Date of Award

Fall 10-1-2013

Document Type


Degree Name

Doctor of Science in Information Systems


Business and Information Systems

First Advisor

Amit Deokar

Second Advisor

Omar El-Gayar

Third Advisor

Viki Johnson

Fourth Advisor

Darryl Plecas


This dissertation addresses the pressing and difficult problem of assessing the risk of re-offending for parolees. The prison system in the state of California has been given a strong mandate to dramatically reduce the prison population. Before final discharge, prisoners often serve a portion of their sentence on parole release, but they are at high risk to re-offend. A number of systems have been developed to aid practitioners in parolee risk assessment, but the recommendations of these systems have not been consistently followed. Field practitioners were skeptical that recommendations adequately accounted for repeat offending histories, and did not believe that the recommendations were logical. We propose a hazard pattern based risk assessment approach to address these concerns. In this work, we demonstrate this approach using real world data, and rigorously evaluate the discovered patterns.

The design science nominal process flow was selected as the methodological framework for this undertaking. The motivating case is a business problem, in context. The search for and development of a solution is documented, including the careful evaluation of existing technologies and development of novel approaches and artifacts where necessary. An IT artifact is developed, demonstrated and evaluated within the context of the motivating case.

The driving question behind this work is this: How can we assess risk of future offending? A substantial body of work has explored this question, reflecting the importance of the question and the difficulty of finding an answer. A number of risk assessment tools have been developed but their accuracy has been moderate and their acceptance by practitioners has been lukewarm. We are thus faced with a need for a way to make accurate risk

assessments that can be justified to field practitioners.

As necessary components of a solution, two key contributions are highlighted in this work: a) hazard patterns, which extend existing work in event sequence patterns, and b) a method of selecting and presenting a relatively small number of interesting patterns that codify the rationale underlying the assessment of risk.

The solution was evaluated according to design objectives of parsimony, generalizability across data sets, meaningfulness, and predictiveness over time. We satisfied the objective of parsimony by selecting only those hazard patterns showing statistically significant differences in relative risk. To demonstrate generalizability and guard against over-fitting, ten-fold cross validation testing was performed. The selected patterns were consistent indicators of increase or decrease in arrest risk across folds in cross-validation trials. To test for meaningfulness, pattern discovery and selection was repeated with the underlying data randomly shuffled. The differences in the resulting output empirically demonstrate that the patterns were dependent on the input rather than on the pattern discovery process. Finally, to test for predictiveness over time, hazard patterns discovered in one time frame were compared to arrest outcomes in subsequent time frames. A moderate relationship between antecedent hazard patterns and future outcomes was observed, with lower accuracy near the beginning and end of criminal careers.