Bayesian Model Selection for a Class of Spatially-Explicit Capture Recapture Models

TL;DR

This study evaluates various Bayesian model selection tools for spatially-explicit capture recapture models in ecology, highlighting their strengths and limitations, and providing practical recommendations for their use based on data information content.

Contribution

The paper systematically assesses 25 Bayesian model selection methods for SECR models, offering guidance on their application in ecological research.

Findings

Bayes Factor performs well for parameter estimation with low information content.

Posterior predictive loss is recommended for model selection when data are limited.

No single tool is best for both model selection and parameter estimation in low-information scenarios.

Abstract

A vast amount of ecological knowledge generated recently has hinged upon the ability of model selection methods to discriminate among various ecological hypotheses. The last decade has seen the rise of Bayesian hierarchical models in ecology. Consequently, popular tools, such as the AIC, become largely inapplicable and other tools are not universally applicable. We focus on a class of competing Bayesian spatially explicit capture recapture (SECR) models and first apply some of the recommended Bayesian model selection tools: (1) Bayes Factor - using (a) Gelfand-Dey (b) harmonic mean methods, (2) DIC, (3) WAIC and (4) the posterior predictive loss function. In all, we evaluate 25 variants of model selection tools in our study. We evaluate these model selection tools from the standpoint of model selection and parameter estimation by contrasting the choice recommended by a tool with a `true' model. In all, we generate 120 simulated data sets using the true model and assess the frequency with which the true model is selected and how well the tool estimates N (population size). We find that when information content is low, no particular tool can be recommended to help realise, simultaneously, both the goals of model selection and parameter estimation. In such scenarios, we recommend that practitioners utilise our application of Bayes Factor for parameter estimation and recommend the posterior predictive loss approach for model selection when information content is low. When both the objectives are taken together, we recommend the use of our applications of Bayes Factor for Bayesian SECR models. Our study reveals that although new model selection tools are emerging (eg: WAIC) in the applied statistics literature, an uncritical absorption of these new tools (i.e. without assessing their efficacies for the problem at hand) into ecological practice may mislead inferences.