This paper discusses the role of theoretical notions in making predictions and evaluating statistical models. The core idea of the paper is that such theoretical notions can be spelt out in terms of priors over statistical models, and that such priors can themselves be assigned probabilities. The discussion substantiates the claim that the use of theoretical notions may offer specific empirical advantages. Moreover, I argue that this use of theoretical notions explicates a particular kind of abductive inference. The paper thus contributes to the discussion over Bayesian models of abductive inference. (C) 2013 Published by Elsevier B.V.