NCJ Number
72276
Journal
Criminal Justice and Behavior Volume: 7 Issue: 3 Dated: (September 1980) Pages: 315-330
Date Published
1980
Length
16 pages
Annotation
Compared within the context of parole decisionmaking are the predictive utility of five statistical methods commonly employed to develop correctional risk-screening devices: (1) two general linear additive models, (2) two configural models, and (3) a model based on a multivariate contingency approach.
Abstract
Samples were drawn up in a manner approximating random selection from data collected previously in a study of parole decisionmaking concerning about 4,500 persons released from Federal prisons in the years 1970-1972. The 1970 release sample was selected to serve as the construction subsample, and the 1972 release sample was designated the validation subsample. With 'failure' defined as (1) any return to prison, or (2) any conviction for any new offense, or 3) death during the commission of a criminal act, or (4) absconding from parole supervision, the base rate for successful performance during the 2 years after release for the 2,382 case 1970 sample was 64 percent. For the 1,124 subjects who were released under parole supervision, the 'success' rate was 70 percent. The 689 subjects who were mandatorily released under supervision had a 'success' rate of 57 percent. Finally, 61 percent of the 569 subjects who were released at the expiration of their sentence and without supervision were 'successful.' Although 29 items appeared to possess both predictive and practical utility for developing risk-screening devices, most of the variables fell into one of four general categories related to (1) the present offense, such as the type of admission or the commitment offense; (2) a subject's history of criminal or delinquent involvements, such as age at first arrest or conviction, and the number of prior convictions; (3) 'social history,' such as the highest completed grade claimed, or living arrangements prior to commitment, and (4) institutional adjustment. Results suggest that no apparent empirical advantage accrues, given use of different models. Even when large amounts of error are randomly added to the predictor item-pool, substantive conclusions do not change. Further, most devices developed are highly intercorrelated. Implications for the practical development of screening instruments are discussed, and further research is suggested. Three text tables, one graph and 24 references are provided. (Author abstract modified)