Model comparisons.

Objective There is currently inconclusive evidence regarding the relationship between recidivism and mental illness. This retrospective study aimed to use rigorous machine learning methods to understand the unique predictive utility of mental illness for recidivism in a general population (i.e.; not only those with mental illness) prison sample in the United States. Method Participants were adult men (n = 322) and women (n = 72) who were recruited from three prisons in the Midwest region of the United States. Three model comparisons using Bayesian correlated t-tests were conducted to understand the incremental predictive utility of mental illness, substance use, and crime and demographic variables for recidivism prediction. Three classification statistical algorithms were considered while evaluating model configurations for the t-tests: elastic net logistic regression (GLMnet), k-nearest neighbors (KNN), and random forests (RF). Results Rates of substance use disorders were particularly high in our sample (86.29%). Mental illness variables and substance use variables did not add predictive utility for recidivism prediction over and above crime and demographic variables. Exploratory analyses comparing the crime and demographic, substance use, and mental illness feature sets to null models found that only the crime and demographics model had an increased likelihood of improving recidivism prediction accuracy. Conclusions Despite not finding a direct relationship between mental illness and recidivism, treatment of mental illness in incarcerated populations is still essential due to the high rates of mental illnesses, the legal imperative, the possibility of decreasing institutional disciplinary burden, the opportunity to increase the effectiveness of rehabilitation programs in prison, and the potential to improve meaningful outcomes beyond recidivism following release.研究目标 目前关于累犯与精神疾病之间的关联尚无定论。本回顾性研究旨在采用严谨的机器学习方法，探究精神疾病对美国普通监狱人群（即不仅包含精神疾病患者）的累犯预测的独特预测效用。 研究方法 本研究的研究对象为从美国中西部地区3所监狱招募的成年男性（n=322）与成年女性（n=72）。为明确精神疾病、物质使用、犯罪及人口统计学变量对累犯预测的增量预测效用，本研究开展了3组基于贝叶斯相关t检验的模型对比分析。在评估用于t检验的模型配置时，共考虑了3种分类统计学习算法：弹性网逻辑回归（elastic net logistic regression, GLMnet）、k近邻（k-nearest neighbors, KNN）以及随机森林（random forests, RF）。 研究结果 本研究样本中物质使用障碍的患病率尤其高（86.29%）。相较于犯罪与人口统计学变量，精神疾病变量与物质使用变量并未为累犯预测带来额外的预测效用。针对犯罪与人口统计学、物质使用、精神疾病特征集与空模型的探索性分析显示，仅犯罪与人口统计学模型提升累犯预测准确率的可能性更高。 研究结论 尽管未发现精神疾病与累犯之间存在直接关联，但鉴于服刑人群中精神疾病的高患病率、法律层面的必要性、降低监狱纪律惩戒负担的可能性、提升监狱内康复项目有效性的契机，以及改善刑释后除累犯外其他有意义结局的潜力，对服刑人群的精神疾病进行治疗仍至关重要。