Progress and peril in model-based ecology

Peter Adler, Department of Wildland Resources and the Ecology Center, Utah State University, Logan, UT, Stephen Ellner, Department of Ecology and Evolutionary Biology, Cornell University, Ithaca, NY, Christina Hernandez, Biology Department, Cornell University, Ithaca, NY, Giles Hooker, Biological Statistics and Computational Biology, Cornell University, Ithaca, NY and Robin Snyder, Biology, Case Western Reserve University, Cleveland, OH

Background/Question/Methods
Ecologists have traditionally relied on models for inference about nature, but, until recently, most of the models were ANOVAs fit to data from simple experiments. We have very clear rules for checking ANOVAs to make sure the model fits the data, and resulting inferences are valid. But traditional experiments and ANOVAs cannot address many of the most important problems in our field, which involve processes operating over large spatial and temporal scales or responses that cannot be directly observed. Recent advances in statistics and computing now allow us to fit complex, dynamic models to empirical data. Furthermore, we are rapidly developing new methods to analyze fitted models, helping us provide empirical answers to questions that in the past were purely theoretical. However, our ability to fit and analyze complex, custom-made models has outpaced our ability--or willingness--to rigorously validate them. Understanding exactly how a model behaves is not the same thing as understanding nature. Here we use examples from our own work on species coexistence to highlight two sources of peril in learning about nature from population models fit to field data.
Results/Conclusions
The first, and most obvious, source of peril is bias in fitting model parameters. Key parameters in our coexistence models are phenomenological estimates of intra- and interspecific density-dependence. Recent theoretical work has demonstrated how measurement errors in both response variables and covariates can bias estimates of density-dependence. Results from field experiments to test model predictions indicate the presence of these biases. Fortunately, we may be able to solve this kind of problem with carefully designed experiments and validation tests. A second, more insidious, source of peril is when seemingly innocent details of model specification, such as the choice of a link function or random effects structure, pre-determines model behavior. Solving this problem is more challenging. Realizing the enormous potential of model-based ecology will require greater commitment to model checking and validation.