Comparing distribution models for small samples of overdispersed counts of freshwater fish

The study of species abundance often relies on repeated abundance counts whose number is limited by logistic or financial constraints. The distribution of abundance counts is generally right-skewed (i.e. with many zeros and few high values) and needs to be modelled for statistical inference. We used an extensive dataset involving about 100,000 fish individuals of 12 freshwater fish species collected in electrofishing points (7 m2) during 350 field surveys made in 25 stream sites, in order to compare the performance and the generality of four distribution models of counts (Poisson, negative binomial and their zero-inflated counterparts). The negative binomial distribution was the best model (Bayesian Information Criterion) for 58% of the samples (species–survey combinations) and was suitable for a variety of life histories, habitat, and sample characteristics. The performance of the models was closely related to samples’ statistics such as total abundance and variance. Finally, we illustrated the consequences of a distribution assumption by calculating confidence intervals around the mean abundance, either based on the most suitable distribution assumption or on an asymptotical, distribution-free (Student’s) method. Student’s method generally corresponded to narrower confidence intervals, especially when there were few (≤3) non-null counts in the samples.