glmer vs glmmPQL vs glmmadmb
Thank you Ben! Indeed there is a problem and I keep coming back to this dater after having learned more about GLMM on other projects. I tried the suggested Gauss-Hermite quadrature. The higher the nAGQ number the more interactions become significant in the summary output. The car::Anova output stays the same though. I have a lot of Zero values both in one of my explanatory variables as well as in about half of my response values. Adding 1 and log transforming both of them (which is admittedly questionable in terms of elegance) reduces the importance of the interaction term but loosens up the cluster in the residual plots. During a previous attempt I had talked to Dave about using a zero inflation model, but we had concluded that this is not the way to go. Now I don't know where to go, because my residual plots are always highly clustered around the many 0 (1) values. Thank you Lena
On 04/04/2016 11:51 PM, Ben Bolker wrote:
If glmer and glmmADMB agree with each other and disagree with glmmPQL, I would generally trust the former (Laplace approximation is better than PQL, esp for binary data). However, (1) you should also try Gauss-Hermite quadrature (nAGQ>1) in glmer, and (2) the very large magnitude of your parameters makes it look like you probably have a complete-separation problem ... On Mon, Apr 4, 2016 at 10:53 PM, Magdalena Wiedermann <mwiederm at mtu.edu> wrote:
Dear List Quick question: Why is the interaction term usingglmmPQL not significant, whereas it is highly significant using glmer and glmmadmb? Thank you! Lena *_ Example:_* resp = resp<-cbind(data$Dead, data$Alive) m1<-glmer(resp~(treatm+log(Tree))^2+block+(1|plot), family=binomial, data = data) summary(m1) m2<-glmmPQL(resp~(treatm+log(Tree))^2+block, random=~1|plot, family=binomial, data=data) summary(m2) |m3<-|glmmadmb|(||resp||~||(||treatm||+||log(Tree)||)^2||+block, random=~1|plot,||||family=||"||binomial||"||, ||data=||data||)| |summary(m3) |*_Results:_*
> car::Anova(m1)
Analysis of Deviance Table (Type II Wald chisquare tests)
Response: resp
Chisq Df Pr(>Chisq)
treatm 82.6249 4 <2e-16 ***
log(Tree) 17992.6841 1 <2e-16 ***
block 3.6873 5 0.5953
treatm:log(Tree) 230.6844 4 <2e-16 ***
>car::Anova(m2)
Analysis of Deviance Table (Type II tests)
Response: resp
Chisq Df Pr(>Chisq)
treatm 114.4015 4 <2e-16 ***
log(Tree) 384.7095 1 <2e-16 ***
block 1.0899 5 0.9550
treatm:log(Tree) 2.9839 4 0.5605
> car::Anova(m3)
Analysis of Deviance Table (Type II tests)
Response: resp
Df Chisq Pr(>Chisq)
treatm 4 68.7297 4.208e-14 ***
log(Tree) 1 1846.9933 < 2.2e-16 ***
block 5 3.0464 0.6928
treatm:log(Tree) 4 117.7358 < 2.2e-16 ***
Residuals 734
p.s.: this it true for summary(m1,m2,m3) too.
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models