Fixed-effect regression with cross-sectionally correlated residuals.

[Karl Lamber]

Hi all,

I currently analyzing a large panel dataset with N=700 T=125. My dataset suffers from both cross-sectional and time-series correlation; therefore, robust standard errors are downward biased because it only corrects for time-series correlation.

In order to adjust standard errors for both issues, I use the following three models:

1) fixed-effect with time dummy and cluster by id:

xtreg y x i.time, fe vce(cluster id)

2) fixed-effect with multiway clustering:

reghdfe y x, absorb(id time) vce(cluster id time)

3) fixed-effect with Driscoll and Kraay standard errors:

xtscc y x, fe

Model (2) and (3) are recommended to me in a previous post.

My main concern is that I am getting a different coefficient of x when I run model (3) - xtscc - than when I run model (1) and (2). Also, the r-squared is significantly larger when I run model (1) and (2).

I would appreciate it if someone can clarify the issue I am facing as I am not an advance stata user.

Thank you.

[Jesse Wursten]

I think you forgot to add the time dummies in (3)?

[Karl Lamber]

Jesse Wursten Yes, the only reason is that I am unable to add categorical variables (i.e. i.time) in model (3). Is the standard error in model (3) - xtscc - adjusted for cross-sectional dependence? Should I add a time dummy?

[Jesse Wursten]

No idea, I've never used the command. You could try adding the xi: prefix, then it might accept factor variables (it is the old syntax, e.g. xi: xtscc depvar i.time, fe).