Model Averaging in Agriculture and Natural Resources: What Is It? When Is It Useful? When Is It a Distraction?

I use two examples to illustrate three methods for model averaging: using AIC weights, using BIC weights, and fully Bayesian analyses. The first example is a capture-recapture study that estimates the population size by averaging over 4 models for capture probabilities. The second is an analysis of a study of logging impacts on Curculionid weevils using a before-after-control-impact (BACI) study design. The estimated impact is averaged over 4 ecologically relevant models. Both examples demonstrate the sensitivity of model weights, or posterior model probabilities, to the choice of prior model probabilities and prior distributions for parameters. The model averaged estimates and their confidence intervals are less influenced by those choices. The BACI-design example also demonstrates the need to carefully choose the model parameterization so that the parameter of interest, the interaction, has the same interpretation for all models in the model set. I also briefly discuss three other frequentist approaches to model averaging: bagging, stacking, and model-averaged-tail-area confidence intervals