Academic due diligence and robustness testing for a Fixed Effects model

[Ayrton Da Silva]

Hello Statalist!

I have a large N, small T panel dataset on which I have run some fixed effects regressions. The model is predicting industry adjusted firm performance based on independent variables from one time period before. The choice of FE is supported by the Hausman test.

I am looking for some advice on which tests one would need to perform to demonstrate the valid model with BLUE estimators? I presume this would be all the Gauss Markov assumptions, but am unsure of how to test them.

I have the following conditions (and in some cases tests):
Normality of Residuals

predict rs
kdensity rs, normal
pnorm rs
swilk rs

Homoskedasticity

predict Fitted, xb
predict Epsilon, e
twoway (scatter Epsilon Fitted), ytitle(Epsilon residuals) xtitle(Fitted values)

Multicollinearity

pwcorr var1 var2 ... vark

Autocorrelation
Test unknown

Serial correlation
Test unknown

Exogeneity
Test unknown

Any advice on the assumptions which need to be tested and additional tests which would need to be performed would be most appreciated.

Thanks all!

Ayrton Da Silva

[Ayrton Da Silva]

Any help on this would be greatly appreciated -- thank you!

[Justin Niakamal]

You can find a lot of discussion on these topics here in Statalist.

For example:
Post #19 in this thread
https://www.statalist.org/forums/for...gression/page2

Post #4
https://www.statalist.org/forums/for...lity-of-errors

etc.

[Joro Kolev]

If you are doing research, the current thinking is than in large N small T the only important test is the Hausman test of random vs fixed effects. And you should calculate standard errors robust to heteroskedasticity and arbitrary within T correlation (vce(robust) option).

If this is a homework, just ask your teacher what tests he wants you to perform.

[Sebastian Kripfganz]

The following Stata FAQ might also be of help:
How do I test for panel-level heteroskedasticity and autocorrelation?
Note that autocorrelation and serial correlation are essentially the same thing. Also, normality is not one of the Gauss-Markov assumptions and not required for consistent estimation and inference.

[Ayrton Da Silva]

Hey Justin Blasongame, Joro Kolev and Sebastian Kripfganz

Thank you so much for your responses. Those resources were super helpful. It is for research.

Have a great day!