Screening for Risk Among Parolees - Policy, Practice, and Method (From Prediction in Criminology, P 54-77, 1985, David P Farrington and Roger Tarling, ed. - See NCJ-99006)

• NCJ Number: 99009
• Author(s): S D Gottfredson; D M Gottfredson
• Date Published: 1985
• Length: 24 pages
• Sale Source: State University of New York Press
• Publication Type: Report (Study/Research)
• Language: English
• Country: United States of America

Annotation

Within the context of parole risk assessment, this study examined the statistical efficiency of five prediction methods: two linear additive models (least squares-multiple regression and the Burgess method), two clustering methods (predictive attribute analysis and association analysis), and a log linear approach (multidimensional contingency table analysis).

Abstract

The methods were compared with respect to the extent to which they account for predictor variable intercorrelations, linearity and/or additivity of relations, and tendency to overfit construction-sample data with concomitant shrinkage on validation. Data used for the comparisons were for 4,500 persons released from Federal prison in the years 1970-72. Given the types and level of sophistication of available data and outcome criteria, no one method for developing operationally useful statistical decisionmaking aids appears to provide a statistical advantage over any other. Thus, those techniques which are simpler and more easily understood and implemented may work as well as more complex techniques. Prediction devices constructed using the various techniques are highly intercorrelated. Finally, those variables that are most vulnerable to legal and ethical attack (e.g. social history variables) may actually add little to the statistical efficiency of recidivism prediction devices. Tabular data and 33 references are provided.