Validating VAST models using simulation-based residuals

Tuesday April 7th at 2:00 PM (PST)

Andrea Havron and Cole Monnahan

Spatiotemporal delta-models are becoming increasingly popular for a wide variety of uses in ecology and fisheries science. These models rely on a hierarchical variance structure (eg., process and observation error) making standard validation measures, such as Pearson residuals, difficult to interpret. Quantile residuals provide a solution where simulations from the model are compared against observations using the cumulative distribution function (cdf) of a given distribution. This method is complicated, however, when the cdf function is not easily specified, as is this case in delta-models which consist of a mixture of zeros and positive catches. Here, we demonstrate model validation on VAST models (conditioned on EBS pollock) using the “empirical cdf” approach of Hartig (2020). We demonstrate how to validate both the process and observation components of the model, and contrast the advantages of this new approach with existing methods for model validation, and model selection tools. We conclude with practical advice for VAST users.