Use of markov-encoded sequential information in numerical signal detection

Abstract

Twelve Ss made binary decisions with feedback on numbers from one of two normal distributions with equal variances and unequal means. Sequences of distribution choices corresponded to first-order two-state Markov processes with probabilities of change of state of p1 = p2 = .50, p1 = p2 = .75, and p1 = p2 = .25. Performance was best (in d’ terms) when p1 = p2 ≠ .50. First-order sequential response dependencies tended to mirror the first-order stimulus dependencies. Violations of a fixed cutoff point decision rule were concentrated in the region of the average critical point, with a bandwidth of about 1/2σ, in which violations were strikingly more frequent than would be expected if they had occurred randomly. These results imply that in this task Ss are using a criterion-band decision rule instead of a fixed cutoff point rule, and that they are basing decisions in the region of the criterion band on information extracted from the sequence of decisions presented to them. The average bandwidth is generally different from the optimum bandwidth used by an ideal O in combining the two sources of information.