Sensitivity and Bias in Covariation Detection: A Direct Approach to a Tangled Issue,

Abstract

Signal detection theory was used to examine the effects of sensitivity and bias in covariation detection. On each trial, participants judged whether a sample of paired data was drawn from a correlated or an uncorrelated population. Average sensitivity was suboptimal compared to an ideal observer, and performance was at chance levels for population correlations less than ρ = .30. The prior likelihood of encountering a sample drawn from a correlated population was also manipulated, resulting in higher proportions of false positive and false negative errors than were expected on the basis of a Bayesian classification rule. These biases did not hinder sensitivity, in contradiction to Alloy and Tabachnik's (1984) theory of covariation detection, nor did they enhance sensitivity, in contrast to Wright and Murphy's (1984) finding that biases can facilitate the detection of covariance. Moreover, although these biases were somewhat more extreme than Bayes' theorem would predict, there was a tendency for observers to shift their decision criteria optimally as a function of the degree of the signal-plus-noise population correlation.