|
Abstract
|
The paper suggests that the conventional wisdoms that we cannot and should not predict violence are wrong. We can predict violence, and we should predict violence. It is the unfortunate case, however, that we cannot do it very well, and this is true whether the predictions are made subjectively or statistically. Since the prediction of violence (and of other antisocial behaviors) is so pervasive in our justice and mental health systems, it is important that we attempt to do it better--that is, more efficiently and more effectively. In this paper we show that there is value to both clinical and statistical strategies toward the prediction problem, and suggest ways by which both may be improved. Attention to issues of fundamental measurement, to the base rate, to selection ratios, and to the methods of combining predictive information will be needed if the suggested improvements are to be realized. Finally, we propose that the statistician and the clinician need to pay attention to and learn from one another.
VioLit summary:
OBJECTIVE:
The purpose of this study by Gottfredson and Gottfredson was to utilize both clinical and statistical information in order to create a way to successfully and effectively predict violent behavior.
METHODOLOGY:
A non-experimental method and review of previous research was utilized. For social scientists, there existed problems of fundamental measurement. If measurement was inaccurate, then so was prediction. "Reliability" was defined as the stability with which measurements could have been made. The relative frequency of an event that occurred in the interested population referred to the "base rate." Inaccurate prediction may occur if the event was either frequent or infrequent. Although it may be common sensical to assume that infrequent events yielded inaccurate predictions, events that occur very frequently were also prone to error. Since the occurrence of violence was low to begin with, frequent events require the experience of extremely rare events, which subsequently posed a difficulty in predicting the base rate. In the past, base rates have failed to be considered when making prediction.
The proportion of individuals or events that were studied and predicted to be a part of the interested criterion classification was referred to as the "selection ratio." Thus, while the base rate provided one marginal distribution for an expectancy table, the selection ratio provided the other. Also, if one was striving for accurate prediction, the samples that were utilized in the construction of the selection devices should be representative of the population on which the device was aimed to be employed. Then, the consideration of the appropriate base rate would be ensured, and predictive power would thereafter be strengthened.
The "additive linear model" was the most frequently used actuarial predictive method. Also, the best known additive linear model was that of least-squares regression. The method involved using weighted linear combination of predictor variables that minimized the sum of squares of errors about regression to some criterion variable.
Clustering models were utilized when interactions and heterogeneities might be expected to reduce the power of multiple regression methods. Nonlinear undesignated interactions and heterogeneities that were possibly present in a population were accounted for in clustering methods.
Multidimensional contingency table analysis needed few assumptions concerning the nature of the variables under study. The method was able to estimate various weights for different predictors, was convenient to predict interaction terms, and gave a way to estimate an optimal model.
The authors contended that statistical methods of prediction were superior to clinical methods for various reasons. Firstly, information was not always utilized reliably with decision-makers. Secondly, informational items may be incorrectly weighted, which could not even turn out to be predictive. Lastly, clinical methods of prediction could be excessively influenced by causal attributions and correlations that were spurious.
Two kinds of errors were made in any predictive decision-making situation. In the context of violent prediction, some persons predicted to be violent in fact will not be (false positives), and some persons predicted not to be violent in fact will be (false negatives).
AUTHORS' RECOMMENDATIONS:
The authors suggested that there is a value to both clinical and statistical strategies toward the prediction of violent behavior, and that these strategies needed to be improved. The authors contended that improved measurement of both predictor and criterion variables was needed. Also, attention to issues of fundamental measurement, to the base rate of selection ratios, and to the methods of combining predictive information will be needed if the suggested improvements were to be realized. Finally, the authors recommended that the statistician and the clinician needed to pay attention and learn from one another. (CSPV Abstract - Copyright © 1992-2007 by the Center for the Study and Prevention of Violence, Institute of Behavioral Science, Regents of the University of Colorado)
KW - Violence Prediction
KW - Juvenile Violence
KW - Adult Violence
KW - Research Methods
KW - Adult Violence
KW - Juvenile Offender
Language: en |