[Stephen Zhang]

thanks

[Sam Murgatroyd]

Hi Carlo,

I was interested in the advice you gave to Jenny in post #10 (quoted) below, and was hoping you could elaborate.

[QUOTE=Things are different with -xtreg-, as you are requested to -xtset- your data first: hence, Stata knows from the start that you're dealing with panel data.
Hence, clustering/robustifying standard errors with -xtreg- is not manadtory: it makes sense if you suspect heteroskedasticity/autocorrelation in your data (by the way, the latter is quite immaterial as long as you're dealing with a large N, small T panel dataset, as it frequently appears to be the case on this forum); otherwise, default standard errors are enough..[/QUOTE]

In my field, researchers often use panel data with fixed country effects to adjust for the associations between outcomes (y); but this doesn’t address the dependence of outcome (serial correlation). To my understanding, this means that using xtreg with country (and year) dummies will still produce biased estimates (because the lag of y has not been included). It is not clear to me how clustering the standard errors addresses the issue of serial correlation and I was hoping you could elaborate on that.

Thank you!

Sam

[Jeff Wooldridge]

Sam: As I mentioned in another thread, you can't conclude that the static model is "biased" just because a lagged value of y is significant. The models identify different parameters under different assumptions. The clustering is when you don't want to include a lagged y -- such as with policy analysis by difference-in-differences -- and you know the idiosyncratic errors, u(i,t), have serial correlation across t.