Response time distributions in memory scanning

Abstract This article reports the results of a memory scanning experiment (S. Sternberg. 1966, Science, 153 , 652-654) in which each of four subjects participated in about 1500 experimental trials per memory set size. These large samples made it possible to test a number of important nonparametric (i.e., model-free) properties of the response time (RT) distributions. These properties place severe constraints on the various memory scanning models and they provide a deeper description of the data than summary statistics or goodness-of-fit values. Five conclusions stood out. First, increasing the size of the memory set induced the strongest possible form of stochastic dominance on both target present and target absent trials. Second, the RT hazard functions were nonmonotonic, thereby falsifying a large class of serial searell models, Third, strong evidence was obtained against art exhaustive search. Fourth, some evidence was found that adding an item to the memory set inserts a stage with exponentially distributed duration into the processing claim, at least on largest absent trials. Fifth, the data supported the hypothesis that three of the subjects stored the representations of the memory set items in a visual short-term memory system and the fourth subject used all acoustic short-term system. To our knowledge, the only extant model of memory scanning that is consistent with all these results assumes that search is parallel, self-terminating, and of very limited capacity.