1. Empirically, parasite distributions are often best described by the negative binomial distribution; some hosts have many parasites while most have just a few. Thus identifying heterogeneities in parasite burdens using conventional parametric methods is problematical. In an attempt to conform to the assumptions of parametric analyses, parasitologists and ecologists frequently log-transform their overdispersed data prior to analysis. In this paper, we compare this method of analysis with an alternative, generalized linear modelling (GLM), approach. 2. We compare the classical linear model using log-transformed data (Model 1) with two GLMs: one with Poisson errors and an empirical scale parameter (Model 2), and one in which negative binomial errors are explicitly defined (Model 3). We use simulated datasets and empirical data from a long-term study of parasitism in Soay Sheep on St Kilda to test the efficacies of these three statistical models. 3. We conclude that Model 1 is much more likely to produce type I errors than either of the two GLMs, and that it also tends to produce more type II errors. Model 3 is only marginally more successful than Model 2, indicating that the use of an empirical scale parameter is only slightly more likely to generate errors than using an explicitly defined negative binomial distribution. Thus, while we strongly recommend the use of GLMs over conventional parametric analyses, either GLM method will serve equally well.
All Science Journal Classification (ASJC) codes
- Ecology, Evolution, Behavior and Systematics
- Generalized linear modelling
- Negative binomial