Using computers to make judgments: Correlations among predictors and the comparison of linear and configural rules

摘要

Statistical prediction is an important method for predicting and describing human behavior. Though linear rules are generally recommended for prediction tasks, configural rules can do well. Their success seems to be related, in part, to whether correlations among predictors are negative. One may wonder how frequently predictors are negatively correlated in real-life settings and whether the addition of interaction terms leads to a meaningful improvement in prediction in such situations. This article addresses the above questions in the context of the diagnosis of schizophrenia. Symptom ratings served as predictors, diagnoses served as criterion scores. Negative correlations were found among the predictors. A configural model made more accurate estimates of the likelihood of schizophrenia than did unit weight and differential weight linear models.