Stopping the revolving door: MDP-based decision support for community corrections placement

• Main Authors: GAO, Xiaoquan, SHI, Pengyi, KONG, Nan
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2023
• Subjects: Analytics for Social Good; Non-memoryless; Superconvexity; Actor-critic Algorithm; Operations and Supply Chain Management; Social Control, Law, Crime, and Deviance
• Online Access: https://ink.library.smu.edu.sg/lkcsb_research/7667; https://ink.library.smu.edu.sg/context/lkcsb_research/article/8666/viewcontent/ssrn_4672337.pdf
• Institution: Singapore Management University

Community corrections (CC) programs offer an alternative to incarceration that can reduce jail overcrowding and recidivism rates. The aim is to address the root causes behind criminal behavior, ultimately breaking the cycle of reincarceration. However, placing all eligible individuals in CC may strain case managers, resulting in reduced supervision, increased violations, and higher recidivism rates, which undermines the intended benefits for all participants in the programs. We take the first step in building a comprehensive analytical framework based on a queueing system to support the placement decisions and related decisions such as capacity planning. We develop a Markov Decision Process (MDP) to systematically study the intricate tradeoffs among individual recidivism risks and the negative effects of overcrowded jail and CC programs. Unlike conventional queueing routing problems, our model incorporates salient features in the criminal justice setting. These include deterministic service times (sentence length) and convex costs that vary with program occupancy, which present significant analytical challenges. To first gain structural insights, we develop a new approach to establish the superconvexity of the value functions. This approach, based on marginal cost decomposition and system coupling, directly bounds the policy deviation in different systems and avoids the extreme tedium using traditional methods. The superconvexity result then provides a theoretical basis for our development of an efficient gradient-based algorithm, an integral element of our whole framework to support practical decision-making. We show the importance of our approach in breaking the cycle of recidivism through a case study using data from our community partner. Notably, the capacity planning recommendations generated by our research have been adopted by the community partner, showcasing the relevance and significance of our work for individuals involved in CC and the broader community.