Panel data logistic regression (xtlogit): Choosing between fixed effects and random effects

[Roger Clements]

HI Helpful List,

I have panel data on ~2000 companies between 2005 and 2010. Of these 2000 companies, about 100 went bankrupt in some year between 2005 and 2010. I have various variables in the dataset, including financial leverage, performance, shareholders, CEO data, etc. I am trying to test the significance of the various variables in predicting company bankruptcy (the outcome variable).

I am trying to decide whether fixed effects (FE) or random effects (RE) models make more sense in my case. I like FE because this controls for omitted variables, but the major limitation I am facing is that FE requires the outcome variable to vary over time (i.e., companies to go bankrupt at some point in time). Therefore, all companies that never go bankrupt in my sample (about 95%) are excluded from the models when I use FE. This limits my sample considerably, and I would think does not make theoretical sense (is there not something important in the data on companies that never go bankrupt?). RE does not have this same requirement and so I can run regressions (using xtlogit, re) with RE for my full sample (~2000) companies.

Am I thinking about this the right way? Any advice on how to go about deciding between FE and RE in this specific situation would be appreciated. Maybe the ideal solution is to report results using both FE and RE; however, if this is the case, I would still like a some stronger footing to know which results to lead with... what makes the most sense...

In advance, thank you.

Roger

[Phil Bromiley]

While you can do a statistical test, I suspect that the issue is more theoretical. You want to differentiate between firms that go bankrupt and those that don't. That is at least equally important to exactly what year they go bankrupt. I think you are almost automatically pushed toward random effects. If you want the benefits of both, consider xthybrid.

[Roger Clements]

Thanks for the quick reply Phil. This is very helpful. I have two follow-up questions, when you have a moment:

1. How does RE consider the year of bankruptcy whereas FE do not? I thought both take into account the year.

2. In the same model, I am trying to include a moderator. I have data on the actions bankrupt firms took just before filing for bankruptcy and I am trying to see if these actions moderate the relationship between CEO characteristics (IV) and the probability of bankruptcy (DV). However, I think my problem is that I only have data on the firms that actually went bankrupt (i.e., treatment group), so there is no variation in the group (between bankrupt firms and non-bankrupt firms) on this moderator variable. Am I right to discern that I cannot include a moderator for which I have data only for the bankrupt firms?

[Roger Clements]

Phil Bromiley: I realize my second question may have been getting off topic for this thread, so I posted the question in a new thread with a different context (https://www.statalist.org/forums/for...on-interaction). Thanks.