Why does clogit offer clustered standard errors while xtlogit does not?

[Richard Herron]

My understanding is that -xtlogit- calls -clogit- as a convenience for users familiar with the -xt- commands. I am one of these users.

I correct standard errors for within-individual clustering with either -vce(robust)- for -xt- commands or -vce(cluster individual)-. However, -xtlogit- lacks these options, so I typically block bootstrap to correct standard errors for within-individual clustering.

Today I read the -clogit- help file and learned that -clogit- offers -vce(robust)- and -vce(cluster individual)-! In my limited testing, clustered -clogit- and block bootstrapped -xtlogit- estimate similar standard errors.

My question: Why does -clogit- offer clustered standard errors while -xtlogit- does not? Did I waste years of my life (OK, hours ) block bootstrapping standard errors with -xtlogit-? Are there cases where -clogit- is not a suitable alternative to -xtlogit-? It seems the -xtlogit- help file should point users to -clogit- for more options, or maybe it does somewhere.

[daniel klein]

This is probably a good starting point to an answer if not already a satisfying one.

[Richard Herron]

Perfect! Thanks!

[Andrew Musau]

The conditional maximum likelihood estimator (CMLE) is inconsistent in the presence of serial correlation and heteroskedasticity - see the following reference provided by Jeff Wooldridge: https://www.statalist.org/forums/for...tandard-errors. xtlogit is exclusive for panel data whereas clogit is more general, and one may have clustered data, in the same way that clogit allows the -svy- prefix.

help svy: clogit

[Richard Herron]

Thanks, Andrew! Your link is very helpful, too. But I am saddened that my search skills are so bad that I missed two prior answers!

[Jeff Wooldridge]

Just to follow up on Andrew's helpful response, I do think that Stata should allow a cluster option with xtlogit (and other xt commands when the estimators are generally inconsistent with serial correlation/heteroskedasticity): one might want to cluster at a higher level. For example, if you have student level panel data, but you're studying a school-level policy, there's a case for clustering at the school level.

Even clustering at the student level makes sense if one takes the sensible perspective that all models are misspecified, and so we want to properly reflect the uncertainty in such cases. But then you're admitted the estimators are inconsistent, and we're usually reluctant to do that.

BTW, the paper that demonstrates the bias in the fixed effects logit estimator with serial correlation, by Kwak, Martin, and Wooldridge, has now been published in the Journal of Econometric Methods -- but it's remarkably hard to find online.

[Richard Herron]

Thanks, Jeff. I appreciate your insight!