Random utility without regularity

Classical random utility models imply a consistency property called regularity. Decision makers who satisfy regularity are at least as likely to choose an option $x$ from a set $X$ of available options as from any larger set $Y$ that contains $X$. In light of ample empirical evidence for context-dependent choice that violates regularity, some researchers have questioned the descriptive validity of all random utility models. In this article, we show that not all random utility models imply regularity. We propose a general framework for random utility models that accommodate context dependence and may violate regularity. Mathematically, like the classical models, context-dependent random utility models form convex polytopes. They yield behavioral predictions for those choice sets from which choices are made, by specifying combinations of preference rankings across two or more contexts. We discuss how context-dependent models can be less or more parsimonious than the classical models. Random utility models with or without regularity can be tested with contemporary methods of order-constrained inference.