In the past couple of months I've come across two cases of the sandwich
estimator giving standard errors of zero. One when someone (not me)
fitted the clustering variable as a fixed effect in a mixed logit model.
Another when someone else (me) used svyprop, subpop for a subpopulation
containing only one cluster. I found one reference to this problem in
the archive (pasted below - thanks, Bryan!) but no suggested solutions.
My limited understanding of the issue is this. The sandwich estimator
takes the pessimistic view that nothing is known about within-cluster
correlation, so it refuses to speculate about the accuracy of the
estimate. Or, algebraically, it collapses the score matrix across the
cluster so the "meat" in the sandwich becomes zero.
The alternative, optimistic, approach would be to model these
correlations and pretend that you've done it properly, but this needs
(I think) information about the joint sampling probabilities which will
not be available. Is there a middle way? Surely something can be said
about the variability of a single cluster estimate, when there are lots
of other clusters in the sample. Any suggestions, references gratefully
accepted. Geoff
st: RE: Svytab subgroup analysis with more than two subgroups
--------------------------------------------------------------------------------
