I am running a logit model on a panel data for analysing the effect of a policy on zombie firms closure. Zombie firms are defined as firms that have been in losses for three consecutive years. Now I want to see the effect on individual and group firms over a period of 10 years. Group firms are the treatment group and standalone firms are the control group. When I try to run the logit model here using an interaction between Treat (whether firm is group firm or individual firm) and Post (Period before 2018 is pre and from 2018 is post-treatment) I am seeing that there is perfect multicollinearity between the interaction term and the fixed effects structure. My fixed effects structure includes time, country, industry and firm fixed effects. When I drop the firm fixed effects, the logit works perfectly fine. However, when I run the model using a linear probability model, the multicollinearity does not exist (even with firm fixed effects). Why do I face this issue when using the logit model and what could be the better fixed effects structure for this study? I also come with this question since when I set up my panel data my panel id is the firm id and by using xtlogit, the panel id fixed effects are automatically included. The idea here is that I expect the heterogeneity in industry and country to impact firm closure but not much on the individual firm characteristics (other than group and standalone firms which is exactly what we are trying to test).
I am not posting the output here since my data is extremely huge and the xtlogit is taking forever to run. However, I tried the same in R where it was much faster and where I could see the issue arising. (I used the feglm estimator in R)
I am not posting the output here since my data is extremely huge and the xtlogit is taking forever to run. However, I tried the same in R where it was much faster and where I could see the issue arising. (I used the feglm estimator in R)
Comment