Risk Assessment and Recidivism: Prediction, Calibration, and False Positives
Author: Ukrainian Psychological Hub · Published: October 11, 2026 · Editorial Policy
Ukrainian Psychological Hub | October 11, 2026
What recidivism risk assessment actually predicts
Recidivism risk assessment is a structured attempt to estimate the likelihood of a defined justice-system outcome during a specified future period for people who meet particular eligibility criteria. The outcome may be a new arrest, conviction, incarceration, technical supervision violation, or a narrower event such as a violent reconviction. The assessment may help inform supervision, intervention planning, resource allocation, or an authorized legal decision. Its scientific meaning depends on the outcome, population, time horizon, measurement process, and validation evidence.
A risk estimate describes uncertainty over an event, not a person's guilt, moral worth, clinical diagnosis, or inevitable behavior. An assessment that separates higher- and lower-risk groups can still misestimate absolute probabilities, misclassify many individuals, or perform differently when transferred to another jurisdiction. These distinctions are especially consequential when the result affects liberty. A major systematic review of criminal risk instruments used at sentencing found methodological limitations in much of the validation literature, particularly around calibration, fairness, and application to specific decision settings.
For a reader asking “How accurately can risk assessment predict recidivism?” the answer is therefore conditional: some instruments show useful statistical discrimination in particular populations, but their performance is bounded, varies across settings, and requires local examination of absolute risk, errors, population differences, and the decision that will follow. A validated statistical association is evidence about groups observed under defined conditions; a particular individual's future remains uncertain.
Define recidivism before interpreting any score
Recidivism is an operational justice-system outcome. The U.S. Bureau of Justice Statistics distinguishes measures involving rearrest, reconviction, and returns to incarceration. Each can produce a different reported rate from the same cohort. Rearrest may occur without a conviction; reconviction depends on charging, adjudication, and available evidence; reincarceration may follow a new offense or a supervision-related event, according to the study's definition.
The clock also matters. Does follow-up begin at arrest, sentencing, release, or admission to supervision? Is the window one, two, three, or five years? Are people who die, move, or remain incarcerated handled consistently? Do studies count the first new event or all events? Are out-of-jurisdiction records available? These choices change the denominator, opportunity for an observed event, and apparent outcome frequency.
The word “reoffending” is often used colloquially as though it means all subsequent criminal behavior. Most administrative datasets measure a recorded criminal-justice contact instead. Unreported offenses are generally absent, while policing patterns, surveillance intensity, access to counsel, local court practice, and data-sharing arrangements shape what gets recorded. A model trained on rearrest is principally estimating an observed rearrest outcome under the conditions represented in its data.
For broader evidence on the factors associated with reoffending, see Recidivism: What Predicts Reoffending?. This article addresses a narrower assessment question: when an instrument returns a number or category, what does it mean, how should its errors be measured, and what conditions justify its use?
What goes into a criminal recidivism risk instrument?
Many instruments combine information about prior justice-system contacts, age, earlier offending patterns, current circumstances, and other variables. Their items may be weighted using a fitted statistical model, a published actuarial rule, or a structured professional judgment procedure. Tools differ substantially in their definitions, scoring protocols, and purpose; a general recidivism scale, a violent recidivism model, and a sex-offense-specific instrument cannot be treated as interchangeable.
Historical or static indicators, such as the number of prior convictions, may help distinguish observed outcome rates in a validation cohort. Dynamic indicators, such as current housing instability or patterns of substance use, may change. Their measurement may help plan support, but an association between a dynamic item and later events does not show that changing the item will necessarily reduce the event. Prins and Reich's critical meta-review explains why construct validity, prediction, and claims about intervention mechanisms require separate scrutiny.
Clinical diagnoses, behavioral observations, self-report answers, arrest records, psychometric results, and professional opinions are different kinds of information. None should silently substitute for another. In particular, the presence of a psychiatric diagnosis does not establish criminal behavior or justify labeling someone dangerous. A risk instrument is also not a competency evaluation, criminal responsibility determination, or diagnostic test. Those questions use different methods and legal standards.
A clinician's structured narrative formulation may be valuable for identifying feasible management steps, while a calibrated actuarial estimate addresses an empirical probability question. Where a procedure blends them, documentation should identify what came from statistical scoring, what came from records, and what remains professional interpretation. The broader process is explained in Forensic Psychological Assessment: Purpose, Methods, and Limits.
Prediction is a claim with six necessary specifications
A meaningful statement such as “estimated two-year rearrest risk is 25%” requires a defined reference population, an outcome definition, a time window, a measurement date, a model and version, and a validation context. If any of these is missing, the number may appear more precise than the evidence allows.
The reference population specifies whom the instrument was developed for: for example, adults released from particular correctional facilities rather than juveniles, pretrial defendants, people receiving inpatient psychiatric treatment, or people with a different offense history. The outcome might be any rearrest, violent conviction, or a technical violation. The window might be two years after release. The instrument's version and scoring rules determine the predictors and thresholds. Validation establishes how that procedure performed in a relevant dataset, including how often the specified event occurred at different predicted probabilities.
A risk classification such as “high” is even more dependent on context. It may reflect a rank relative to a reference group, a cutoff selected for an operational purpose, or an absolute probability band. Without the underlying definition and observed event rate, the label cannot tell a judge, clinician, or person assessed how many comparable individuals actually experienced the outcome.
Discrimination: can the instrument rank cases?
Discrimination measures whether people who later meet the specified outcome tend to receive higher scores than those who do not. The area under the receiver operating characteristic curve, usually called AUC, is common. An AUC of 0.50 is equivalent to chance-level ranking under its standard interpretation; 1.00 is perfect ordering. An AUC of 0.70 means that, for a randomly selected outcome case and non-outcome case from the evaluated dataset, the outcome case receives the higher score approximately 70% of the time, with ties handled by the usual convention.
That is a pairwise ranking interpretation. It does not mean that a person assigned a 70 score has a 70% probability of reoffending, that 70% of high-risk individuals will reoffend, or that 70% of the instrument's predictions are correct. Nor does AUC reveal whether a published risk estimate is numerically accurate.
In a 2022 systematic review of criminal risk assessment tools used at sentencing, external-validation results in larger samples showed AUC values ranging approximately from 0.57 to 0.75. That range suggests some predictive separation, with substantial variation across tools and studies. It also underscores why there is no defensible single accuracy percentage for recidivism assessment as a whole. Different samples and outcomes produce different estimates.
Discrimination can remain similar even when an instrument systematically overpredicts or underpredicts event probabilities. Consequently, a study reporting only AUC supplies a partial answer to the question most consequential for allocating risk-related consequences.
Calibration: do predicted probabilities match observed outcomes?
Calibration concerns agreement between forecast probabilities and actual frequencies in comparable groups. If an instrument assigns 100 comparable people an estimated two-year event probability of 20%, a well-calibrated prediction would correspond to roughly 20 observed events in that group over that horizon, allowing for sampling variation. It would not identify which 20 people. Calibration is evaluated across many probability ranges, rather than by checking one person's outcome after the fact.
A calibration plot compares predicted risks with observed frequencies. Ideally, points align closely with the reference line. Calibration-in-the-large assesses systematic overprediction or underprediction. A calibration slope can reveal forecasts that are too extreme or insufficiently separated. Other methods, including observed-to-expected ratios and suitable probabilistic scoring rules, provide additional information when used and interpreted properly.
Absolute risk can drift when the outcome prevalence, supervision regime, policing environment, eligibility rules, available interventions, or record completeness changes. Even a model with respectable discrimination may need recalibration before its numeric predictions travel. In an external validation of the OxRec violent-reoffending model, researchers examined discrimination and calibration in a new English prison-release cohort and reported important differences between expected and observed risks that warranted attention to model transportability.
Calibration requires a careful denominator. Suppose a model says that 30% of an observed group will be rearrested within two years. If 18% are rearrested, the forecast may be poorly calibrated for that setting, or the follow-up and ascertainment methods may differ from those in development. The evidence must distinguish those possibilities; the quoted risk percentage should not simply be carried forward unchanged.
Sensitivity and specificity: what does the threshold do?
Risk classifications require a threshold. Sensitivity is the proportion of people who later meet the defined outcome who were classified as positive or high risk. Specificity is the proportion who do not meet the outcome who were classified negative or lower risk. An instrument can achieve different combinations of sensitivity and specificity depending on where its cutoff is placed.
Lowering a threshold usually captures more people who will meet the outcome, increasing sensitivity while often raising false positives. Raising it often increases specificity but misses more later outcome cases. No cutoff eliminates the basic trade-off when the distributions of scores overlap. The appropriate trade-off is a decision question requiring a specified objective and the costs of errors.
A false positive is someone classified as likely to meet the defined outcome who does not meet that outcome within the stated follow-up. A false negative is someone classified lower risk who does meet it. These are retrospective classification terms, not proof that a particular person would or would not have committed an offense in some alternative circumstance. Risk management and supervision can also influence observed outcomes after classification, complicating simple interpretations.
Positive predictive value: why a “high-risk” label can mislead
Positive predictive value, or PPV, is the percentage of people given a positive classification who actually experience the specified event during the observation period. Negative predictive value, or NPV, is the percentage given a negative classification who do not experience it. Unlike sensitivity and specificity, predictive values depend strongly on the event's base rate in the population.
Consider a hypothetical cohort of 1,000 people in which 10% experience the defined event. Imagine a test with 80% sensitivity and 80% specificity. Among 100 outcome cases, it correctly flags 80 and misses 20. Among 900 non-outcome cases, it correctly clears 720 but flags 180. Of the 260 people called positive, only 80 experience the event: PPV is about 30.8%. Meanwhile, NPV is 720 of 740, approximately 97.3%. This is a deliberately illustrative arithmetic example, not the published performance of any named forensic instrument.
A technically respectable sensitivity and specificity can therefore coexist with a majority of positive classifications being false positives where the event is relatively uncommon. The same test at a different base rate would have different predictive values. Reports should present the anticipated prevalence and the actual confusion matrix whenever a threshold informs a consequential decision.
A 2012 systematic review and meta-analysis in the BMJ of violence risk tools reported a median positive predictive value of 41% for violent outcomes and a median negative predictive value of 91% across the included studies. These figures concern the studied violence outcomes and populations, rather than a transferable estimate for all forms of criminal recidivism. The authors cautioned against relying on the instruments alone to make decisions about detention or release.
A complete example: 1,000 people, two possible thresholds
Suppose a fictional two-year rearrest model is used for 1,000 eligible individuals and 200 actually experience the recorded outcome. Under threshold A, sensitivity is 75% and specificity is 70%. The confusion matrix contains 150 true positives, 50 false negatives, 240 false positives, and 560 true negatives. PPV is 150 divided by 390, approximately 38.5%, while NPV is 560 divided by 610, approximately 91.8%.
If the institution lowers the cutoff to threshold B, sensitivity could increase to 90% while specificity falls to 50%. The resulting illustrative matrix would contain 180 true positives, 20 false negatives, 400 false positives, and 400 true negatives. The model identifies 30 more future outcome cases but subjects 160 additional non-outcome cases to a high-risk classification. Its PPV falls to about 31.0%.
The second threshold may be defensible for offering low-cost, voluntary outreach where false-positive costs are limited. It would present a profoundly different ethical and legal proposition if the label controlled detention, harsher punishment, or denial of release. The stakes of an intervention are part of the assessment's legitimate interpretation.
These examples also demonstrate why reporting “accuracy” alone can obscure the error structure. Even a method that labels nearly everyone lower risk may look accurate when the outcome is rare. An audit should report the confusion matrix and predictive values alongside discrimination, calibration, and uncertainty intervals.
Population differences and the limits of transportability
A prediction tool validated among adults in one state is not automatically validated among juveniles, people released from different facilities, defendants at a pretrial hearing, people with different criminal-justice histories, or populations living under different legal conditions. Differences in age distribution, eligibility, referral patterns, policing, outcome recording, and intervention practices can alter both ranking and absolute probability.
External validation tests an instrument on data distinct from those used to develop it. Temporal validation tests later cohorts. Geographic validation tests a new site or jurisdiction. Subgroup validation examines whether performance differs across relevant groups. A meaningful validation study specifies how missing data were handled, whether outcomes were measured consistently, which exclusions occurred, and how much uncertainty surrounds estimates.
Risk categories derived in one reference population can become misleading even when the arithmetic scoring rules remain identical. Research on Static-99R and Static-2002R found variation in absolute recidivism rates across validation samples, an illustration of why group-specific observed rates matter. Because those instruments address a specialized offense domain, their evidence should not be generalized to every recidivism tool.
The strongest claim justified by a local validation is bounded: this specified instrument, in this specific version, demonstrated a measured pattern of performance for these eligible individuals, recorded outcomes, and observation period. Broader claims require additional validation evidence.
Subgroup fairness: different definitions can produce different answers
Fairness has statistical, procedural, and legal dimensions. Statistical audits may ask whether predicted probabilities are calibrated within demographic groups, whether false-positive and false-negative rates are comparable, whether people with the same observed outcome receive similar classifications, or whether data and decisions reproduce patterns of unequal criminal-justice contact. These are related questions, but they are not identical measures.
A peer-reviewed analysis by Alexandra Chouldechova established an important mathematical tension: when outcome prevalence differs across groups and prediction is imperfect, several attractive fairness criteria generally cannot all be satisfied simultaneously. Kleinberg, Mullainathan, and Raghavan formalized related incompatibilities. This means that describing a tool as simply “fair” without naming a fairness criterion is scientifically incomplete.
Observed outcome prevalence may itself be influenced by differences in enforcement, surveillance, access to resources, and administrative measurement. A prediction of recorded rearrest can correlate with criminal-justice contact patterns even when the target construct in a public debate is actual offending. Evaluators should examine what the outcome data reflect before treating statistical parity as proof of substantive fairness.
The United States provides a current real-world example. A federal report on review and revalidation of the First Step Act's PATTERN tool examined predictive performance and group differences. Such findings should be read in the context of federal corrections and the instrument's authorized application. They do not establish how a different assessment would behave in a state court, another country, or a clinical population.
A defensible fairness audit reports subgroup sample sizes, event rates, missingness, discrimination, calibration, threshold-specific errors, and uncertainty. It also assesses procedural rights, data quality, access to correction, human review, and the concrete burdens imposed by a score. Group-based metrics must not be used to pronounce an individual's guilt or future behavior.
Algorithms, actuarial models, and professional judgment
An actuarial instrument applies a specified rule to measured inputs. A statistical or machine-learning model may be more complex, but complexity alone provides no guarantee of better external prediction. Structured professional judgment systems ask a trained evaluator to consider specified evidence and formulate risks and management needs; their outputs may take a qualitative or categorical form rather than a calibrated numerical probability.
These approaches answer partly different questions. A model may be designed to rank people by an administrative outcome. A professional risk formulation may emphasize plausible scenarios, relevant circumstances, intervention needs, and uncertainty. Neither procedure becomes legally authoritative merely because it appears structured, computerized, or professionally administered.
Research comparing approaches should distinguish the model's statistical discrimination from the transparency and reproducibility of its decision process. The sentencing instrument systematic review is particularly relevant because it focuses on validation evidence for tools used in a legally consequential setting. Its results do not warrant declaring one category of instrument universally superior across all forensic tasks.
For this article's intent, the central question remains whether any approach offers supported probability statements, realistic estimates of classification errors, and relevant local validation. A clinician's confidence or a model's numerical precision cannot replace those checks.
What can go wrong with records, predictors, and feedback?
Administrative records are incomplete representations of behavior. Arrest history may mix suspected acts with disputed, dismissed, or unproven allegations. Coding inconsistencies and duplicate entries can change numeric scores. Missing treatment information and time-varying circumstances may produce additional error. Sensitive variables and geographic proxies require particular care where their use may reproduce unequal surveillance or disadvantage.
Outcome data may also be affected by decisions based on the assessment itself. A person rated higher risk may receive more intensive supervision and thus have more opportunities for detected technical violations; another may receive more services and experience fewer harmful outcomes. This feedback can make subsequent validation difficult unless studies describe the interventions and observation processes that occur after prediction.
A predictive factor should never automatically be called a causal factor. Even if an item consistently predicts rearrest, removing it from an individual's circumstances may not lower criminal activity, and changing the item in a form may not meaningfully alter the causal process. Prins and Reich identify these conceptual problems in the criminogenic-risk literature. Effective prevention requires appropriately designed intervention studies and outcome evaluation, not merely an elevated association in a model.
The legal question: how may a score actually be used?
A score has no universal legal effect. States and countries vary in rules governing pretrial release, sentencing, parole, probation, confidentiality, due process, admissibility, and human discretion. The evaluator's scientific claim and the court's legally authorized use should be identified separately.
In State v. Loomis (Wisconsin Supreme Court, 2016), the court addressed consideration of COMPAS at sentencing. It allowed consideration under the limitations and cautions described in its opinion, and it emphasized that the assessment was not determinative in the sentencing decision. That holding is a decision under Wisconsin law in its specific procedural setting, not a universal permission or scientific endorsement of every risk tool.
A scientifically responsible expert identifies the legal referral question, describes the instrument's population and outcome, explains observed reliability and validity, and conveys limitations that may be material to the decision. A risk score is neither a finding of guilt nor evidence that a person committed a particular charged offense. It cannot establish criminal responsibility, mental capacity, or the existence of a psychiatric diagnosis.
The American Psychological Association's Specialty Guidelines for Forensic Psychology emphasize competence, appropriate methods, understanding the forensic context, and communication of relevant limitations. The Guidelines are professional guidance; legal obligations remain jurisdiction-specific and are determined by applicable law and court rules.
Can a recidivism score justify incarceration or harsher sentencing?
A score may be considered within some legally authorized decisions, but its statistical properties do not independently settle the normative question. The seriousness of a possible harm, availability of alternatives, presumption and burden standards, procedural fairness, and rights of the person assessed cannot be inferred from a regression coefficient or AUC.
For a liberty-restricting use, false positives create an especially serious concern because many people who would not have experienced the defined event may nonetheless receive a punitive consequence. False negatives may also matter when deciding how to prevent harm, and their consequences depend on the outcome being predicted and intervention considered. Selecting a threshold is therefore an explicit policy decision, subject to law and oversight.
Professionals should avoid statements like “this person will reoffend” when the evidence supports only a group-based probability estimate. They should also avoid turning a score into an apparently precise recommendation without confirming the instrument was validated for that specific decision. The BMJ meta-analysis of violence risk tools illustrates why false-positive burdens limit the legitimacy of using predictions as a stand-alone basis for coercive decisions.
When can risk assessment support rehabilitation and supervision?
A calibrated estimate may help allocate limited services if the specified action is lawful, proportionate, beneficial, and evaluated. Risk and needs information can prompt attention to housing, access to treatment, employment barriers, social support, and supervision conditions. The individual formulation should consider current circumstances and practical capacity to change, rather than treating historical indicators as an immutable identity.
A service-oriented decision requires evidence that the service works for the population and that use of the assessment improves outcomes. A good prediction model can coexist with an ineffective intervention; a poorly calibrated score can misdirect resources. Research on interventions and recidivism addresses a separate empirical question from prediction accuracy.
The rights and preferences of the person assessed, informed explanation, opportunities to challenge inaccurate records, and the availability of less restrictive responses are relevant to ethically sound implementation. Agencies should track not only reoffending outcomes but also service uptake, adverse consequences, decision disparities, and the effect of the whole assessment-and-intervention process.
What a transparent validation report should contain
A technically credible report identifies the tool and version, purpose, legal decision context, eligible population, data sources, time at risk, follow-up horizon, and exact endpoint. It shows how many individuals were evaluated and how many experienced the defined outcome, as well as exclusions, missing data, and confidence intervals.
It then reports discrimination with AUC or appropriate comparable measures; calibration with observed versus predicted frequencies; threshold sensitivity and specificity; PPV and NPV; the confusion matrix; and subgroup performance where sample sizes permit. Where an intervention occurs after the score, the report explains how supervision, treatment, detention, or censoring might affect outcome measurement.
Independent external validation is more persuasive than performance only in the development cohort. Regular revalidation is necessary when policy, population composition, behavior, and data systems change. The U.S. Bureau of Justice Assistance's risk assessment validation resources support the principle that validated implementation and continuing monitoring matter to corrections and pretrial risk work.
A report should also describe the likely consequences of both types of error and who has authority to select the decision threshold. Even when a model is technically reproducible, its legitimate use requires institutional responsibility and the opportunity for meaningful review.
Practical checklist for a judge, lawyer, evaluator, or policymaker
Start by asking what event the risk tool predicts, what observation period and population the estimate refers to, and whether local validation has been completed. Ask whether the reported number is an absolute probability or merely a score that ranks individuals. Request calibration evidence rather than only an AUC.
Where a categorical label is used, ask which cutoff produced it and how many false positives and false negatives arose at that cutoff in a relevant population. Examine fairness metrics across groups, particularly differences in outcome measurement and confidence interval width. Check the scoring record for missing or disputed inputs and clarify what further assessment, if any, contributed to the opinion.
Finally, identify the legally permitted decision and the proportionality of the intended response. A score intended for broad service planning should not quietly become a stand-alone reason to deprive a person of liberty. Record the assumptions, uncertainties, limitations, and mechanisms by which a result can be corrected.
Common misconceptions
“My risk score is my personal probability of committing a crime.” It may be a rank or category rather than a calibrated probability. Even if an instrument outputs a probability, the estimate applies to a defined event within a horizon and is subject to uncertainty and transport limitations.
“An AUC of 0.75 means 75% of people are correctly predicted.” AUC is about pairwise ranking and cannot be interpreted as the proportion of individually accurate forecasts. Overall classification accuracy depends on cutoff and prevalence.
“High risk means the person is guilty or dangerous.” A score describes an imperfect association with a defined future administrative outcome; it does not establish past misconduct, a mental health diagnosis, or criminal guilt.
“If the same tool is applied to everyone, it must be fair.” Identical scoring rules may still yield different errors and consequences across groups. The substantive fairness inquiry must specify the outcome measure, relevant comparison, and legally permissible use.
“Adding more data or using AI necessarily makes the model better.” An additional feature or complex algorithm can overfit, amplify errors in administrative records, or fail when deployed in a different population. External validation and calibration supply the evidence of practical performance.
“Risk scores are useless because they sometimes produce false positives.” Their possible value depends on the accuracy demonstrated in context, the decision being supported, the costs of errors, available alternatives, and monitoring. Scientific assessment demands that these conditions be documented.
Frequently asked questions
What is recidivism risk assessment?
It is a structured estimation or classification of the chance of a clearly defined future justice-system outcome, such as reconviction within two years, for a specified population. The tool may use criminal history and other validated indicators. Its interpretation requires information about the event definition, horizon, reference group, and error rates.
How accurate are recidivism prediction tools?
Accuracy varies by instrument, population, outcome, and metric. A systematic review of tools used at sentencing reported external-validation AUC values of approximately 0.57–0.75 in larger samples, while also identifying major reporting limitations. AUC does not establish calibration or individual certainty.
What is a false positive in risk assessment?
A false positive occurs when someone crosses a specified risk threshold but does not experience the defined outcome during the stated follow-up. The percentage of all positive classifications that are false positives depends in part on the outcome base rate and cutoff.
What is a false negative?
A false negative occurs when someone is classified below the positive threshold but experiences the defined event during follow-up. Its frequency can be expressed relative to all actual outcome cases and should be considered together with the effects of the decision threshold.
What is calibration, and why does it matter?
Calibration is the agreement between predicted probabilities and observed event frequencies in comparable groups. It matters when agencies communicate numeric risk or use fixed probability thresholds. Good discrimination cannot substitute for calibration.
Can a risk tool predict one person's exact future?
A tool can produce an estimated probability or risk category based on patterns in groups. It cannot establish exactly what any particular person will do, and even a correct group-level estimate remains uncertain at the individual level.
Are risk assessments biased against demographic groups?
The answer depends on the model, dataset, population, outcome, decision threshold, and definition of fairness. Researchers examine group-specific calibration, false-positive and false-negative rates, representation, and the underlying outcome process. Because fairness criteria can conflict when outcome prevalence differs, an assessment needs multiple disclosed measures rather than a single label of “unbiased.”
Are clinical diagnoses used to predict recidivism?
Some specialized evaluations consider clinically relevant information, but a diagnosis alone does not establish offending or future dangerousness. Any diagnostic item requires evidence of relevance to the validated instrument and referral question. Clinical diagnosis, legal culpability, and statistical prediction remain separate determinations.
Do risk scores determine parole or sentencing?
Some legal systems permit scores to inform specified decisions, while others limit their use. Their authority and weight depend on the applicable jurisdiction and procedure. The Wisconsin Supreme Court's Loomis opinion illustrates a particular legal treatment of one tool, not a nationwide rule.
Can predictive tools be improved?
Yes, through better data quality, independent external validation, clear outcome definitions, calibration and recalibration, subgroup checks, transparent threshold setting, meaningful human review, and evaluation of actual decision outcomes. Improvement must be measured in the intended application rather than assumed from a model update.
Related Articles
References
American Psychological Association. (2013). Specialty guidelines for forensic psychology. American Psychologist, 68(1), 7–19. https://doi.org/10.1037/a0029889
Beaudry, G., et al. (2023). Prediction of violent reoffending in people released from prison in England: External validation study of a risk assessment tool (OxRec). Journal of Criminal Justice. https://doi.org/10.1016/j.jcrimjus.2023.102061
Bureau of Justice Assistance. Risk assessment validation resources. U.S. Department of Justice. https://bja.ojp.gov/program/psrac/validation/risk-validation
Bureau of Justice Statistics. Recidivism program. U.S. Department of Justice. https://bjs.ojp.gov/recidivism-program
Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. https://doi.org/10.1089/big.2016.0047
Fazel, S., et al. (2022). The predictive performance of criminal risk assessment tools used at sentencing: Systematic review of validation studies. Journal of Criminal Justice, 81, 101902. https://doi.org/10.1016/j.jcrimjus.2022.101902
Fazel, S., Singh, J. P., Doll, H., & Grann, M. (2012). Use of risk assessment instruments to predict violence and antisocial behaviour in 73 samples involving 24,827 people: Systematic review and meta-analysis. BMJ, 345, e4692. https://doi.org/10.1136/bmj.e4692
Helmus, L., et al. (2012). Absolute recidivism rates predicted by Static-99R and Static-2002R sex offender risk assessment tools vary across samples. Criminal Justice and Behavior. https://doi.org/10.1177/0093854812443648
Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. Innovations in Theoretical Computer Science. https://doi.org/10.4230/LIPIcs.ITCS.2017.43
Office of Justice Programs. (2024). 2023 Review and Revalidation of the First Step Act Risk Assessment Tool. U.S. Department of Justice. https://www.ojp.gov/library/publications/2023-review-and-revalidation-first-step-act-risk-assessment-tool
Prins, S. J., & Reich, A. (2022). Criminogenic risk assessment: A meta-review and critical analysis. Punishment & Society. https://doi.org/10.1177/14624745211025751
State v. Loomis, 2016 WI 68, Wisconsin Supreme Court. https://www.wicourts.gov/sc/opinion/DisplayDocument.pdf?content=pdf&seqNo=171690