A recidivism rate is the share of people who commit a new offense after contact with the justice system, measured over a fixed follow-up period — and the single most cited American figure is Bureau of Justice Statistics' finding that 68 percent of prisoners released in 2005 across 30 states were arrested within three years, rising to 83 percent within nine. The number is rigorous and often quoted; the trouble starts when readers compare rates that were never measuring the same thing.
This site publishes information, not legal advice.
What does a recidivism rate actually count?
Every recidivism statistic requires three choices: what event counts as reoffending, which population is followed, and how long the follow-up runs. The event can be an arrest, a conviction, or a return to prison; each is a different threshold on the same behavior. Arrest is the loosest — it counts police action, not adjudicated guilt. Conviction requires a court finding. Return to prison can include technical supervision violations with no new offense at all, which is why states that count revocations report higher rates than states that count only new convictions.
Per BJS's published methodology, the 68-and-83 percent figures use arrest as the event, drawn from state and federal criminal history records. The same studies report far lower figures on stricter definitions — about half the cohort returned to prison within the observation windows — and both numbers describe the same people.
Why do states' published rates disagree so much?
Because each state picks its own definition. Some report rearrest within three years of release, others convictions within two years of community supervision, others any return to custody within five. Per state correctional and sentencing-commission publications, neighboring states have published recidivism rates differing by a factor of three while measuring similar populations — differences produced almost entirely by definitional choices, not by behavior.
The follow-up window matters as much as the event. Reoffending risk is front-loaded: BJS's published survival analyses show the largest share of rearrests occurring in the first one to two years, with the curve flattening but never reaching zero. A two-year follow-up and a five-year follow-up of the same release cohort can produce rates differing by 20 percentage points or more. Comparing a state's two-year conviction-based rate to the federal three-year arrest-based figure, as press coverage often does, compares apples to an orchard.
How does the measurement window change the number?
| Follow-up period | Arrested, per BJS 2005 cohort (30 states) |
|---|---|
| Within 1 year | 44 percent |
| Within 3 years | 68 percent |
| Within 5 years | 77 percent |
| Within 9 years | 83 percent |
The table shows the pattern BJS's published analyses document: roughly a quarter of all nine-year rearrests happen after the standard three-year window closes. A program evaluated at three years and a rival evaluated at one year will produce different conclusions even if both work identically.
Related stories: Why single-cause explanations of the crime drop rarely survive scrutiny · What clearance data shows about homicide cases going cold.
What does the arrest-based rate miss?
Three things, in opposite directions. It misses offenses never detected — the true reoffending rate is by definition higher than any arrest rate, and victimization surveys confirm substantial unreported crime. It overcounts conduct that ends without charges, since most arrests do not lead to conviction, and an arrest is not guilt. And it blurs severity: the BJS cohort studies show a large share of rearrests are for public-order and property offenses, while violence accounts for a minority, so one flat rate hides very different public-safety realities.
The rate also says nothing about whose risk was measured. Release cohorts differ by age, offense history, and sentence length, and age is among the strongest predictors of desistance in the criminology literature. A state releasing many young burglary cases and a state releasing older drug cases are not comparable cohorts, whatever their published rates claim.
Who gets counted out of the denominator?
Cohort composition quietly decides the headline. Most published recidivism rates follow people released from state prison, which excludes everyone diverted before conviction, everyone on probation who never entered prison, and everyone held in local jails — a much larger population whose reoffending is usually measured separately or not at all. Per BJS's published population data, probationers outnumber prisoners several times over, so a state's "recidivism rate" typically describes a minority slice of its justice caseload.
Selection works inside the cohort too. People serving long sentences exit older and rarer in their offense types; people released after short stays are younger and higher-risk on most actuarial dimensions. Comparing a state with long mandatory sentences to one with short decentralized sentences compares different age structures as much as different policies, and per the criminology literature, age alone can account for a large share of cross-state rate differences.
The event definition also moves with the law. When a state reclassifies an offense or expands diversion, the population reaching release changes and the published rate shifts even if individual behavior does not. Analysts who follow the figures longitudinally therefore watch for statute changes in the same years they watch the rates, because a legislative session can move the number as much as any program.
What would a better-informed reader ask for?
Three attachments should travel with any recidivism figure: the event definition, the follow-up window, and the comparison baseline. Per the published evaluation standards that BJS and the National Institute of Justice encourage, credible program claims also name their study design — random assignment, matched comparison, or none. A percentage quoted without those attachments is not wrong so much as uninterpretable, and most public argument about recidivism is argument about the attachments, not the arithmetic.
Can recidivism studies compare program effects fairly?
Only with the cautions above applied. Credible evaluations — including BJS-supported random-assignment studies and strict quasi-experimental designs — fix the definition, the window, and the comparison group before looking at outcomes. Per the published evaluation literature, programs that look strong under a two-year revocation-based measure can look modest under a matched three-year arrest measure, and vice versa. The honest summary of any recidivism claim names all three: the event counted, the window used, and the baseline cohort compared against.
The number itself is not broken. BJS's longitudinal studies are among the most careful in the federal statistical system. What breaks is the habit of quoting a single percentage without its definition attached — a habit this arithmetic cannot support.
For more context, read Why single-cause explanations of the crime drop rarely survive scrutiny.
For more context, read homicide clearance rate.
For more context, read clearance rate.