Interpreting an F-test result to choose between FE and POLS

[Fernando Carnot]

I am working on a panel model and I have to choose between FE, RE, and POLS.

My first step was to perform the Hausman test (between FE and RE) and the results allowed the rejection of the null hypothesis (Prob > chi2 = 0.0001), favoring the FE model.

The second step was to test RE x POLS (xttest0), and the test favored the POLS model (Prob > chi2 = 1.0000) - I was not expecting such a strong result.

The third test was to check the FE's F-test (that all U[i] = 0). The test's result was 1.72 (Prob > F = 0.0602). My intention was to use this test to choose between FE and POLS. However, I don't know how to interpret this result. Is it low enough to favor the FE model? Should I try another approach/test?

Thanks in advance.

By the way, the POLS model has normal errors while the FE's errors are far from being normal. Does it matter in this case?

[George Ford]

https://www.statalist.org/forums/forum/general-stata-discussion/general/1523680-pols-fe-and-re

[Fernando Carnot]

Thank you for your reply, George.

I had already seen the post you mentioned but I could not find an answer to my question there.

Am I missing something?

[George Ford]

Is it a random sample?

[George Ford]

The FE are significant at the 10% level, but not at the 5% level (the null is jointly zero). So, it depends on how you feel about that. As a likely non-random panel dataset and the null at the margin of 5%, I'd probably include the FE.

[Fernando Carnot]

To be honest with you I don't know the answer. I know that elements from each panel change on every t. However, I have no access to the criteria employed to select them (probably a cluster random sampling).

The problem is: Neither I know how do I feel about the significance level. On the one hand 0,06 is pretty close to 0,05. On the other hand I am assuming a 5% level of significance on all of my tests.

Thank you again, George.

[Jeff Wooldridge]

This has never been a good testing strategy. I bet RE and POLS are exactly the same estimator in your case. Have the variables been differenced? If you reject RE in favor of FE then you logically have to reject POLS, too. The F test does not test what you want.

The only test that makes sense is the to use the Mundlak robust test.

Most importantly, FE is the most convincing method. If you’re getting usable results with it then stick with it. There’s no need to use a less robust estimator if FE gives precise enough estimates.

[George Ford]

Ain't gonna argue with Jeff.

[Fernando Carnot]

Thank you very much, Jeff.