NCJ Number
214470
Journal
Journal of Quantitative Criminology Volume: 22 Issue: 1 Dated: March 2006 Pages: 31-59
Date Published
March 2006
Length
29 pages
Annotation
This article uses simulation evidence to examine model selection criteria in the field of criminological research, in the application of finite mixture models and in making a determination of the number of components to include in the mixture.
Abstract
In the majority of the three-group simulations, the Akaike Information Criterion (AIC) outperformed the Bayesian Information Criterion (BIC) in selecting the correct number of components in the mixture. However, when the components were well separated and the samples were large, BIC outperformed AIC. The findings suggest the BIC will perform best when the information in the data suggest a clear separation of the components. The overall analysis suggests that mixture models with relatively few components are required to accurately reproduce the theoretical distributions used to generate observable data. Finite mixture models have become an important analytic tool in criminological research. They have been increasingly widely used by criminologists because they provide a tractable and flexible methodology for identifying meaningful clusters of individuals and developmental trajectories of criminal and antisocial behaviors. However, a persistent point of uncertainty is the problem of choosing the number of components for the mixing distribution. Using simulated data, this paper compares the model selection performance of the widely used BIC and AIC. Figures, tables, and references