Marte Lilleeng <mlilleeng <at> gmail.com> writes: > > Hi, I am new to this list.
The r-sig-mixed-mod...@r-project.org mailing list would be more appropriate for this question -- please direct any further questions there ... > I have a question regarding including both spatial and temporal random > factors in lmer. These two are not nested, and an example of model I > try to fit is > > model1<-lmer(Richness~Y+Canopy+Veg_cm+Treatment+(1|Site/Block/Plot)+ > (1|Year), > family=poisson, REML=FALSE), > where > richness = integer > Y & Treatment = factor > Canopy & Veg_cm = numerical, continous > Site/Block/Plot= factor > Year = integer Fine, but REML=FALSE is unnecessary/irrelevant for generalized linear mixed model (family!="gaussian") fits. > > I get the following warning message: > > Warning messages: > 1: In mer_finalize(ans) : > Cholmod warning 'not positive definite' at > file:../Cholesky/t_cholmod_rowfac.c, line 432 > 2: In mer_finalize(ans) : singular convergence (7) > > Is this due to the nature of my fixed/random factors or the way I put > up the random factors? Hard to tell exactly. It's probably due to overfitting and/or lack of balance (glmer handles lack of balance, but extreme lack of balance can lead to technical difficulties like this one). > In lme I could include a component for autocorrelation, > ex:cor=corAR1(form=~Year|Site/Block/ID). Does the equivalent exist for > lmer? No, sorry. Crossed random effects are possible in lme (see p. 165?) of Pinheiro and Bates 2000, and glmmPQL in the MASS package can handle a Poisson response, so that might be the best way to go. However, I would also strongly encourage you to do some graphical exploration of your data and make sure there aren't outliers, almost-empty blocks, etc. Ben Bolker ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
