Panel Data, random effects or fixed effects, heteroscedasticity test

[Nick Baradar]

Hello Everyone,
I have firm-level data for 20 years, where my dependent and independent variables are count variables. My DV is the number of specific type of financial announcements, and my IV is the number of tax inspections in a given year. I also have other control variables that are firm characteristics over the years - such as their domain and financial performance indicators. I log-transformed both my IV and DV, and lagged my IV by one year. I used xtreg command. At first, I tried running it with fixed effects, but it was dropping out the variables that are consistent over time, such as the domain where the companies are. So I tried running it with random effects, which I am not very sure is the right thing to do... It's been a while since I have been trying to understand if it was a good idea to use xtreg command, especially with random effects. Moreover, I am quite unsure if I should use the robust option as I couldn't find a way to check for heteroscedasticity after using xtreg command. I would highly appreciate it if someone could direct me in this process.

Thank you very much in advance,
Nick Baradar

[Carlo Lazzaro]

Nick:
I'd start with -xteg,fe- amd test for groupwise heteroskedasticity with the community-contributed module -xttest3-.

[Nick Baradar]

Dear Carlo, thank you for your advice. xtreg,fe command drops time-invariant variables such as the domain of the company. However, that information is very valuable for my analyses. Also, I am not sure what you mean "I'd start with", could you please share what your next steps would be?

[Carlo Lazzaro]

Nick:
usually, panel data analysis with continuosu regressand starts off with the -fe- estimator (within panel variation).
Then, you may wnat to go -re- and compare the two estimators via -hausman- (provided that you use default standard errors).
While it's true that the -fe- estimator wipes out time-invariant variables, the -re- one, in turn, is based on the assumption that the u component of the residual is not correlated with the vector of regressors (that may be untenable in real world research).

[Nick Baradar]

Dear Carlo, thank you for further clarification. I indeed performed Hausman test that shows preference for FE esimator. However, as I mentioned earlier, I have some time-invariant variables that are important from the research point of you. I was wondering if there are ways I could use to tackle that issue.

[Carlo Lazzaro]

Nick:
see Mundlak correction (The Stata Blog » Fixed effects or random effects: The Mundlak approach).