Multiple fixed effects using logit models

[Helen Chang]

Hi,

I have a firm-year level panel data. The dependent variable is a dummy variable. I would like to include firm and year fixed effect with firm-level clustering (if possible). Anyone knows how to do this?

[Andrew Musau]

xtset firm year
xtlogit y x1-xn i.year, fe

will estimate a conditional fixed effects logit model with year dummies. With panel data, the reason for clustering is usually due to concerns about serial correlation and heteroscedasticity. If this is the case, you should not use the conditional FE logit estimator. See #2 of https://www.statalist.org/forums/for...tandard-errors.

[Maxence Morlet]

Andrew has provided great advice! You may want to compare your results to a benchmark / baseline linear probability model (LPM), especially if most of your regressors are binary as well. The LPM model is fit using

regress

[Helen Chang]

xtset firm year
xtlogit y x1-xn i.year, fe

will estimate a conditional fixed effects logit model with year dummies. With panel data, the reason for clustering is usually due to concerns about serial correlation and heteroscedasticity. If this is the case, you should not use the conditional FE logit estimator. See #2 of https://www.statalist.org/forums/for...tandard-errors.

Thanks, Andrew. If I need to use both firm fixed and year fixed effects, then should I just use the following? Or "fe" already stands for the firm-fixed effect?
xtlogit y x1-xn i.year i.firmID, fe

[Andrew Musau]

You declare the panel identifier in the xtset command

xtset firm year

So yes, the firm fixed effects are conditioned out of the likelihood with

xtlogit y x1-xn i.year, fe

[Maxence Morlet]

Helen Chang you may want to check out the community-contributed commands

logitfe

and

probitfe

[Helen Chang]

Thank you everyone!! Just a follow up question. I noticed that the number of observation using the logit model is much smaller than the sample size using the OLS regression. Is it normal?

[Helen Chang]

Thank you everyone!! Just a follow up question. I noticed that the number of observation using the logit model is much smaller than the sample size using the OLS regression. Is it normal?

"note: multiple positive outcomes within groups encountered.
note: 1,356 groups (10,895 obs) dropped because of all positive or
all negative outcomes."
This is what I received from Stata.

[Maxence Morlet]

See this thread for an answer to your question: https://www.statalist.org/forums/for...n-lead-to-bias.