Questions about deviance
Prof Brian D Ripley <ripley at stats.ox.ac.uk> writes:
On Mon, 24 Jul 2000, halvorsen wrote:
I have experimented with the cheese data example from McCullagh&Nelder, page 175. With a proportional odds model they obtain a residual deviance of 20.31. Estimating the same model with polr(MASS) gives a residual deviance of 762.11 !, while using ordglm(gnlm) gives a deviance of 523.94. Can anybody explain these differences?
It depends on the choice of saturated model. See MASS3 pp 230-1, including how to compare them. It is a standard problem with discrete glms, and for this problem (which is not a glm). I maintain that our choice, which amounts to minus twice sum of log predicted probabilities, is by far the most interpretable. (Peter McCullagh had a thing about their lack of value, but he is not supported by the whole prediction assessment industry.)
I think Peter McC's point was that they are useless as measures of discrepancy between model and data, which is pretty obvious once you realise that the deviance for the model with p=.5 for all observations is a constant, so the deviance for any data is equivalent to the test for "all p==.5".
O__ ---- Peter Dalgaard Blegdamsvej 3 c/ /'_ --- Dept. of Biostatistics 2200 Cph. N (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk) FAX: (+45) 35327907 -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-help-request at stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._