Correlation, base-rates, and the predictability of behaviour

Abstract

Two properties of data are examined that affect correlations with dichotomous measures of behaviour. These properties: (1) effects of coarse grouping and (2) skewness in distributions or base-rates, have important implications for interpretations of low or moderate “person” correlations and lack of cross-situational consistency in behaviour. Evidence is presented that indicates in many cases these properties seriously restrict the maximum correlation that can be possibly obtained between any variable and behaviour dichotomies. It is also shown that the manifest value of the correlation coefficient can underestimate the latent relationship between variables under these conditions. These properties have implications for aggregations of behaviour as well. The limitations imposed by extreme base-rates on predicting infrequent behaviours are also discussed. Alternative models of prediction are examined and suggestions for future research are provided that may be helpful in resolving some of the epistemological difficulties presented by dichotomous variables.