Construct Measurement Quality Improves Predictive Accuracy in Violence Risk Assessment: an Illustration Using the Personality Assessment Inventory
Article first published online: 23 MAY 2013
Copyright © 2013 John Wiley & Sons, Ltd.
Behavioral Sciences & the Law
Volume 31, Issue 4, pages 477–493, July/August 2013
How to Cite
Hendry, M. C., Douglas, K. S., Winter, E. A. and Edens, J. F. (2013), Construct Measurement Quality Improves Predictive Accuracy in Violence Risk Assessment: an Illustration Using the Personality Assessment Inventory. Behav. Sci. Law, 31: 477–493. doi: 10.1002/bsl.2066
- Issue published online: 4 AUG 2013
- Article first published online: 23 MAY 2013
- Manuscript Accepted: 7 DEC 2012
- Manuscript Revised: 25 OCT 2012
- Manuscript Received: 1 JUN 2012
- National Institute of Mental Health. Grant Number: 1 RO1 MH63783-01A1
Much of the risk assessment literature has focused on the predictive validity of risk assessment tools. However, these tools often comprise a list of risk factors that are themselves complex constructs, and focusing on the quality of measurement of individual risk factors may improve the predictive validity of the tools. The present study illustrates this concern using the Antisocial Features and Aggression scales of the Personality Assessment Inventory (Morey, 1991). In a sample of 1,545 prison inmates and offenders undergoing treatment for substance abuse (85% male), we evaluated (a) the factorial validity of the ANT and AGG scales, (b) the utility of original ANT and AGG scales and newly derived ANT and AGG scales for predicting antisocial outcomes (recidivism and institutional infractions), and (c) whether items with a stronger relationship to the underlying constructs (higher factor loadings) were in turn more strongly related to antisocial outcomes. Confirmatory factor analyses (CFAs) indicated that ANT and AGG items were not structured optimally in these data in terms of correspondence to the subscale structure identified in the PAI manual. Exploratory factor analyses were conducted on a random split-half of the sample to derive optimized alternative factor structures, and cross-validated in the second split-half using CFA. Four-factor models emerged for both the ANT and AGG scales, and, as predicted, the size of item factor loadings was associated with the strength with which items were associated with institutional infractions and community recidivism. This suggests that the quality by which a construct is measured is associated with its predictive strength. Implications for risk assessment are discussed. Copyright © 2013 John Wiley & Sons, Ltd.