Influence of the underlying distribution and parameter estimation approach on the performance of N-mixture models for unmarked mammal data

Abstract

Abundance estimation of wildlife populations is a central task in ecology. Camera trapping is a non-invasive technique that can be used to estimate animal abundance using three classes of statistical models: mark-recapture models, distance sampling and occupancy models. N-mixture models are a type of occupancy based models that divide the study area into independent sites and estimate abundance across these sites using count data of observations. This work uses camera trap data collected from 2016 to 2022 on 10 mammal species in northern Spain, ranging in size from 300 grams to 100 kg, to evaluate differences in the performance of two N-mixture models fitted using maximum likelihood estimation (MLE), one assuming a Poisson distribution and one assuming a negative binomial distribution, and two N-mixture models assuming a Poisson distribution, one fitted using ML and one using Bayesian inference. It also relates the differences between models with sample size (number of observations per species and number of sites were a species was detected) and species characteristics (body mass and home range). The negative binomial model differed in abundance estimates, their precision, and bias compared to the two models assuming a Poisson distribution, whereas these latter models were relatively similar in their performance across most species. Neither sample size nor species characteristics appeared to be strong predictors of performance difference between models. These results suggest that while negative binomial models may capture overdispersion better than simpler Poisson models, the latter may offer more robust estimates even in mildly over-dispersed datasets. It highlighted that the selection of an appropriate distribution may be more influential than the choice of parameter estimation for this particular type of model.