Comparison of Multivariate Techniques in the Prediction of Juvenile Postparole Outcome

Abstract

Several strategies for the prediction of institutional outcome have been attempted by researchers in juvenile and parole prediction. The majority of researchers have utilized a variation of multiple regression, while a few have made use of discriminant functions and factor analysis. The present study contrasted the three approaches in terms of predictive efficiency for juvenile recidivism. A total of 579 juvenile cases were selected from the Washington state juvenile archives during the calendar year of 1971. Forty predictive variables from personal histories and behavioral records of the Office of Juvenile Rehabilitation were collated and quantified for the study. The total sample was divided randomly into two groups for purposes of double cross-validation. On each of the two sample halves, a set of predictors of juvenile outcome were generated by each of three methods, factor analysis, stepwise regression, and stepwise discriminant functions. Validation of the three methods was assessed by (1) D2 or discriminability of each predictive variable on the cross-validation sample, and (2) the accuracy of classifying each juvenile based on discriminant analyses. The results showed clearly that stepwise discriminant functions provided the more consistent selection of variables with seven of 10 variables the same for all samples. Also, discriminant functions was shown to manifest the most predictive efficiency with significantly fewer misclassifications and greater discriminability. By contrast, stepwise regression and factor analysis as strategies manifested more misclassifications and less discriminability. In addition, the two methods were found to be unreliable in the selection of predictive variables for the two validation samples. Results were discussed in terms of the shrinkage problems of regression analyses and the homogeneity of variance associated with factor analysis.