[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Convergence time of logit with fixed effect

From   "Kyle K. Hood" <>
Subject   Re: st: Convergence time of logit with fixed effect
Date   Sun, 08 Mar 2009 20:09:22 -0700

I don't know if anyone has yet replied, but there are a couple of solutions that may improve performance:

(1) This is not a sign that it will not converge -- you only see up to 2 decimals (the numbers are not the same, and my guess is that the maximization routine will quit when the gradient is small).

(2) You are allowed to choose maximize options. If you want, you can set the tolerance. You can also set the max number of iterations to 100 or something, and see what you end up with.

(3) You should not use xi with logit. One reason that stata has an xtlogit and not an xtprobit, is that you can remove fixed effects in a logit with preprocessing, much as you remove the FE in a regression my de-meaning the data by groups. Regression on dummies does not take advantage of this. Try:

. xtlogit y x1 x2, fe i(x3)

(4) Breaking the process right now will do nothing (you will get no result). However, you're probably better off breaking it now and trying other approaches (IMO, but it comes with no warranty.)

Glenn Goldsmith wrote:
Hi Anupit,

This doesn't really answer any of your specific questions, but -help clogit-
might be useful to you as an alternative way of fitting your model.

"Supnithadnaporn, Anupit" <> writes:

Dear all,

I run the logit model with some fixed effect using

xi:logit y x1 x2 i.x3

My data is quite large around 686,000 observations.
And x3 has around 700 categories. I use Stata MP on
the server and it has been running for 2 days until now.
The current iteration is 402 but the log pseudolikelihood
does not change from the iteration 31.

Iteration 31:  log pseudolikelihood = -149641.22
Iteration 402:  log pseudolikelihood = -149641.22

Would anybody please advise?
1) Is this a sign that it will not converge?
2) Should I keep waiting? How long should I wait?
3) If I interrupt, will I get any result at all?

Thank you very much,

*   For searches and help try:

Kyle Hood
Department of Economics
Yale University
New Haven, CT

fn:Kyle Hood
org:Yale University;Economics
adr;dom:;;;New Haven;CT

© Copyright 1996–2021 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index