Hausman or Mundlak?

[Sundus Al Riyami]

Hi,

I have a panel data with T>N. I want to test whether FE or RE is more appropriate. My Hausman test Indicates RE whilst the mundlak test indicates FE.
Is the Mundlak approach better if my standard errors are heteroskedastic?

[Rhys Williams]

Yes, the Mundlak test is generally considered better when one has heteroskedastic errors.

The following articles may be of interest: https://blog.stata.com/2015/10/29/fi...dlak-approach/
https://stats.stackexchange.com/ques...ects-retrieved

The Mundlak help file points out that:

The command mundlak estimates random-effects regression models (xtreg, re) adding group-means of variables in indepvars which vary within groups. This technique was proposed by Mundlak (1978) as a way to relax the assumption in the random-effects estimator that the observed variables are uncorrelated with the unobserved variables. Additionally, the degree of statistical significance of the estimated coefficients on the group means can be used to test whether such assumption holds for individual regressors. See also Chapter 10 in Wooldridge (2010) and Chapter 11 in Greene (2011).

Best,
Rhys

[Joro Kolev]

You should show exactly what you typed in Stata, and exactly what Stata returned. At the moment you are telling us your interpretations of what happened, and we are supposed to interpret your interpretations instead of actual Stata output/results.

In general the Hausman test is not valid if you have heteroskedasticity or some sort of correlation (auto, or cluster correlation), so if you have those, the Mundlak's approach is the only option.

[Sundus Al Riyami]

Thank you Rhys and Joro!

[Carlo Lazzaro]

Sundus:
if you actually have a T>N panel dataset with a continuos regressand, -xtreg- is not appropriate. See -xtgls- and -xtregar-, instead.

[Sundus Al Riyami]

I will be using Driscoll and Kraay standard errors (-xtscc-)... is -xtgls- more appropriate?

[Carlo Lazzaro]

Sundus:
-xtscc- looks ok to me assuming that you also have across-panel correlation.

[Joro Kolev]

I slightly here disagree with the way how you are stating your data structure, and Carlo did not call you out on that.

I think the key is what are you ready to assume that goes to infinity?

You say T>N, but it still can be that you have 50 US states over 51 years... The statement T>N says something, but not that much after all.

-xtscc- and -xtgls- are appropriate if you are ready to assume that your T grows large.

If you are ready to assume that both your N and T grow large, a lot of methods become appropriate, including -xtscc- and -xtgls- but also the methods for wide panels with short T.

I will be using Driscoll and Kraay standard errors (-xtscc-)... is -xtgls- more appropriate?

[Sundus Al Riyami]

Dear Joro,

I have 18 countries and around 40 years.

[Joro Kolev]

Yes, this does sound like a long panel, -xtscc- and -xtgls- are appropriate for such panels.

The only headache with long panels is that you need to worry about the time series properties of your variables, that is, whether you have deterministic or stochastic trends.

Dear Joro,

I have 18 countries and around 40 years.

[Carlo Lazzaro]

Sundus:
I do share Joro's helpful comment. As headaches are annoying to bear, this is probably why most of the methological literature is about N>T panel datasets.
On a different tone, Stata Press may consider a new textbook on how to deal with long panel datasets with Stata.