ABSTRACT
Generalized linear models are fit using maximum likelihood estimation, and implemented in software using an iterative algorithm known as iteratively weighted least squares that generalizes the least squares method for classical linear models. The gamma and inverse-Gaussian families are distributions useful for modeling a continuous and positive response variable with no upper bound. A simpler model can be constructed using color as a numeric variable, and either width or weight to represent female size. In addition to overdispersion, many sets of empirical data exhibit a greater prevalence of zero counts than can be accommodated by the Poisson or negative-binomial models. Zero-inflated models, introduced by Lambert as the zero-inflated Poisson model, provide an attractive solution to the problem of dealing with an overabundance of zero counts. The emphasis shifts here from fitting and comparing models with different distributional forms and link functions to selecting terms for an adequate descriptive and explanatory model.
