Comparative Evidence-Based Model Choice: A Sketch of a Theory

An extensive literature on decision theory has been developed by both subjective Bayesians and Neyman-Pearson (NP) theorists, with more recent contributions to it from evidential decision theorists. The last-mentioned, however, have often been framed from a Bayesian perspective and therefore retain a subjectivist orientation. By contrast, we advance a comparative evidence-based model choice (CEMC) account of epistemic utility, which is explicitly non-subjective. On this account, competing models are assessed by the degree to which they are supported by the data and relevant background information, and evaluated comparatively in terms of their relative distances. CEMC thus provides a philosophical framework for inference that integrates the complementary epistemic goals of prediction and explanation. Our approach proceeds in two stages. First, we articulate a framework for non-subjective, non-NP-style, comparative, evidence-based model choice grounded in epistemic utility. Second, we identify statistical tools appropriate for measuring epistemic utility within this framework. We then contrast CEMC with non-comparative evidential decision-theoretic approaches, such as interval-based probability, pioneered by Henry Kyburg, which do not necessarily share the dual aims of explanation and prediction. We conclude by considering the interrelations between prediction, explanation, and model selection criteria, and by showing how these are closely connected with the central commitments of CEMC.