Tobias,

The GLIMMIX macro(s) which was initially constructed under the direction of Russ Wolfinger did report a model deviance. When the GLIMMIX procedure was released, this model deviance was no longer reported. In David Murray's 1998 manuscript "Design and Analysis of Group-randomized Trials", this is said about the deviance term reported for the GLIMMIX procedure:

The deviance and scaled deviance statistics can be interpred as goodness-of-fit chi-square statistics for the conditional model given the random effects ("Littell et al., 1996). In addition, nested models can be compared in terms of their deviance or scaled deviance to compute an improvement-in-fit chi-square with df equal to the difference in the number of random effects in the two models. McCullagh and Nelder (1989, pp. 33-36) warn that testing procedures based on the deviance statistics are not well studied in the context of the Generalized Linear Model and should be interpreted cautiously. They are even less well studied in the context of the Generalized Linear Mixed Model, so analysts should treat them accordingly."

The GLIMMIX procedure has never produced an estimate of the model deviance. I don't know if there has been research between the time that Murray wrote the above paragraph and the time that the GLIMMIX procedure was introduced that would indicate that the deviance returned for the GLIMMIX macro is not valid for testing nested models. I haven't seen such research. I would dearly like to know why the deviance term reported in the macro is not reported in the GLIMMIX procedure.

Dale

--------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------

--- On Wed, 11/19/08, Tobias Jeppsson <tobias.jeppsson@EKOL.SLU.SE> wrote:

> From: Tobias Jeppsson <tobias.jeppsson@EKOL.SLU.SE> > Subject: Deviance in glimmix > To: SAS-L@LISTSERV.UGA.EDU > Date: Wednesday, November 19, 2008, 10:04 AM > Hi > > Is it possible to compute deviance in GLIMMIX, especially > for models using > pseudo-likelihoods? I cannot find any information about it > in the help. Or > does the pseudo-likelihood prohibit use of deviance to > evaluate model fit, > as it does using ie AIC to compare models? > > thanks > Tobias