Panel model with fixed time effects, residuals not normally distributed

[Jovana Ju]

Hello, everyone, nice to be member of this group :-)

First of all, I would like to apologize in advance if I ask too basic questions, but working with panel data and Stata is completely new to me.

I have data for 550 companies for period 2009-2018 (since it is unbalanced panel, I have 4576 observations). I am trying to find appropriate model specification, so I conducted F test and BP test for individual effects, and they showed that individual effects are not statistically significant (both, fixed and random). Because of large N and small T dimension, I did F test for fixed time effects, that turned out to be statistically significant. I included them in model through time dummies and run the regression (OLS). I hope that all I have done is correctly :-)

Tests showed me that model has problem with heteroskedasticity and autocorrelation. But, what worries me more is the fact that residuals are not normally distributed (according to Shapiro-Wilk test). How can I deal with these problems?

I need to mention that I am trying to repeat existing model from literature in terms of model variables: dependent, independent.

Thank you all for reading. If someone can give me advice, I would be grateful.

[Carlo Lazzaro]

Jovana:
1) if you have panel data with a continuous regressand, go -xtreg,fe- if you're interested in -fe- specification;
2) if you detected heteroskedasticity and/or autocorrelation, simply invoke -cluster- or -robust- (they'll give you back identical results);
3) simply ignore the non-normality of residual distribution: with 4756 observations any minimal departure from the gaussian makes Shapiro-Wilk test crying wolf.

[Jovana Ju]

Dear Carlo,

Thank you very much for help.

I have dilemma about something you said - xtreg, fe specification. I have already tested fixed individual effects, using F test, and it showed me that they are not statistically significant. Other researchers, who did the same analysis, used pooled specification (often). I thought it is correctly to do these tests first (individual and time effects) and then determine specification that fits the data, based on their results. Am I right?

Thanks once again, I read many of your comments and they were very useful to me.

[Carlo Lazzaro]

Jovana:
yes, your approach is correct.
if there's no evidence of group-wise effect, pooled OLS is the way to go.

[Jovana Ju]

Thank you, Carlo, once again.