Identification of potential recoccurring DUI-related conviction recidivists following successful completion of outpatient substance abuse treatment in the State of Florida: A comparison of two statistical models

Driving under the influence (DUI) is a serious problem in Florida; consequently, courts and driver-licensing regulators try to ensure that repeat convictions do not occur. One way that the State of Florida attempts to eliminate such recidivism is that persons convicted of a DUI receive education and assessment regarding drugs---including alcohol. If such a person is assessed as having an alcohol or other drug problem, he or she is required to complete a course of treatment. If treatment is completed successfully, the State of Florida assumes that the risk of further DUI convictions is minimized. This assumption is made without the use of any statistical predictive model. It is the intent of this dissertation to compare the accuracy of two statistical methods (logistic regression vs. discriminant analysis) in predicting group (recidivist or nonrecidivist) membership, given the same set of predetermined predictor variables. Specifically, the primary focus is to answer the question: How good are the logistic regression and discriminant function models to the original data set; which is to ask, What percentage of cases are correctly classified? A secondary question seeks to determine whether the order of prediction for stepwise discriminant analysis and stepwise logistic regression are similar in terms of the order of entry of variables. The results demonstrated that the logistic regression and discriminant function models served to correctly classify approximately the same percentage of cases. In addition, neither model detected any statistically significant contribution to outcome by the predictor variables; however, the use of the predictor variables in the logistic regression model did approach p<0.05 (p = 0.07) in terms of contribution. A large effect size was found between the two methods; however, it was due to chance only. No gender or age-related effects were found using Pearson's chi 2 statistic.