I run a logit regression with xtlogit.

Specifically my command is: Xtlogit DepVar Var1-Var5 dummies

where VAr1-Var5 are 5 continuous variables while dummies are 700 (seven hundred) binary variables reporting information about cities, legal type, industry, treatments and many other aspects. I never seen the results of this regression because my STATA 12 works for hours and hours without any results. I leave STATA working for the whole night but, after 6,8...14 hours, the screen is stopped to Iteration 1 or 2 of full model. Of course, I use “set more off” command .

I think (I’m sure) that the reason of this problem is related to the huge number of dummies.
However, my question is on a different topic. Since I need to read some results, also if very preliminary and not robust, I run the previous regression with “logit”. After 4-5 minutes, I obtain the results.
My question now is, which type of estimation i obtain this when i use logit rather xtlogit? Obviously, it is not a panel data estimation (I’m using logit and not xtlogit) even though data are organized with xtset. Is it a Pooled Logit Estimation?


Many Thanks in advance.

Mark Debragian

P.S.
Please, do not consider at this stage the opportunity or not to insert 700 dummies in one regression. I can sure that my problem with xtlogit persist also with 200 dummies and less.

[Clyde Schechter]

Yes, using -logit- gives you a pooled estimation. Which is probably not going to be useful: it ignores variation among firms, and it incorrectly treats all the observations as independent.

If I were you, I wouldn't be so sure that 200 dummies wouldn't be a lot faster than 700. The size of the VCE matrix goes up as the square of the number of variables. And, if I recall, the kinds of calculations done with it are at an even higher power. It really might make sense to reduce the number of dummy variables either by omitting altogether some effects that might be unimportant, or combining categories of some polychotomous variables. With so many dummies, in addition to having a very large VCE matrix to deal with, it is likely that some, perhaps most, of the dummy variables are identifying just a handful of observations. When that happens, there is very little information in the data set about the effect of that dummy, and this leads to requiring more iterations to get to convergence on those coefficients. In the end, of course, only you can judge whether the computational effort required to estimate this model is justified in terms of what you might learn from it.

[Mark Debrigian]

Many thanks Clyde. I will follow your suggestion.