Negative hausman test

[Ward Bruurs]

Hello Stata-fan

I am conducting a research with a xtpoisson regression. Now I have performed a hausman test (without exposure or robust which I do need in the regression) to determine whether I should use an FE or RE model. As a result I get a negative chi (-20) which is actually not possible in theory. Now I'm wondering how to determine which model (FE or RE) I need. In my mind, an FE model seems more appropriate as the dependent variable is the number of pharmaceutical companies in a country.

Thank you
Ward Bruurs

[Carlo Lazzaro]

Ward:
a negative -hausman- outcome tells you that you cannot reject the null (see Example 2, -hausman- entry, Stata .pdf manual, page 1032).
However, if you go -robust-, you cannot do -hausman-.

[Joao Santos Silva]

This is an interesting problem.

Ward Bruurs, I suggest you use FE because it is a much more robust estimator (Wooldridge, 1999).

Carlo Lazzaro, thanks for pointing out that example. An alternative interpretation is that under the null the test has a chi-2 distribution and therefore a negative value is evidence that the null is not valid. Maybe others can provide additional references?

Best wishes,

Joao

[Carlo Lazzaro]

Just downloaded (with only a quick glance given) The Hausman Test Statistic can be Negative even Asymptotically (degruyter.com)

[Ward Bruurs]

Carlo Lazzaro Joao Santos Silva

Thank you very much for the clarification and sources.

[Joao Santos Silva]

Thank you very much, Carlo Lazzaro; that seems to confirm the view that a negative test statistic is not compatible with the null.

Best wishes,

Joao

[Carlo Lazzaro]

Joao Santos Silva: you're right! Thanks for starting this interesting literature-based debate (far to be solved, I fear).