Ecological datasets are intricate -- sparse, overdispersed, spatially correlated, measured with error. The standard approach to model selection asks which model best fits the data overall, using criteria like AIC or BIC that assess global goodness-of-fit. Gerda Claeskens, Celine Cunen, and Nils Lid Hjort argue this is the wrong question. Their focused information criterion selects the model that best estimates a specific quantity of interest, not the model that best describes the entire data-generating process. Applied to bird species abundance and Antarctic minke whale body condition, the focused approach produces different model rankings than global criteria -- and different substantive conclusions.
The structural claim is that model quality is not a property of the model alone but of the model-quantity pair. A model that best predicts mean abundance may not best predict variance in abundance. A model that best estimates the rate of whale body condition decline may not best estimate the spatial pattern of that decline. Global selection criteria treat the model as a unified object and ask whether it fits; focused selection treats the model as a tool and asks whether it works for the specific job. This reframes model selection from an ontological question (which model is true?) to an instrumental one (which model serves this inference?). The practical consequence is that researchers using a single model for all downstream analyses are implicitly assuming that global fit implies local precision -- an assumption the focused criterion shows to be false. The best model depends on what you need from it, and different needs may point to different models even within the same dataset.
(arXiv:2603.16896)