Submission to New Zealand Ministry of Business, Innovation and Employment 'Future Pathways' process.

Binomial and poisson models both assume that totals of events that occur independently. For the binomial, event probabilities are assumed the same for all trials. For the poisson, rates are assumed constant (e.g., per unit time). These assumptions are too often made without careful thought and checking. Examples are given of data where no explanatory variables are involved, and where they are clearly false. Alterrnatives to the binomial that are explored are, for the binomial, the betabinomial and double binomial. For the poisson, alternatives compared are the negative binomial, the poisson inverse gamma, and the zero-inflated poisson. Functions are demonstrated, mainly from the 'gamlss' package, that are available for checking.

Overdispersion, i.e., greater than binomial variation, is an issue for experimental data that examines insect mortality. I am not aware of published work that has taken account of a common pattern of change in which dispersion is highest at midrange exposures and/or mortalities. This document demonstrates the use of new and interesting abilities that have recently become available in the glmmTMB package for R. Further analyses should include careful forms of model check, and exploration of implications for the calculation of the lethal time estimates and associated confidence intervals that are of interest for data of this type.

Talk to Canberra R Meetup, Sept 2 2014