Bayesian inference in ecology: informative, non-informative, or dis-informative.
Lele, Subhash*,1, 1 University of Alberta, Edmonton, Alberta, Canada
ABSTRACT- Ecological models are complex but data that can be used to parametrize and validate them are difficult to obtain. Bayesian approaches are becoming popular in ecology because of: 1) Possibility of augmenting observed data by expert opinion, 2) Computational simplicity of statistical inference with the advent of MCMC methods, and 3) Intuitive appeal. The purpose of this paper is to discuss and inform ecologists about the caveats and subtleties underlying the subjective and objective Bayesian approaches. In particular, I will discuss 1) Different definitions of non-informative priors and their effect on statistical inference, 2) Various difficulties associated with the convergence of the Markov chain used in MCMC and their effect on statistical inference, 3) Discuss the concept of identifiability of parameters and potential for dis-informativeness of the MCMC based Bayesian inference, 4) Proper interpretation of credible intervals, and 5) Potential problems with elicitation of priors. These issues will be discussed using ecological examples, simulations and not simply by presenting abstract mathematical arguments.
Key words: Markov Chain Monte Carlo, Prior, Hierarchical models
All materials copyright The Ecological Society of America (ESA), and may not be used without written permission.