Fast, accurate categorization is fundamental to survival

(Ashby & Maddox, 1998). Whenever we define an object as a “kind” of thing, we are categorizing. In keeping with its important role in perception and cognition, several powerful theories have been proposed and model-based instantiations developed to predict categorization performance. These include, among others, prototype (Anderson, 1991; Homa, Dunbar, & Nohre, 1991; Reed, 1972), exemplar (see Estes, 1994, and Nosofsky, 1992, for reviews), and decision bound models (e.g., Ashby, 1992a; Ashby & Maddox, 1993, 1998; Maddox, 1995; Maddox & Ashby, 1993). Nearly all these models focus exclusively on categorization accuracy as the dependent variable. Although categorization accuracy provides important information about the process of categorization, the observer’s response time (RT) often provides a richer source of information. For example, an observer might respond with the same accuracy level for two category exemplars, but with different RT distributions (e.g., Laming, 1968; Luce, 1986; Welford, 1968). To date, few rigorous theories of categorization RT exist; however, some are currently being developed, and initial tests are being conducted (e.g., Ashby & Maddox, 1994; Maddox & Ashby, 1996; Nosofsky & Palmeri, 1997). The goal of this article is to provide a rich base of categorization RT data that can be used to guide the development and testing of new and emerging models of categorization RT. Specifically, the aim is to identify a set of empirical constraints that must be predicted by any viable theory of categorization RT. A robust empirical finding is that correct-response mean RT tends to decrease as the distance between the exemplar and the category boundary increases (Bornstein & Monroe, 1980; Cartwright, 1941; see also Ashby, Boynton, & Lee, 1994). That is, exemplars that are far from the category boundary yield (on average) fast categorization responses, and exemplars near the category boundary yield (on average) slow categorization responses. Ashby and Maddox (1991, 1994) formalized this notion and called it the RT–distance hypothesis. This is an important empirical finding, but it has been tested only on a fairly weak statistic of the data—namely, mean RT. Higher order statistics, such as the RT distribution and RT hazard function, provide a richer source of information about categorization RT and thus yield more powerful empirical constraints on