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

RE: st: clogit with fixed effects

From   jverkuilen <>
To   <>
Subject   RE: st: clogit with fixed effects
Date   Sat, 21 Mar 2009 01:43:41 -0400

Elaborating slightly on Steve's excellent post, the problem with estimating the intercepts is that unless the number of observations within groups is relatively large, they suffer from the inconsistency noted by Neyman and Scott where the numbers of estimated parameters grows linearly with N. 

In psychometrics, most of us prefer marginal maximum likelihood estimation followed by empirical Bayes prediction (or fully Bayesian posterior prediction when the posterior mode is a poor summary) to estimating fixed effects, which we refer to as "joint maximum likelihood." The latter are unbiased in a sense, but the shrinkage from MML is judged to be worth it to avoid inconsistency. JML works when there are many observations per group because each within-group average is reasonably well-defined. 

As Steve noted, -xtmelogit- gives MML and -predict- gives EB predictions post-estimation. 

-----Original Message-----
Sent: 3/20/2009 11:59 PM
Subject: Re: st: clogit with fixed effects


Please, next time, as the Statalist FAQ request, show exactly what
-xtlogit- command you would like -clogit- to duplicate.

I think you misunderstand the relation of -xtlogit- with the "fe" option
to -clogit-. They are equivalent, as the manual entries make clear. Run
the code below for a demonstration. In fact, the full title of -clogit- is
"Conditional (fixed-effects) logistic regression."  These commands
suppress intercepts for the panels (-xtlogit-) or groups (-clogit-). Since
you give no details , I cannot tell if "individuals" and "countries" are
the group variables you have used in -clogit-.  If you wish to add dummy
variables and estimate the intercepts  explicitly, use ordinary -logit-.

This is a bad idea if you do not have enough observations to estimate the
intercepts with good precision.  Even if you have enough observations, it
is probably still a bad idea.  What are you going to do with the
intercepts?  With hundreds (or thousands) of countries and individuals,
you will get pages of output.  Instead, consider -xtmelogit-, which will
allow you to model random slopes in addition to random intercepts.  Unlike
-clogit- and -xtlogit, fe-, -xtmelogit- followed by -predict- will
estimate useful predicted probabilities, including probabilities for
groups, considered as random effects.


*************BEGIN CODE********************
webuse union, clear
xtset idcode
xtlogit union age, fe
clogit union age, group(idcode)
**********END CODE*********************

On Mar 19, 2009, at 3:08 PM, wrote:

Dear Statalisters,

I would like to run a conditional logit and add individual and countries
fixed effects, but I dont find the command which corresponds to
*xtlogit...fe* for the conditional logit model: it seems that it does not
exist. So I am to create dummies and add them in the traditional
conditional model. Do you think it is a good idea?

*   For searches and help try:

*   For searches and help try:

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