Variance explained by random factor
On Thu, Aug 14, 2008 at 4:43 AM, Douglas Bates <bates at stat.wisc.edu> wrote:
On Thu, Aug 14, 2008 at 11:24 AM, Renwick, A. R. <a.renwick at abdn.ac.uk> wrote:
Many apologise but the glm model I compared was ma not ma1 and thus did have the interaction term:
ma<-glm(RoundedOverlap~sess+breedfem+sess:breedfem ,family=poisson,data=Male) mixed<-lmer(RoundedOverlap~sess+breedfem+sess:breedfem+(1|Site),family=poisson,data=Male)
In that case it could be that the deviance or log-likelihood is not being evaluated correctly in glmer. Look at the slot named 'deviance' in the lmer fit. It should be a named numeric vector. The names of interest are 'disc', the discrepancy for the generalized linear models (this is the deviance without the compensation for the null deviance), 'ldL2', the logarithm of the square of the determinant of the Cholesky factor of a second-order term, and usqr, the squared length of the transformed random effects. For a mixed-effects model in which the variance of the random effects is estimated as zero, both 'ldL2' and 'usqr' should be zero. You can check these values in mixed at deviance
But isn't possible that the log-likelihoods are incompatible between glm and glmer? Maybe one chose to drop an irrelevant constant and the other didn't. lmer currently doesn't let you specify a model without a random effect, but I think this gets pretty close: counts <- c(18,17,15,20,10,20,25,13,12) outcome <- gl(3,1,9) fixed <- rep(1, 9) d.AD <- data.frame(treatment, outcome, counts) m_glm <- glm(counts ~ outcome, family = poisson) m_lmer <- glmer(counts ~ outcome + (1 | fixed), family = poisson)
logLik(m_glm)
'log Lik.' -23.38066 (df=3)
logLik(m_lmer)
'log Lik.' -2.564571 (df=4) Although they do have the same deviance:
deviance(m_lmer)
ML 5.129141
deviance(m_glm)
[1] 5.129141 Hadley