Correcting autocorrelation by adding year dummy variables?

[Kane Smith]

Hello everybody!

For my master dissertation, I'm dealing with panel data. In addition, I detected heteroskedasticity and autocorrelation in my dataset.
Until now, I only checked for heteroskedasticity as I had never heard of autocorrelation before. Therefore, I used White robust standard errors for all my models (until now).
(White robust standard errors = heteroskedasticity vs vce(cluster) = heteroskedasticity + autocorrelation)

The approach I use is OLS regression. In the first model, I run the standard regression. In the second model, I check for year dummies as well. According to someone in the literature, I avoid autocorrelation when checking for year dummy variables. Thus, in the second model I should only use White robust standard errors, in contrast to the first model. Do you agree?
However, if I use vce(cluster) in model 1, I observe big differences in the t-statistics compared to my second model (where I use White robust standard errors).

What should I do?

Thanks in advance!

[Carlo Lazzaro]

Kane:
welcome to this forum.
I'm not clear with your using -regress- instead of -xtreg- (as first choice, at least) when dealing with panel data (I assume that it is an N>T one).
The -robust- and -cluster()- options under -xtreg- deal with both heteroskedasticity and/or autocorrelation.

[Kane Smith]

Dear Carlo
I'm using both regress and xtreg. I'm familiar with the exact results that robust and cluster compute with xtreg.
So my question is about regress and yes N >T.

In OLS 1, I run the standard regression, whereas, in the second model, I check for year dummies as well. According to someone in the literature, I avoid autocorrelation when checking for year dummy variables. Thus, in the second model I should only use White robust standard errors, in contrast to the first model. Do you agree?
However, if I use vce(cluster) in model 1, I observe big differences in the t-statistics compared to my second model (where I use White robust standard errors).


Thank you!

[Carlo Lazzaro]

Kane:
in OLS, -robust- and -cluster- give back different satndard errors, because -robust- assumed that epsilon are uncorrelated across regression (as assumption that might be problematic with panel data and deserve investigation, even with a N>T array), wheres -cluster()- relax this assumption. In sum, whie -robust- is OK if you detect heteroskedasticity only, -cluster()- shoud be invoked every time you detect autorrelation (and, possibly, heteroskedasticity).
Eventually, I'm not sure that -i.year- can accomodate for autocorrelation in epsilon distribution.