Yoshio Takane Tadashi Shibayama McGill University
Stimulus identification data have attracted considerable attention from many researchers (e.g., Ashby & Perrin, 1988; Keren & Baggen, 1981; Nosofsky, 1985b; J. E. K. Smith, 1980, 1982; Takane & Shibayama, 1986; Townsend & Ashby, 1982; Townsend & Landon, 1982). In a stimulus identification experiment one of n stimuli is randomly selected and presented on each trial, and the subject's task is to identify the stimulus. The basic data thus consists of a set fj/i (i = 1, . . . , n; j = 1, . . . , n) of frequencies of response j when stimulus i is presented, with fi = ∑jfj/i the total number of presentations of stimulus i.
Various models have been proposed for stimulus identification data (e.g., Ashby & Perrin, 1988; Keren & Baggen, 1981; Luce, 1963a; Nakatani, 1972; Shepard, 1957; Townsend, 1971). Typically, these models attempt to predict pj/i, the probability of response j when stimulus i is presented. The models are distinguished by different submodels assumed for pj/i. Two major classes of models have been proposed. (We exclude, from our account, the more recently proposed general recognition model by Ashby and Perrin, 1988, since it is treated in Chaps. 6-8, and 16. Also, see Ashby and Lee, 1991.) One class is similarity-choice models, and the other is sophisticated guessing models ( J. E. K. Smith, 1980; Townsend & Landon, 1982).
In the similarity-choice models, a model of stimulus similarity is postulated, and the strength of a response when a stimulus is presented is defined as a function of the stimulus similarity and the bias for that response. A response is assumed chosen with probability proportional to its response strength relative to other alternative responses. This class of models includes the unrestricted