Thanks to persistent queries from "KTTK" on sourceforge, we now have
some fixes and refinements for the fixed-effects and random-effects
For extended background see
The following changes are in current git and snapshots.
1) I think this has been mentioned before but I'll state it for
completeness: when you give the --robust option with fixed effects,
the accompanying test for difference of group means is now also a
2) Also pertaining to fixed effects with the robust option: we show
clustered (Arellano-type) standard errors and the joint F-test, based
on the covariance matrix, is also robust. However, we were using the
wrong denominator degrees of freedom for this test. In the clustered
case this should be based on the number of clusters (that is, groups,
in the panel case) and not the total number of observations. Now
3) In the case of random-effects estimation (unlike most estimators),
we were not showing a joint test on the regressors. We now show a Wald
chi-square test by default.
4) While working on the above, I noticed a problem that hasn't been
reported yet: if you gave the --time-dummies option with random
effects, you could get a spurious "insufficient degrees of freedom"
error. That's because when we were checking the degrees of freedom for
the "between" model we were wrongly including the time dummies in the
parameter count -- wrongly, because of course these dummies will wash
out of the between model. Now fixed.
5) KTTK asked if it's just an oversight that there's no robust option
for random effects. No, not just an oversight; it's not so obvious
what "robust" variance estimation should look like in this case.
However, we'll take another look at this.