Prediction of psychopharmacologic effect in man: development and validation of a computerized diagnostic decision tree.

Abstract Of 336 psychiatric patients classified according to Klein's diagnostic scheme ( Klein, 1967 ), 305 were randomly assigned to double-blind treatment with chlorpromazine, imipramine or placebo. Multiple regression equations failed to predict drug response, but analysis of global improvement scores revealed specificity of drug action within diagnoses. Because of the fallibility of clinical diagnoses an inductive, computerized group assignment procedure was developed and applied to standard interview data, yielding a high degree of validity. Multiple regression and factorial prediction techniques were considered inappropriate because of the curvilinear and heteroscedastic bivariate relationships frequently obtained in ratings of psychopathology. The decision tree method is equal to discriminant function analysis in group assignment, superior in face validity of item selection, and can be applied without violation of statistical assumptions. It approximates clinical logic, and shows good, but not perfect, agreement with clinical diagnoses in detecting drug effect.