On Wed, Jan 26, 2005 at 02:03:25PM +0100, Peter Dalgaard wrote: > Göran Broström <[EMAIL PROTECTED]> writes: > > > On Wed, Jan 26, 2005 at 11:01:03AM +0100, Peter Dalgaard wrote: > > > Göran Broström <[EMAIL PROTECTED]> writes: > > > > > > > > Try it with > > > > > > > > > > b=c(rep(0,8),rmultinom(1,24,rep(1/8,8))) > > > > > > > > This is the conditional distribution, given the sufficient statistic. > > > > > > Yes (and that was of course the point). I was surprised that there is > > > a difference at all (between this and rep(3,8) and also between > > > different calls to rmultinom). > > > > That surprises me :-) Really, if the observed values are exactly equal to > > the predicted values or not must make a difference? > > To the deviance, yes. It is less obvious why the table of coefficients > comes out different, given that everything inside of it is a function > of the same sufficient statistics. The reason is the termination > criterion in glm.fit, which is on the relative change in deviance, > > (I'm not too happy about that in principle.
I had the same thought. Would it be possible to automatically reduce to sufficient statistics inside glm, and would it be desirable? Should I, as a user, reduce to sufficient statistics before running glm? It shouldn't really matter, but I get different deviances and AIC, and that can be confusing. Note that this is the case also in situations where the fitting algorithm converges nicely. -- Göran Broström tel: +46 90 786 5223 Department of Statistics fax: +46 90 786 6614 Umeå University http://www.stat.umu.se/egna/gb/ SE-90187 Umeå, Sweden e-mail: [EMAIL PROTECTED] ______________________________________________ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html