Improving comparisons between models for CPUE

Abstract

In a recent article published in this journal, Shono (2008) discuss an issue of great relevance in fisheries: how to model catch per unit effort (CPUE). CPUE data are non-negative and severely skewed right, suggesting gamma or log-Normal models. However, often no fish are caught, which produces exact zeros in the data and neither a gamma or log-Normal model can then be used. These exact zeros cannot be ignored and contain important information. For modelling this type of CPUE data, Shono compares three different approaches commonly used in the literature with the relatively new approach of using a Tweedie generalized linear model (GLM). The four models compared all lie in the GLM methodology (McCullagh and Nelder, 1989): (i) model CPUE using an ‘ad hoc’ method in which the logarithm of CPUE is modelled using standard linear regression after first adding an (arbitrary) small constant to all CPUE values; (ii) model catch using a Poisson or negative binomial GLM; (iii) model CPUE using a two-step model, where a binomial (logit) GLM is first used to estimate the zero-catch, and then a log-Normal regression model is used to model the positive catch; (iv) model CPUE using a Tweedie GLM. These models are defined and well-explained in Shono, so will not be discussed further here.