Panel Data Question I could not find in any book

[Maxy Smith]

Hi All,

I have read several manuals on panel data econometrics but I never found the answer to these very basic questions.

I have a panel data with about T=1,000 days (per stock) and N=5,000 stocks. I want to regress daily cumulated returns (over a 21-horizon) over a set of variables (some of which are persistent).

1. How do I decide whether to include time fixed effects (i.e. time dummies)?
2. If so, do I need to do so based on the frequency of the data (i.e. include daily dummies if I have daily data and so on)?
3. I’m interested in making inference about the CROSS-SECTION i.e. whether stocks with a certain characteristic have higher or lower returns. If I include stock fixed effects does my coefficient have more a time-series interpretation instead because it is like using the "within-estimator"?
4. If I cluster the standard errors in the time dimension, do I need to do so based on the frequency of the data (i.e. cluster by day if I have daily data and so on)?
5. Can I double cluster standard errors even if I do not include stock and time dummies?

Thanks a lot!

Max

[Carlo Lazzaro]

Maxy:
welcome to this forum.
It's strange that your teacher/professor/supervisor gave you no methodological hints about how to analyze your data.
Anyway:
1) you have a large N, large T panel dataset. In this instance, the categorical variable to be included among predictors related to -panelid- not time. In your case, I would consider -i.industry-.
2) see 1);
3) you may want to consider -xtregar,fe-.
4) standard errors shoud be clustered on -panelid-, -not -timevar-.
5) see the user-written programme -reghdfe-, that you can spot and install by typing -search reghdfe- from within Stata.