Hausman Test negative

[silver rupa]

Dear All:
I am facing the following problem with Hausman Test
chi2(4) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= -0.00 chi2<0 ==> model fitted on these
data fails to meet the asymptotic
assumptions of the Hausman test;
see suest for a generalized test

What does this mean? Is this result OK ? should I use random effects?
Thanks in advance for your help!
Regards,
silver

[Carlo Lazzaro]

Silver:
singletons in your preditctors might be the answer.
That said, you can go -xtreg,re- and test if thsi is the way to go via the community-contributed modue -xtoverid-.
As per FAQ, please post exactly (and comprehensively) what you typed and what Stata gave you back. Thanks.

[silver rupa]

Thanks a lot for your answer. I will post with more details.
Regards,
silver

[Maxence Morlet]

You can also run the Mundlak test (1978): https://blog.stata.com/2015/10/29/fi...dlak-approach/.

[silver rupa]

Thanks Maxcene Morlet for your help
Regards

[Carlo Lazzaro]

Silver:
Maxence made a good point in advising you about the Mundlak approach.
On of its advantage is that you can impose non-default standard errors (something the -hausman- does not allow you to).
However, this procedure is more demanding than it would seem at its face-value.
Therefore, be 100% sure of getting all its requirements (e.g, including all the panel-speciifc means of the time-varying predictors) and -test- their joint statistical significance after -xtreg,re-.
If -test- outcome shows evidence of statistical significance, the main assumption of -xtreg,re- (that is, zero correlation between -u- and the vector of regressors) is rejected and -fe- is the way to go, whereas you should stick with -re- specification if -test- does not reach statistical significance..

[Anthony Macedo]

You can also run the Mundlak test (1978): https://blog.stata.com/2015/10/29/fi...dlak-approach/.

Can I use the Mundlak test with model regressions through xtlogit, fe and xtlogit, re ?
I got a negative chi2 and could follow Example #2, -hausman- entry, Stata .pdf manual, and interpret that result as no rejection of the null hypothesis, but I would feel more confident if I could find another way to test the RE.
For instance, without being sure that the approach is adequate for xtlogit, I tried the Mundlak test and the result also suggest that I can go with RE.
Note that I cannot use the sigmamore option of Hausman and Suest does not work because I don’t have a score function.

[Carlo Lazzaro]

Anthony:
as per your description, I would go -re- without further tests.

[Anthony Macedo]

Thank you Carlo.
But do you know if can use the Mundlak approach with Logit? It would allow me to consider robust standard errors.

[Carlo Lazzaro]

Anthony:
see: https://stats.stackexchange.com/ques...ression-with-d
https://www.statalist.org/forums/forum/general-stata-discussion/general/1631045-mundlak-correction-for-logit-probit-model

[Anthony Macedo]

Thank you Carlo. So it seems that it is not possible.

[Carlo Lazzaro]

Anthony:
yes, it seems so.
I'd take a look at -clogit-.

[Anthony Macedo]

Unfortunately, I think it does not help me much. I still can't test the validity of RE while using robust standard errors (at least with commands such as rhausman and xtoverid).

[Carlo Lazzaro]

Anthony:
would a linear probability model be feasible with your data?

[Anthony Macedo]

I'm afraid not because there is a risk of the predicted values falling outside the range 0 and 1. My dependent variable is binary, 0 means a decrease in popularity in month m, 1 means an increase.