Unbalanced Panel Data

[Harsh Tuli]

Q1) Due to some missing data for firms, my panel data is unbalanced. Should one run fixed effect for the same?
Q2) For a sample of 157 from a possible 500 listed index firms. The literature says for a not-so-random sample one should use FEM and not REM. Is this correct?
Q3) if autocorrelation, multicollinearity is present, do we use the cluster command with FEM or REM?

Thanks in Advance.

[Carlo Lazzaro]

Harsh:
1) panel unbalancedness has no bearing on the -fe- or -re- specifications;
2) I've never heard such an advice;
3) if autocorrelation and /or heteroskedasticity are detected, going -robust- or -vce(cluster panelid)- is ok. Multicollinearity is a different beast, that rarely hits hard (see https://www.hup.harvard.edu/catalog....9780674175440; Chapter 23).

[Harsh Tuli]

Carlo thanks for your prompt response. I am looking for a justification to what prof. Wooldridge and others suggested in an earlier thread that I should use vce cluster command with time fixed effects. Here is the link to the post https://www.statalist.org/forums/for...relation/page2

Also did he actually suggest that I should run fixed effect. I'm trying to find a justification for the same. Please see if you could find some literature for the same. I'm a little lost with it.

[Carlo Lazzaro]

Harsh:
1) a sample of 157 panel is enough to go -vce(cluster panelid)- standard errors;
2) if I'm not mistaken, one of Jeff Wooldridge 's favourite remarks on (linear) panel data regression is that researchers switch from -fe- to -re- when on the brink of desperation!

[Harsh Tuli]

I completely understand the same Carlo but i needed a justification as to why I didn't run the hausman and just chose fixed as my way to go. As most research papers compare the two. I hope you can help me with this.

Regards
Harsh

[Carlo Lazzaro]

Harsh:
while it is true that most papers report both specification, the issue is that the -fe- estimator is always consistent (but less efficient if -re- is the way to go), whereas the -re- estimator is not consistent if -fe- is the way to go.
In addition, the main assumption supporting the -re- estimator (zero correlation between u and the vector of regressors) seldom holds in practice.
All that said, probably the best advice is to abide by the tribal rules of yor research field.