Probit/Logit with Panel Data. Should I use probit or xtprobit?

[Nick Jenkins]

Hi everyone,

I am using STATA 14 to work with a panel data set of the United States from 2007 to 2015. I want to estimate a discrete choice model but I am not sure whether I should use:

probit dep indep ..., vce (cluster stateid)

or:

xtprobit dep indep ..., pa vce(robust)

I am concerned about serial correlation in my data which is why I am shying away from using a Logit model with fixed effects and using vce(bootstrap) doesn't seem to work:

xtlogit dep indep ..., fe vce(bootstrap)


Essentially, my question is what estimation method to use. Both probit and xtprobit give very different results. Any suggestions would be greatly appreciated!

[Clyde Schechter]

If you have panel data, you need to use panel-data models. Your observations are not independent, and -probit- does not account for that. -xtprobit- (or -xtlogit-) does. If after running -xtprobit- you find that rho (at the very end of the output table) is very close to zero, then it would be acceptable to say that the extent of intra-panel correlation is small enough to ignore and if there is some other advantage to using -probit-, you could then use -probit- (and would get nearly identical results). But otherwise, it is wrong to use a one-level model such as -probit- on panel data.

Added: What do you mean by "I am concerned about serial correlation in my data which is why I am shying away from using a Logit model with fixed effects and using vce(bootstrap) doesn't seem to work:" -xtlogit, fe- definitely allows the -vce(bootstrap)- option. If you aren't able to run it, you may be doing something wrong or there may be a problem with your data. Without knowing exactly the full command you tried and seeing the exact output you got, it is impossible to advise more specifically.

[Carlo Lazzaro]

Nick:
as an aside to Clyde's superb advice, you're seemingly dealing with a large N, small T panel dataset; hence, autocorrelation is probably a minor concern.