Signal Detection Theory

Abstract

In perceptual, cognitive, and diagnostic tasks, accuracy depends on both limited sensitivity and the application of a decision process. By making explicit assumptions about the nature of the cognitive representation, signal detection theory (SDT) measures the contributions of these two components of performance. (Other approaches to assessing accuracy also constrain possible representations, although not all are explicit about their assumptions.) The receiver‐operating characteristic (ROC) curve, which takes account of response bias in decision‐making by examining conditions in which this bias varies, is a valuable tool for sensitivity measurement: The ROC curve provides an assumption‐free index of accuracy, and is also used to distinguish among competing representations. The simplest form of SDT assumes that different stimulus possibilities lead to distributions of strength on a decision axis, and this model can be applied to many experimental paradigms in which a single stimulus is presented on each trial. Using more complex representations in which distributions lie in a multidimensional space, SDT predicts the relation between different paradigms for measuring accuracy (e.g., two‐alternative forced‐choice, same‐different). Multidimensional representations also describe the classification and identification of stimuli from complex stimulus sets, and are particularly useful in analyses of perceptual interaction and attention.

Reaction time signatures of discriminative processes: Differential effects of stimulus similarity and incentive

“…The straight functions from the red stimulus correspond to the ROC curves expected from high threshold theory. See Macmillan, 2002, and Blough, 1967, 2001, for further explanation and discussion. )…”

“…In the signal detection theory (SDT) implementation of this hypothesis, sensory information is represented by an internal distribution of input values, preexisting information is represented by a criterion, and percentage of responses is determined by the proportion of the sensory distribution that falls beyond the criterion (e.g., Blough, 2001;Macmillan, 2002). This SDT scheme handles the effects of stimulus similarity on the percentage of responses observed in these and related experiments quite well, as I have discussed at length elsewhere (see Blough, 1978Blough, , 2001.…”

“…The most familiar formalization of this dichotomy is signal detection theory, which provides a recipe for separating and measuring the effects of these variables on response emission and choice (e.g., Green & Swets, 1966;Macmillan, 2002). Originally applied to human performance, detection theory soon influenced research with animals (e.g., Blough, 1967;Nevin, 1965).…”

Kindergartners’ fluent processing of symbolic numerical magnitude is predicted by their cardinal knowledge and implicit understanding of arithmetic 2 years earlier

“…Signal detection methods, i.e. d-prime (z scores for hits – z scores for false alarms) (MacMillan, 2002), were used to assess task fluency (Geary et al, 2007; 2009). Cronbach’s alphas range from .70 to .90 for hits, misses, correct rejections, and false alarms (Geary et al, 2007).…”

Using fuzzy signal detection theory to determine why experienced and trained drivers respond faster than novices in a hazard perception test

“…The traditional method for determining which of the two proposed models of hazard perception is true would be to conduct a signal detection analysis (Green and Swets, 1966;Macmillan, 2002;Macmillan and Creelman, 1991) in which measures of correct hits (when a participant correctly identifies a hazard) and false positives (when a participant incorrectly identifies a nonhazard as being a hazard) are entered into formulas to determine separate measures of sensitivity and response bias (Stanislaw and Todorov, 1999). However, there are a number of reasons why this approach is both conceptually inappropriate and practically difficult for hazard perception-like tasks.…”