UMass study finds DOC's classification score predicts misconduct but nondiscretionary restrictions alter outcomes

The Special Commission on Correctional Consolidation and Collaboration heard a detailed presentation on Jan. 12 about the Massachusetts Department of Correction's objective point-based classification tool and a UMass Chan study of its predictive accuracy.

UMass Chan researchers told the commission they analyzed DOC reclassification data from 2019 through 2022, following each person from a classification board to their next board or release. The sample for the male reclassification analysis included about 7,611 classification events. Dr. Spencer Lawson, the study's analyst, said the scored custody level produced by the tool "accurately predict[ed] an individual's likelihood of engaging in institutional misconduct" and that the tool's AUC measures were comparable to industry-standard risk assessments.

But the study also found that when DOC applied its operational overrides and restrictions (the department's final, override-informed custody level), predictive accuracy declined. Lawson showed a pair of survival-curve comparisons: the scored custody level produced clear separation among minimum, medium and maximum groups; after overrides were applied the minimum and medium curves overlapped more, reducing discrimination between those two groups.

UMass separated two classes of adjustments. Discretionary overrides are case-by-case professional judgments that raise or lower custody and were rare in the sample (about 7.6% were moved higher and about 8.2% moved lower). Nondiscretionary overrides are categorical restrictions set by policy (DOC uses eight codes that automatically exclude people from minimum or medium). The researchers reported that roughly 43% of the sample had at least one nondiscretionary restriction applied; the single largest contributor was the civil-commitment restriction ("code C"), which applied to about 18% of the total sample. When UMass removed cases with nondiscretionary restrictions (and especially when removing code C cases) the predictive separation returned to levels similar to the scored-tool result.

Commission staff and DOC officials emphasized that some nondiscretionary restrictions reflect risk-of-flight and legal constraints (for example pending civil-commitment reviews or immigration detainers) and that the regulations implementing the classification process are memorialized in 103 CMR 4.20. Assistant Deputy Commissioner Abby Nelligan described the department's practice: nondiscretionary codes cannot be disregarded by commissioner designees; discretionary overrides require a detailed rationale and a six-month review.

UMass and DOC also discussed other metrics reported in the brief: violent misconduct during the follow-up window was under 5% in the male reclassification sample, while broader misconduct (nonviolent disciplinary incidents) ranged approximately 30% to 45% depending on the subgroup. Lawson emphasized the historical nature of the analysis (2019'2022) and said the team plans to replicate the same analyses on the DOC's updated classification system following recent revalidation work.

Members asked whether placement itself (the environment of a maximum versus medium institution) alters behavior and whether the analysis disentangled score-produced custody from actual placement. UMass said its primary predictive test focused on the scored custody output (the tool's score), though the researchers also ran parallel analyses using the override-informed custody level and performed subgroup checks that showed nondiscretionary restrictions drove the observed degradation of predictive accuracy.

Advocates at the hearing cautioned that high rates of nondiscretionary restrictions can amount to overclassification and harm people who otherwise would score lower; Claire Mazinton of the Health Legal Advocates Committee said the data "is consistent with overclassifying to medium" and urged the commission to consider the harms caused by keeping people at higher custody levels than their scores suggest. Commissioners asked DOC and UMass to provide additional targeted data and analysis examining the role of specific nondiscretionary codes and step-down dynamics.

The commission scheduled follow-up items (the parole board was set to appear at the next hearing on Feb. 9) and planned a series of facility visits so members could observe classification and placement practices in person. The meeting adjourned after the chair moved to close the session.

Next procedural steps noted by the commission included soliciting additional data requests from members to the DOC and UMass team and setting future agenda items to examine civil-commitment procedures, available minimum/prerelease bed capacity, and the relationship between classification scores and actual placement transitions.