Difference between robust standard errors with xtreg fe and reg with dummy variables

[Emmanouil Tranos]

Hello,
I am estimating the same FE model twice: once with xtreg fe and once with reg including dummy variables. For both cases I ask for robust standard errors. Although the coefficients are the same as they should be the standard errors are different. Why?

This is the sample code:

xtreg y x1 x2 x3 if sample==1 [aw=weight], fe robust
reg y x1 x2 x3 i.countryid if sample==1 [aw=weight], robust
Thank you,
Em

[Sebastian Kripfganz]

xtreg, vce(robust) is equivalent to reg, vce(cluster countryid) because xtreg always takes the panel structure explicitly into account.

[Carlo Lazzaro]

Em.
welcome to the list.
Sebastian touched on a relevant issue.
Whereas using -vce(robust)- or -vce(cluster clusterid)- makes no difference under -xtreg-, these options performs quite differently under -regress-.

[Emmanouil Tranos]

Thank you both, I think understand the difference.

My data is an unbalanced panel of countries for 15 years. Given that I don't have any theoretical reasons to believe that there is an underpinning structure in my data apart from the one depicted by the FE (countries in this case), should I continue with xtreg, fe vce(robust) or with with reg and robust standard errors?

Best regards,
Em.

[Nick Cox]

Em Qwert: Please note our strong preference here for full real names.

http://www.statalist.org/forums/help#realnames

http://www.statalist.org/forums/help#adviceextras #3

[Emmanouil Tranos]

Point taken and I have requested from the admin to change it. Apologies.

[Nick Cox]

Thanks for #6.

[Carlo Lazzaro]

Em:
thanks fo re-registering.
You should go -xtreg,fe- if you have both a -panelid- and a time variable (or, put differently, you have repeated weaves of data obtained via measuring, in general, the same sample of units >1 time about a given set of variable).
If you have one wave of data only (that is, one measure only of a given set of variables for different sampe units), you should go -regress- (here and above I assume that you have a continuous -depvar-).
-xtreg,fe- advises you to switch to -regress- whenever the F-test at the foot of -xtreg,fe- outcome tables fails to reach statistical signigìficance.
Unfortunately, that test is availableif -xtreg,fe- is performed with default standard errors.

[Emmanouil Tranos]

Carlo,

Many thanks for this. It is the first: -panelid- and a time variable. The LHS variable is indeed continuous and is the outcome of another estimation hence the weighting. So, my question is that given all the above (panel, fe, weighting), do I need to report the standard errors or the robust standard errors?

Best,
Em

[Carlo Lazzaro]

Emmanouil:
under -xtreg- robustified (or clustered) standard errors make sense if you suspect heteroskedasticity and/or autocorrelation (the latter is usually a minor concern if you're dealing with a large N, small T panel dataset).
I would also recommed you to take a look at -help weight- where you can find how -aweight- are internally treated by Stata.

[Emmanouil Tranos]

thank you again!