I have just committed some code to the r-devel branch to implement the Rao
efficient score test. This is asymptotically equivalent to the LRT, but there is
some indication that it might have better properties in smaller samples since it
is based more directly on the distribution of the sufficient sums under the null
hypothesis (e.g., if you have a divergent fit to the model under the
alternative, the chi-square distribution of the likelihood ratio may be dubious
and the Wald test is surely wrong).
It works fairly obviously by using test="Rao" in various methods. See
for instance the last part of example(anova.glm).
While going over the code, I took the liberty of standardizing the labeling of
the p-value column to Pr(>Chi) and Pr(>F). This may break some scripts if
they extract p-values using constructs like
anova(glm.D93,test="Chisq")$"P(>|Chi|)"[2]. Apologies for
the disturbance, but this sort of change just doesn't let itself be
implemented with the usual deprecation sequence. The workaround should be
straightforward (if you need something that works in multiple versions, then try
both and choose the result that isn't NULL).
--
Peter Dalgaard
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk Priv: PDalgd at gmail.com