Compare random effects to fixed effects?

[Tabea Mueting]

Hi all,

I estimated a model with fixed effects, using data for Germany (the Hausman Test suggested me to use fixed instead of random effects). There is an existing paper which does exactly the same regression as I do, but which uses random effects and data for Switzerland. If possible, I'd like to compare my results to the results of that paper. Is it possible to do a quantitative comparison if one model is estimated with FE and the other one with RE? I'd say no, but I'm not quite sure if there isn't a way I haven't thought about.

Thanks!

[Phil Bromiley]

If you want to directly compare results, use RE on your model. Your positive Hausman indicates your parameters for FE differ from those for RE, so there is no reason to believe they will be the same as someone else's RE estimates. Did the Switzerland study report a Hausman test?

By the way, you'll increase your chances of a helpful answer by following the FAQ on asking questions.

[Tabea Mueting]

Thanks a lot, I'll check the FAQ.
The Switzerland study did not report a Hausman test, they just stated that they use RE. So you suggest to re-estimate my model with RE (despite the Hausman test) and to compare those estimates with the results of the Switzerland study?

[Carlo Lazzaro]

Tabea:
choosing between RE because Others used this specification in the past, or going -FE- because, as far as your data are concerned, -hausman- outcome points you that way is your own judgement call.
As you cannot take for granted that your data are similar to those reported for Germany, I would focus on the Swiss sample and go -FE-.

[Tabea Mueting]

Thanks a lot!

[Eric de Souza]

If the paper using Swiss data included time-invariant variables in the model, this could be a reason why the author(s) used RE. If this is the case then a Hausman test for RE versus FE would not be useful without correction because the number of regressors in the FE and the RE versions of the model would not be the same. Also you cannot use the Hausman test if you use robust standard errors (option robust or cluster).

[Carlo Lazzaro]

Tabea:
as an aside to Eric's helpful advice, if you plan to robustify/clustering your standard errors, you should replace -hausman- with the user-written programme -xtoverid-, which you can spot by typing -search xtoverid- from within Stata.

[Tabea Mueting]

Thanks a lot, using xtoverid gives me the same results as the Hausman test does. This suggests using fixed effects, but I am a bit concerned about the Nickell bias, as my N is really large and I only have yearly data from 2007 to 2016.

[Eric de Souza]

You did not say that you have a dynamic panel data model (the dependent variable lagged as explanatory variable). In that case neither FE nor RE is suitable.

[Carlo Lazzaro]

Tabea:
Eric's point highlights once more the need to post what you typed and what Stata gave you back (via CODE delimiters, please).
We exchanged 7 emails before discovering that you are dealing with a dynamic panel data model that requires (say) -xtabond- instead of -xtreg-.