1

I want to analyze the effect of 2 treatments on the variation of the abundance of a plant species along a time gradient.

The experimental design consist of exclosures (treatment = no moose), paired with control plots (treatment = moose). A site consist of 1 exclosure + 1 control plot. There are 15 sites (so 15 exclosures + 15 paired plots = 30 experimental units). Each experimental units has is unique "id". The abundance (proportion : continuous value between 0 and 1) of fir has been monitored for 5 years (once every year = repeated measures). So the repeated measures is nested inside the "id".

I want to analyze the effect of treatments on fir abundance while controlling for the paired structure of the data and the repeated measures.

Since the response data is a proportion (abundance between 0 and 1), I have been directed toward GLMM with family=binomial.

However, I am not entirely sure on how to code my model. I tried :

fit = glmer(fir ~ treatment*time + (time|id) + (1|site), data=dat1, family=binomial)

If I get this right, "(1|site)" would control for the paired structure, and "(time|id)" would control for the repeated structure (since "time" is nested into "id").

but I receive the following errors :

Warning messages:
1: In eval(family$initialize, rho) :
non-integer #successes in a binomial glm!
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
unable to evaluate scaled gradient
3: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Hessian is numerically singular: parameters are not uniquely determined

Any idea what goes wrong with my model and what these error messages mean ?

Thx

Laurent.DV
  • 41
  • 5
  • It would be easier to help with a [reproducible example](https://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example). It sounds like it might not be possible to estimate the parameters of the model you desire with the data you have. – MrFlick Aug 21 '17 at 19:06
  • would you not code the nested effects like "(1|time/id)"? – yenats Jan 24 '18 at 13:55

1 Answers1

1

I have learned a way to include both paired structure and repeated measures.

But first, I learned that you cannot directly use proportions in a glmer, since you loose information about the "solidity" of that proportion. For example, if you have fir in 1 out of 2 plots, you obtain 50% of abundance, but that 50% is less "trustable" then having 60 plots with fir out of 120 plots (which also give 50% abundance). Additionally, it is not possible to insert a Correlation parameters in glmer. In our experiment, 4200 was the effort of sampling :

dat1$fir2 = round(dat1$fir*4200);
dat1$N = 4200

To include both paired structure and repeated measures, you can use Generalized Linear Mixed Models using Penalized Quasi-Likelihood (or glmmPQL) from library(MASS). The repeated measures are included via a correlation parameter with corCAR1().

The model for the above situation would be:

fit2 = glmmPQL(cbind(fir2, N-fir2) ~ treatment*time,
     random=~1|site/treatment, data=dat1, family=binomial,
     correlation=corCAR1(form = ~time|site/treatment))

Cheers !

Laurent.DV
  • 41
  • 5
  • This is excellent - solved 5 hours of searching. The only question I have is that the corCAR1 doesn't seem to work if the design is unbalanced (if a combination is missing). – Grubbmeister Sep 26 '18 at 05:00