OLS vs FE

[Donovan Pollack]

Greetings,

In stata I can get FE results by using OLS with time and country dummies. Is it ok that time invariant things are estimated in this OLS model with FE?

I have: reg y x1 x2 i.year i.country, and I cluster by country. if x2 doesn't vary overtime, it does not drop out of this model, but interpretation is confusing?

thank you,
Donovan

[Carlo Lazzaro]

Donovan:
welcome to the list.
Pooled OLS is rarely a good alias of -xtreg, fe- exception made for the lack of statistical significance of the F-test appearing at the foot of the -xtreg, fe- outcome table (if default standard errors are imposed; otherwise, the abovementioned test is too difficult to calculate and Stata omits it). See Example 2 under -xtreg- entry in Stata .pdf manual.

[Donovan Pollack]

Thank you Carlo. I will use xtreg, fe in my model. To clarify, you are saying that stata omits the F-test in this case unless default SE are use --how would one do this? Are you also saying that in the pooled model, we cannot interpret the time-invariant coefficients?

thank you,

[Carlo Lazzaro]

Donovan:
q1)

xtreg y x1 x2, fe///uses default standard errors (ie, no robust, cluster or other options)

q2) -fe- machinery rules out time-invariant predictors, whereas pooled OLS does not do that. However, in most real-life panel database is wiser to rely on -xtreg, fe- than making your life more difficult with a pooled OLS (see my previou reply).

[Donovan Pollack]

Thank you, any suggestions on how to also measure the time invariant variables, I have seen FEVD methods, or is this inappropriate?

[Carlo Lazzaro]

Donovan:
if your goal is to measure time-invariant variable, -fe- will not support you; -re- will do it. You do not say if you compared -fe- vs -re- specification via -hausman-.

[Donovan Pollack]

ok, thank you for all the advice