Individual fixed effects for xtlogit and xttobit

[David Johansen]

Dear Statalist,

I run both a logit (xtlogit) and a tobit (xttobit) model for my research. I want to control for individual fixed effects, but I struggle with how to do this. For the year fixed effects I use i.year, however when I use i.id all the individuals are omitted because they predict success/failure perfectly.
I have an unbalanced data set, with five years of data --> N=8500.

[Jackson Monroe]

Do any of your outcomes vary within IDs? It seems like they don't. Use dataex to post some sample data so we can see what you're working with.

[David Johansen]

Hi Jackson,

Indeed I have 4 control variables that are assumed to stay constant across all the 5 years. The other variables do differ each year.

[Carlo Lazzaro]

David:
if you go -xtlogit,fe- what you should be interested in, as postestimation, outcome, is -pc1- (the predicted probability of a positive outcome conditional on one positive outcome within Group).
Hence, I fail to get the need to include -i.panelvar- among your predictors.

[David Johansen]

Hi Carlo,

It's a random effects model, does that make a difference?

[Carlo Lazzaro]

David:
if -xtlogit,re-, go -xb- as postestimation command.
See also -xtlogit postestimation- helpfile (abd/or Stata .pdf manual entry) for more details.

[David Johansen]

Hi Carlo,

Thanks for your prompt reply. Unfortunately, I do not see how that will obtain me individual fixed effects? Is it unusual to add individual fixed effects to both a Xtlogit, re and a Xttobit? I can't find any suggestion in literature either.

[Carlo Lazzaro]

David:
1) outside the -xtreg- realm, you will get conditional fixed effect (due to incidental parameter bias; see if interested

http://www.econ.brown.edu/Faculty/To...meters1948.pdf) with the -fe- specification in -xt-
2) the most relevant post estimation outcome in -logistic- models at large (being cross-sectional or panel datasets) is probability of a positive outcome (as the regressand is categorical, this makes sense).

[David Johansen]

Thank you very much Carlo! Is i.ID the only way to incorporate the individuals?

[Carlo Lazzaro]

David:
yes, but with pooled logistic the incidental parameter bias strikes back (https://stats.stackexchange.com/ques...ameter-problem).