Intuitive Numerical Prediction

Subjects were trained, with feedback, to predict a numerical criterion from each of two separate cues and then asked, without feedback, to predict it from a pair of independent cues or a single cue.Their intuitive predictions were qualitatively inconsistent with an additive model, since the effect of one cue varied inversely with the number of cues available, and with a constant-weight averaging model, since the effect of one cue varied inversely with the validity of the other cue.The data were consistent with a relative-weight averaging model, which assumes that subjective cue values are averaged using weights that depend on cue validities.Normative and descriptive theories of intuitive prediction are compared.Early research on intuitive statistics focused on the normative accuracy of subjective statistical estimates, predictions, or decisions.This research is discussed from different points of view in review articles by Peterson and Beach (1967), Slovic and Lichtenstein ( 197 1 ) ,Hammond and Summers (1972), Bremer (1974), Anderson (1974), and Levin (1975).One issue in subjective statistics that received considerable attention is multiple-cue prediction.I n the early work, it was concluded that intuitive predictions could be well approximated by an additive model of multiple linear regression, based on what seemed to be large multiple correlations between the theory and the data.Recent research departs from the earlier work in two ways.First, since correlations of fit can be misleading (Anderson, 1972; Birnbaum, 1973 Birnbaum, , 1974b)), it is difficult to draw theoretical conclusions from much of the previous research on this topic, which used the additive model both as a theory and a statistical device.Therefore, more recent work uses improved techniques to separate model testing and measurement.Second, the new concern is with testing descriptive psychological theory rather than assessing normative accuracy, or 'performance.'Recent work has