reghdfe vs xtlogit fe: difference in assumptions

[Iuliia Svetetskaia]

Dear STATA users, greetings!

I have a very limited understanding of when it is appropriate to use a linear regression with a binary dependent variable, so I was wondering if someone could share some sources or their own knowledge about this topic.

I was advised to use reghdfe (a community-contributed module) instead of xtlogit fe, because the latter has very strong assumptions that are not satisfied in my research. I have a high-dimensional panel data (7 years of observation), and I also have a dependent variable that represents firms' strategic choice that is not finite in its nature (I basically found out that under some circumstances foreign firms might change their ownership type), so I guess this is why xtlogit is not the perfect fit for my research. But I still struggle to find more information about the cases where using a linear regression with a binary dependent variable was found to be an optimal choice.

I would really appreciate any insights about this topic~

Thank you for your replies in advance!

[Carlo Lazzaro]

Iuliia:
linear probability model is a case in point (OLS with a binary regressand), despite some drawback; -xtlogit- is a panel data command that applies the logistic function to longitudinal dataset (aka panels).
That said;
a) a 7-year panel dataset is probably a short one (N>T);
b) I'm not sure that your dependent variable is in fact binary (a binary variable is composed of two levels, usually coded 0 and 1);
c) the community-contributed module (as the FAQ kindly request you to mention) -reghdfe- basically elaborates on -xtreg,fe-.allowing >1 time-unrelated fixed effects.

[Iuliia Svetetskaia]

Iuliia:
linear probability model is a case in point (OLS with a binary regressand), despite some drawback; -xtlogit- is a panel data command that applies the logistic function to longitudinal dataset (aka panels).
That said;
a) a 7-year panel dataset is probably a short one (N>T);
b) I'm not sure that your dependent variable is in fact binary (a binary variable is composed of two levels, usually coded 0 and 1);
c) the community-contributed module (as the FAQ kindly request you to mention) -reghdfe- basically elaborates on -xtreg,fe-.allowing >1 time-unrelated fixed effects.

Dear Carlo,

Thank you for your reply!

My dependent variable is binary (joint venture=1; wholly-owned subsidiary=0), which is why I initially thought that xtlogit is the only appropriate option. But I guess it is not as black and white as I assumed.

I will have to look into more literature.

Thank you again!

[Carlo Lazzaro]

Iuliia:
thanks for confirming that your regressand is in fact binary.
As an aside to my previous reply, please note that -xtlogit,fe- implies conditional fixed effect, which differs from the -fe- estimator of the -reghdfe-.