[R-sig-ME] Strange predictions from binomial glmm - multi average approach

Quentin Schorpp quentin.schorpp at thuenen.de
Fri Apr 22 19:44:45 CEST 2016


I am analysing the dominance of a Species, i.e. its relative abundance
(Proportion) in a community. Since these data are proportions I use
binomial models. However, the predictions from these models are
surprisingly non-sense. Can somebody help me finding the mistake?

Background: I sampled a Chronosequence of agicultural fields. Sampling
took place in two consecutive years and 12 of the fields have been sampled
repeatedly. However, the reference fields had a different plant species
and due to the field management, reference fields have been switched
between the years. Therfore 6 of the fields have been sampled only once.
Due to the repeated measures design I use mixed models with the field.ID
as random factor (although refernce-field.IDs appear only once, all others
appear twice). The age of the fields was not very different (2-3 years
difference). Since I took samples in intervals of 1 year, I want to
consider both time dependent factors: Year of sampling and Age of the
field (class of the chronosequence). Therefore I calculated time since
establishment (or: cpl - cultivation period length) of the field for all
samples and regress relative abundance against cpl. Besides time dependent
factor I use environmental measures as (nuisance-) covariables.

The strange thing about the predicitions is, that the prediciton lines
reaches values, that ahsve not been observed during the study, i.e. they
are far too high.

Ths is my approach:
fml.glb   <- as.formula(cbind(scs,fail) ~ cpl + I(cpl^2) + pH + mc + ats1
+ nitrogen + (1|field.ID))

con = glmerControl(optimizer="bobyqa", optCtrl=list(maxfun=2e5),
calc.derivs = TRUE, check.conv.grad="ignore")

G1 <- glmer(fml.glb, df1, family=binomial(link="logit"),control=con)

GD1 <- dredge(G1, m.lim = c(NA,3), subset = dc(cpl,I(cpl^2)))

delta4 <- get.models(GD1, subset = delta < 4, REML=T)
M.avg <- model.avg(GD1, subset= delta < 4, fit=TRUE)

pred.se <- predict(M.avg, type="response", se.fit=TRUE, re.form=NA,
full=T, newdata)
newdata$fit <- pred.se$fit
newdata$SE <- pred.se$se

I also posted this Question ión stack overflow, where i provide a dataset
, addititional explanations and R-code and the prediction plots as well as
the results table for Averaged Coefficients:

Any Suggestions or ideas why the predicitions are out of the range of
observed proportions?

More information about the R-sig-mixed-models mailing list