Panel regression: Deciding on fixed effects and random effects - Hausman

Dear members ,I have a problem with my thesis while using regression.
In short: My fixed effects model is not significant and has no meaningfulness whatsoever.
My random effects model is significant and could be used.
But: Hausman tests strongly suggests, that I should use fixed effects and, using logic on my data, I would say so myself.

What now?

Longer story:
I want to assess the relationship between companies economical performance and its ecological one, using CO2 emissions. I have 3 countries and their largest firms in panels of 8 years. Using fixed effects, as far as I understand, only regards data and change in my panels, which is what I want to do. It does not make sense to compare company A´s CO2 reduction and its effect on financial performance with B´s, because their individual features (e.g. manufacturing or service industry) are way to different. But, as I told you, there is no result in the fixed effects regression.Another way would be using pooled method, but, as far as I see, it is wrong because pooled cross sections refer to multiple years where individuals are different. In my data, however, individuals are identical.

Any suggestions? Thanks.


Active Member
Hausman test is only one of criteria when comparing the fixed effects framework to a random effects framework. You should also check how one approach fares against the other in terms Akaike Information Criterion, Bayesian Information Criterion, goodness of fit or something similar.

If fixed effects is indeed the way to go, you should embrace this framework and then drop the non-significant terms one by one.


No cake for spunky
I have never understood how a fixed effects model would not be significant and the random effects be significant. It seems like a parameter being significant when the whole model is not (the model F test is not). That seems nonsensical.