Date of Award

6-8-2012

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Criminology

First Advisor

Erika Frenzel, Ph.D.

Second Advisor

Jennifer J. Roberts, Ph.D.

Third Advisor

Jennifer L. Gossett, Ph.D.

Fourth Advisor

Bitna Kim, Ph.D.

Abstract

The purpose of this research was to determine if the decision tree analytical technique offered improvement in predicting intimate partner violence outcomes. Two dependent variables were used to examine the research question. The first variable represents a dichotomy; reassault versus no reassault. The second variable included five categories; no reassault, controlling behaviors only/ threatening reassault, one-time reassault, and repeat reassault. Logistic regression and decision trees were used to assess the outcome and were compared to one another for predictive accuracy. For logistic regression, there were two models; dichotomous logistic regression and multinomial logistic regression. For the decision tree models there were four algorithm-based models; CHAID, Exhaustive CHAID, CART, and QUEST. The models were ranked on AUC, overall classification, sensitivity for the target category, and selection of the splitting variable. The results suggested that decision trees offer an improvement over logistic regression in the prediction of intimate partner violence reassault and repeat reassault. The CART algorithm was found to be most effective in predicting the outcomes associated with intimate partner violence. The decision tree models selected "controlling behaviors" as the most influential variable in predicting intimate partner violence.

Share

COinS