Choosing between parsimonious and predictive ecological models

Background/Question/Methods Ecologists often use mathematical models to either help them identify and understand ecological mechanisms, or predict ecological patterns. For the former case a model may be considered useful if it is parsimonious (i.e. it trades off model complexity with predictive ability). However, when the objective is ecological prediction one may be less interested in model simplicity. In this talk I present results from simple ecological examples that illustrate the trade-off between model simplicity and model prediction. In particular, I consider a definition of model parsimony that is consistent with Akaike’s Information Criterion (AIC), which is widely adopted by ecologists when selecting models.

Results/Conclusions Using simulated data sets I show that when AIC is used correctly it can reliably identify the relatively simple models that incorporate the most important ecological factors or processes. However, often the models selected during an AIC analysis perform worse at predicting new data when compared with some of the more relatively complex models not selected. These results highlight the importance of identifying the objective of model selection when choosing a model selection approach.