A comparison of observation-level random effect and Beta-Binomial models for modelling overdispersion in Binomial data in ecology & evolution

• DOI: 10.7717/peerj.1114
• Publication Date: 2015-07-21T08:29:45Z
• Authors: Xavier A. Harrison
• Keywords: Overdispersion; Quasi-likelihood; Binomial distribution

Overdispersion is a common feature of models biological data, but researchers often fail to model the excess variation driving overdispersion, resulting in biased parameter estimates and standard errors. Quantifying modeling overdispersion when it present therefore critical for robust inference. One means account add an observation-level random effect (OLRE) model, where each data point receives unique level that can absorb extra-parametric data. Although some studies have investigated utility OLRE Poisson count doing so Binomial proportion are scarce. Here I use simulation approach investigate ability both Beta-Binomial recover unbiased mixed effects under various degrees overdispersion. In addition, as ecologists fit intercept terms sample size low (<5 levels), performance types range sizes present. Simulation results revealed efficacy depends on process generated overdispersion; failed cope with from mixture leading slope estimates, performed well by adding noise linear predictor. Comparison those its corresponding readily identified were performing poorly due disagreement between sizes, this strategy should be employed whenever used assess their reliability. across all contexts, showed tendency underestimate modelling non-Beta-Binomial Finally, contained <5 levels term, especially estimating variance components, appeared independent total size. These suggest useful tool they do not perform circumstances take care verify robustness models.