I am trying to analyze data from an experiment in which I measured the learning of a colour preference in birds under two treatments. 40 Individuals were organized into 8 groups, and 4 groups were assigned to each treatment (i.e. individuals experienced only one treatment). I ran 70 trials which gave me 14 repeated observations (trials) on each individual (individuals were not measured in every trial but were focals once in every 5 trials).

I’m using glmer with a binomial distribution, where the response variable is the proportion of visits to feeders of the correct colour. I include Treatment and z-corrected trial number as fixed effects and group and individual as random effects, i.e.: GLMMHA12z <- glmer(cbind(Nb.correct.vis, Nb.vis.total) ~ Treat + TrialZ + (1|Group) + (1|Ind), data = d, family = binomial)

A histogram of the proportion looks like this: histogram of the proportion correct

which would suggest over dispersion, but when we look for over dispersion there seems to be under dispersion:


chisq ratio rdf p

155.2787661 0.2870217 541.0000000 1.0000000

So, my questions are: 1) Could the fact than we have under dispersion when we might have expected over dispersion be due to the weight given to the observations? Many of the 0 and 1 proportions are cases where the bird only made 1 visit during the trial(82% when proportion = 0 and 72% when proportion = 1), so they should receive a lower weight in the model (if I understand correctly).

2) How should I deal with the under dispersion? I have tried to add an observation level random effect, but it doesn’t change anything:

d$obs<-as.factor(1:dim(d)1) GLMMHA12zObs <- glmer(cbind(Nb.correct.vis, Nb.vis.total) ~ Treat + TrialZ + (1|Group) + (1|Ind)+(1|obs), data = d, family = binomial)

overdisp_fun(GLMMHA12zObs) chisq ratio rdf p 155.2787657 0.2875533 540.0000000 1.0000000

anova(GLMMHA12z,GLMMHA12zObs) Data: TestGroupeEssai Models: GLMMHA12z: cbind(Nb.correct.vis, Nb.vis.total) ~ Treat + TrialZ + (1 | Group) + (1 | Ind) GLMMHA12zObs: cbind(Nb.correct.vis, Nb.vis.total) ~ Treat + TrialZ + (1 | Group) + (1 | Ind) + (1 | obs) Df AIC BIC logLik deviance Chisq Chi Df Pr(>Chisq) GLMMHA12z 5 985.03 1006.5 -487.51 975.03
GLMMHA12zObs 6 987.03 1012.9 -487.51 975.03 0 1 1


This is by all means not a direct answer, I am new to GLMM and have similar problems to yourself. However according to http://ase.tufts.edu/gsc/gradresources/guidetomixedmodelsinr/mixed%20model%20guide.html which mention the overdisp_fun function overdispersion is indicated by a p-value smaller than 0.05. So according to your overdisp_fun output p-value your data is not overdispersed.

I have the opposite, a ratio which (if I am interpreting it correctly) should mean that the model is ok but a p-value smaller than 0.05 indicating overdispersion.

overdisp_fun(M2) chisq ratio rdf p 4.510896e+04 1.016242e+00 4.438800e+04 7.998600e-03

That said I am not sure if this overdispersion test is the correct way to go about checking binomial glmm models. If you got any clearer idea after posting this question, I would appreciate knowing what you found out.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.