New Zealand Statistical Association 2024 Conference

Peter Green

AgResearch

Adapting the linear probability model to GLMMs

Binomial GLMMs with logistic links can struggle numerically when probability estimates for some combinations of factors are zero. On the other hand using a linear link is just asking for trouble with out-of-bounds predictions becoming increasingly likely when probabilities are close to zero. By using a hybrid link equation we can accommodate boundary estimates for our fixed effects without the risk of random effects exceeding those boundaries.