Eighty-five female depressed patients were treated for 4 weeks with amitriptyline in a flexible dose (median, 150 mg daily). Predictive analyses of outcome were carried out, using both typological and regressional models based on the same 29 predictor variables.The typological prediction utilized a previously described four-group typology derived by multivariate cluster analysis and comprising psychotic depressives, anxious depressives, hostile depressives, and young depressives with personality disorder. Patients were assigned to groups by discriminant functions on the basis of scores on the predictor variables. The four groups were found to predict outcome significantly. Psychotic depressives, initially the most severely depressed, improved the most; anxious depressives, a little less severely ill initially, showed the least improvement and the most residual illness at 4 weeks; hostile depressives and young depressives with personality disorders, initially the least severely ill, showed intermediate improvement and were least ill, by a small margin, at 4 weeks.The 29 predictor variables also predicted outcome significantly in linear multiple regression analyses. Five variables showed regression weights significantly different from zero: high initial severity predicted high improvement but more residual illness; initial insomnia, more previous depressions, longer illness, and hypochondriasis all predicted low improvement and more residual illness.Predictions by using typological and regression models were compared, and their interrelationships were examined. The regression prediction accounted for more variance in outcome measures than did the typology, but the statistical significances of the regression and typological predictions were of the same order. Two variables appeared particularly important—high initial severity of illness which characterized the psychotic depresives and increased number of previous depressions which characterized the anxious depressives. Typological prediction was not solely due to these two variables in linear regression model sicne it remained significant when the outcome measure was adjusted for both by analysis of covariance.Both typological and regressional models provide useful approaches to prediction, with only limited overlap. The typological approach, which has been relatively neglected in recent years, warents more frequent application as an alternative to multiple regression.