AIC and the challenge of complexity: A case study from ecology

• Predictive accuracy has justified simplicity in model selection.
• Model selection criteria generally entail a uniformity of nature assumption.
• Complexity threatens this assumption when predictions are extrapolated.
• This challenges simplicity's epistemic value when modeling complexity.
• In complex systems simplicity might improve explanation rather than prediction.

Philosophers and scientists alike have suggested Akaike's Information Criterion (AIC), and other similar model selection methods, show predictive accuracy justifies a preference for simplicity in model selection. This epistemic justification of simplicity is limited by an assumption of AIC which requires that the same probability distribution must generate the data used to fit the model and the data about which predictions are made. This limitation has been previously noted but appears to often go unnoticed by philosophers and scientists and has not been analyzed in relation to complexity. If predictions are about future observations, we argue that this assumption is unlikely to hold for models of complex phenomena. That in turn creates a practical limitation for simplicity's AIC-based justification because scientists modeling such phenomena are often interested in predicting the future. We support our argument with an ecological case study concerning the reintroduction of wolves into Yellowstone National Park, U.S.A. We suggest that AIC might still lend epistemic support for simplicity by leading to better explanations of complex phenomena.