I have tightened the tolerances in glm.control in R-devel (aka 1.8.0 Under Development) from epsilon = 1e-4 to 1e-8, and increases maxit from 10 to 25. Normally the effect is to do one more iteration and get more accurate results. However, in cases of partial separation several more iterations will be done and it will be clearer from the results which parameters are converging and which are diverging (to +/-Inf). I have been meaning to do this for some time (the defaults seemed appropriate to computer speeds at the 1991 origin of glm in S3), but have only this time remembered at the beginning of a release cycle. -- Brian D. Ripley, ripley@stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595
>>>>> "BDR" == Prof Brian Ripley <ripley@stats.ox.ac.uk> >>>>> on Fri, 9 May 2003 12:08:10 +0100 (BST) writes:BDR> I have tightened the tolerances in glm.control in BDR> R-devel (aka 1.8.0 Under Development) from epsilon BDR> 1e-4 to 1e-8, and increases maxit from 10 to 25. BDR> Normally the effect is to do one more iteration and get BDR> more accurate results. However, in cases of partial BDR> separation several more iterations will be done and it BDR> will be clearer from the results which parameters are BDR> converging and which are diverging (to +/-Inf). BDR> I have been meaning to do this for some time (the BDR> defaults seemed appropriate to computer speeds at the BDR> 1991 origin of glm in S3), but have only this time BDR> remembered at the beginning of a release cycle. Very good! Being on this topic: What about enhancing summary.glm() and changing print.summary.glm() quite a bit such that o summary.glm() for back compatibility still computes the Wald tests but also does all the "drop1" tests for each coefficient and o print.summary.glm() would print these tests instead of the (too often misleading) Wald ones. Related: How hard would it be to compute a ``Hauck-Donner diagnostic'' warning? (If it's an open research problem then it's "too hard" for R 1.8 :-) Martin Maechler <maechler@stat.math.ethz.ch> http://stat.ethz.ch/~maechler/ Seminar fuer Statistik, ETH-Zentrum LEO C16 Leonhardstr. 27 ETH (Federal Inst. Technology) 8092 Zurich SWITZERLAND phone: x-41-1-632-3408 fax: ...-1228 <><