Discriminant Functions Applied to the Problem of Predicting Cerebral Damage from Behavioral Tests: A Cross-Validation Study

Abstract

Linear discriminant functions were applied to 23 behavioral test scores from each of 304 patients who were divided into control, left cerebral damage, right damage, and diffuse or bilateral damage groups. Random-half and temporal-half (early versus late test period) cross-validations were performed by generating weights from the scores of each of the foregoing halves and applying them to the appropriate counter-halves. The reduction in accuracy of criterion prediction attributable to cross-validating, i.e., the “noise” in the system, varied between 10% and 20% of the uncrossed accuracy. Validated accuracy was about 80% for comparisons between controls and (a) all the damaged groups together, (b) diffuse or bilateral damage groups, (c) left damage groups, (d) right damage groups, and (e) between left and right damage groups; it was about 70% between the left damage and diffuse or bilateral damage groups; about 60–65% (essentially no better than chance) between the right damage and diffuse or bilateral damage groups. It was concluded that when greater validated accuracies of correct prediction are obtained, for situations like those of this study, they must rest at least partially upon the basis of nonlinear relationships that exist in the data. The obtained accuracies (except for lateralized damage versus non-lateralized damage groups) were considered to be high enough to indicate that the discriminant function has practical value.