My suggestion is: run your -xtreg, re- model. Then download -xtoverid- from SSC and run it: if the test is significant, then you should prefer to run -xtreg, fe-. The p-value should be very similar from that coming out of:
xtreg y xs, fe
est store fixed
xtreg y xs
hausman fixed .
where xs include all variables in both cases (Stata will drop the time-invariant variables in the FE model for you). I like software taking decisions instead of me, above all in the cases it comes from trustful sources like prof. Schaffer & Stillman :-)
Nicola
P.S. please note that Baltagi et al. (Economics Letters 79, pp. 361-369) suggested to do two Hausman tests, in order to choose between RE, FE, and Hausman-Taylor estimators.
At 02.33 17/04/2008 -0400, "Pavlos C. Symeou" wrote:
>Dear Statalisters,
>
>I have a question regarding the Hausman specification test. I want to include in my OLS model 7 time-constant variables which depict the industry classification
>of a firm. In order to use the Hausman test, I first estimate the FE model, I store the estimation results, then run the RE model and finally ask Stata for the
>estimation of the Hausman test. As it is expected, the time-constant variables are dropped from the FE model but not from the RE. This implies that the Hausman
>test compares two different models; one with time-constant variables and one without. The Hausman test rejects the RE model. If I omit the time-constant
>variables from the two models, then the Hausman test compares comparable models and still rejects the RE model. Yet, inclusion of the time-constant variables in
>the OLS model is important for its correct specification so I interact these variables with year dummies and include the interaction variables (and exclude the
>original time-constant variables) in both FE and RE models. I rerun the Hausman test which now rejects the FE in favor of the RE model. I am confronted here
>with a dilemma where I can only perform the Hausman test if I interact my time-constant variables with year dummies, but what I actually need from my final
>model is the inclusion of the original time-constant variables, which however leads to the wrong estimation of the Hausman test.
>
>I would appreciate if you could advise on this. Should I arbitrarily choose the RE model that can involve the time-constant variables in its estimation? Otherwise?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
