Recidivism is a definition, not a number

Ask three agencies for their recidivism rate and you will get three numbers, all defensible, none comparable. The variation is rarely dishonesty. It is definition: what counts as the event, how long the window runs, who is in the cohort, what happens to time spent unavailable to reoffend, and which denominator is used.

Six choices that decide the answer

• Event — rearrest, reconviction or reincarceration. Rearrest measures police activity as much as behaviour.
• Window — 6, 12, 24 or 36 months. Curves separate late; a 12-month figure hides the period where aftercare matters most.
• Cohort — release date, sentence type, facility, programme participation.
• Opportunity — time actually in the community, adjusted for death, migration and other custody.
• Denominator — everyone eligible, everyone who started, everyone who completed, everyone released. Completer-only denominators flatter programmes.
• Fairness — access differences, follow-up differences and subgroup uncertainty.

XReform requires a registered definition manifest before a cohort figure can be published, and prints that manifest beside every result. It is the cheapest integrity control in analytics.

Static factors are context. Dynamic factors are where services act.

Age at first offence, prior custodial terms and offence type are stable predictors — and nothing a prison can change. Employment at 90 days, substance-use relapse, housing stability, antisocial peer contact, programme dosage met and family support are the factors that move, and they are the ones worth building services around. A dashboard that ranks people on static history is a dashboard that has given up.

Measure the chain before claiming the outcome

Assessed, plan agreed, programme access, sufficient dosage, completed, release linkage. If access is 48% and dosage fidelity is 79%, an outcome claim is a claim about the fraction of people who got the full intervention — say so. US correctional-education evidence from the RAND meta-analysis associates participation with substantially lower odds of recidivism; that is a population result about a service, not a prediction about a person.

Publish the complete definition manifest beside every result. Two honest dashboards can disagree — the manifest is how a reader can tell why.

The line we do not cross

Recidivism analytics in XReform are population instruments. Using a cohort model to make an adverse decision about an individual — a placement, a privilege, a release recommendation — is a red-lane function, blocked by policy gate rather than discouraged by guidance. Small cells are suppressed, subgroup confidence intervals are shown, and research extracts run through a governed workspace with privacy controls and reproducibility requirements.