Random effect GLS despite hausman test suggesting fixed effects

[Simon Frei]

Hello everyone!

I'm doing an analysis of the impact of board composition (different types of diversity, etc) on company performance. As the data is in a panel structure (80 companies over 6 years) I intended to use xtreg.

Now, in order to chose either random effects and fixed effects I did a Hausman test which clearly indicated that fixed effects were preferrable (chi2(11) = 548.65 /// Prob>chi2 = 0.0000). Since the variables I'm using (e.g. board size, gender diversity, etc.) are however rather stable for each company over the observation period, a random effect model seems to be preferrable.

A number of studies I read chose the random effect model based on the argumentation that their data only shows little variance. However in their cases, the Hausman test never rejected H0 and was thus in favor of random effects.

Thankful for any advice on how I could tackle this issue.

Best regards,
Simon

[Sebastian Kripfganz]

You might want to look into so-called "correlated random-effects" / hybrid / Mundlak models.

[Phil Bromiley]

A significant Hausman can be interpreted as suggesting that the effect of stable firm differences differs from the transient effect over time. In many cases, we know this makes perfect sense - transients in income influence some individual behaviors differently than stable income for example. In firm data, year-to-year variation in sales may influence boards differently than the base size of the company. Firms don't change their board sizes based on year to year changes in sales, but firm size as measured by sales could influence board size. In addition to Mundlak, you might look at the Stata routine xthtaylor.