I am MSc student in the process of analysing data from project.
I carried out data collection in the coffee triangle of Colombia, looking at a pest of coffee, the coffee berry borer. I studied 6 different farms in three different altitudinal bands.
The data i collected is count data of coffee berries, infested berries, and counts of coffee berry borer. (amongst other things but those are the response variables).
I want to see if the responses differ among altitudes but i want to use farm as a random effect. The data is over dispersed so i must account for that with a quasi poisson/binomial but lme4 package in R does not allow it with a mixed model (GLMM).
i have found a way around the problem but i still do not understand you can't account for over dispersion with quasi family in a mixed model.
I have googled the answer but i can not find really satisfactory answer.
Thanks
Lawrence
There's an answer to the question why quasi-families were removed from lme4, and what alternatives you have here:
https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#fitting-models-with-overdispersion
I think in your case, you can use the glmmTMB package with nbinom1 / nbinom2 family to address overdispersion.
I think 6 farms is too few to use in estimating a random effects model ( you have a sample size of six to estimate the variance) - I would treat as fixed effect and therefore include a dummy identifying each farm.
" Treating factors with small numbers of levels as random will in the best case lead to very small and/or imprecise estimates of random effects; in the worst case it will lead to various numerical difficulties such as lack of convergence, zero variance estimates, etc. "
https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#should-i-treat-factor-xxx-as-fixed-or-random
First of all, it would be better to check which family distribution is more suitable for your data/error distribution. Poisson, Quasi-Poisson, and Negative Binomial all are typical family distributions for count data. To your question, I see people use “glmer.nb” function with the Quasi-Poisson family. However, for this function, there is no inference available for the dispersion parameter theta. Therefore, I would try “glmer” with the “negative.binomial” family from the “MASS” package and specify the overdispersion parameter. E.g., “glmer(...,family=MASS::negative.binomial(theta=1.75))” (taken from "lme4" manual)
Lawrence Whittaker I know this thread is from a few years ago, but if possible how did you find your way around this problem? I have the same situation! thanks!
Hi Riona McNamara !
I just had the same problem. Adding an observation-level random effect apparently works fine. I was modelling binomial data and this reference helped a lot:
Harrison, X. A. (2015). A comparison of observation-level random effect and Beta-Binomial models for modelling overdispersion in Binomial data in ecology & evolution. PeerJ, 3, e1114.
Cheers!
- Badges
- Science topic
- Similar topics
- Correction