Best practices for GLMMs with zero-inflation – unique challenges for repeated-measures data and fitting covariates in binary process

The GLMM working group continues its efforts to compare the implementation of zero-inflation models across SAS and R software platforms with the analysis of this data set where the counts of beneficial insects were measured weekly from sugar beet plots exposed to different insecticidal treatments following a randomized complete block design. SAS uses proc nlmixed to fit zero-inflation models with random effects. However, there are limited capabilities for the covariance structures with nlmixed to model the repeated measures data. We used the R package glmmTMB which allows for flexibility in fitting zero-inflation models with random effects and various covariance structures. Overall, the estimation of covariates in the binary process is prolonged to determine optimal model fit statistics. This must be done iteratively since methods like drop1() function only can be used to test terms in the observed counts data generating process. Furthermore, the value of assessing the impact of the covariates on the probability of observing a zero is likely limited as the zeros are often considered a nuisance in the data or not the primary focus of the analysis.