I came across a discussion of the paper by Alywyn Young that is linked below on Andrew Gelman's blog:
http://personal.lse.ac.uk/YoungA/ChannellingFisher.pdf.
Young points out on pages 20 and 21 that when computing a cluster-robust covariance matrix, "xtreg, fe" does not include the individual fixed effects in the degrees of freedom correction applied to the estimated covariance matrix. See also this blog post:
http://amarder.github.io/clustered-standard-errors/
I understand that there is an option "dfadj" that can be used to change the degrees of freedom correction so that the individual fixed effects are included. However, dfadj appears to be a hidden option, as it does not appear in the xtreg help menu or the manual, as far as I can tell. The number of observations will often be large enough that the impact on inference is negligible, but this seems like something that should be made obvious in the xtreg help menu rather than requiring one to happen to read about this issue in a paper, as I did.
I realize this has been mentioned on Statalist before. I also realize that proper degrees of freedom corrections are far from a settled issued with respect to robust covariance matrices, but it seems odd that the default is to make standard errors more precise than they would be if Stata incorporated what is usually agreed upon as the more correct degrees of freedom correction term, and that the alternative is to use an obscure option.
http://personal.lse.ac.uk/YoungA/ChannellingFisher.pdf.
Young points out on pages 20 and 21 that when computing a cluster-robust covariance matrix, "xtreg, fe" does not include the individual fixed effects in the degrees of freedom correction applied to the estimated covariance matrix. See also this blog post:
http://amarder.github.io/clustered-standard-errors/
I understand that there is an option "dfadj" that can be used to change the degrees of freedom correction so that the individual fixed effects are included. However, dfadj appears to be a hidden option, as it does not appear in the xtreg help menu or the manual, as far as I can tell. The number of observations will often be large enough that the impact on inference is negligible, but this seems like something that should be made obvious in the xtreg help menu rather than requiring one to happen to read about this issue in a paper, as I did.
I realize this has been mentioned on Statalist before. I also realize that proper degrees of freedom corrections are far from a settled issued with respect to robust covariance matrices, but it seems odd that the default is to make standard errors more precise than they would be if Stata incorporated what is usually agreed upon as the more correct degrees of freedom correction term, and that the alternative is to use an obscure option.