sktest displaying strange results

[Solomon Lin]

Hi,

I am running the sktest after using "predict varname, e" after running a fixed effects regression. I find that when I use sktest, the results are:

Pr(Skewness)=0.0000 or 1.0000
Pr(Kurtosis)=0.0000
adj chi2(2)= .
Prob>chi2= .

My sample is has 8,000 observations and I read online that sktest may not be appropriate for large samples. Is there an alternative test I can use (I know swilk also can't be used for large samples) that is appropriate for panel data?

[Joseph Coveney]

You could try

pnorm varname

and

qnorm varname

and even

histogram varname

and make a judgment call whether the plots indicate something large enough to worry about.

If you want some quantitative guidance, then perhaps you could bin your residuals and generate a similar distribution of residuals in simulations under the null hypothesis of scientific interest in order to see how badly the test size is affected. And under the alternative hypothesis of minimum effect size of scientific interest in order to see how badly power is affected.

You can make the bin width as narrow as you want, but if binning is still too coarse for you, then you could use the user-written pair of programs jnsn on the observed residuals followed by ajv to generate a similar distribution of residuals for use in simulations. Both are available from SSC (search jnsn).

[Nick Cox]

I agree with Joseph. The best thing to do is look at qnorm results.

More generally, many discussions can be found on Statalist of the limitations of these tests. Here is one: http://www.stata.com/statalist/archi.../msg00306.html