Input Variables and Validation Studies for the VPRAI and Similar Pretrial Risk-Assessment Tools

What the evidence says

The sources agree on what goes into the tools. The VPRAI takes age, sex, and arrest history, according to a reverse-engineering of public documentation. Another source describes it as using current charge, criminal history, employment status, and drug use. Race, ethnicity, and gender are not explicit inputs. A model fitted to estimate true pretrial failure associates prior failures to appear, unemployment, a history of drug abuse, and more violent arrests with higher risk of reoffending or failing to appear. One review adds that variables such as financial stress, the criminality of friends or family, parental neglect, and domestic violence exposure can increase a risk score, and that risk-tool developers do not widely publicize them.

The main measured outcome is a racial gap in scores. Among individuals matched on age, sex, and arrest history, the VPRAI gives Black individuals scores 2.76 percentage points higher than White individuals. When the comparison uses estimated total crime history instead of arrests alone, the gap grows to 10.72 percentage points, a 7.96-point increase. The mechanism the authors name is that arrest history stands in for race. Arrest records reflect differences in arrest rates between groups, so the tool can carry those differences into scores even though race is not an input. A sensitivity test supports this. Multiplying Black arrest rates by 5 increases the gap from 10.72 to 18.22 percentage points. These estimates rest on semi-synthetic crime data built from national surveys, applied to a Harris County cohort.

A second study measures the same pattern as detention recommendations. Among defendants estimated to be likely to reoffend or fail to appear, the VPRAI recommends detention for 86.0% of Black defendants and 49.3% of White defendants. This depends on a latent-variable model of true outcomes, so the direction is better established than the exact figures. A legal analysis reaches the same direction for commonly used algorithms. It reports that nearly all include variables correlated with race, which produce proxy effects that disadvantage Black individuals. It also reports that two proposed algorithms that remove these proxy effects decrease the number of Black defendants detained. Neither the size of the proxy effect nor the size of the decrease is quantified in the excerpts here.

Evidence on predictive validation is thinner and mostly descriptive. Models built on data from one location show decreased predictive performance when applied to another. The datasets differ in available variables: Broward has "age at first offense" and "prior juvenile charges," which the Kentucky data cannot supply. They also differ in how recidivism is defined, with non-convicted charges in Broward and convicted charges in Kentucky. No baseline algorithm consistently performs best in two-year prediction, and simple linear models sometimes outperform black-box models. The Arnold PSA's New Criminal Activity score fails calibration for higher score values. For the federal PATTERN tool, the development report relied on AUC as its primary validity metric. It reports an AUC of .77 to .78 for violent recidivism, on a sample where 15% were rearrested for suspected violence within three years. The general recidivism sample is roughly balanced, with 45.4% returning to custody or rearrested.

No claim here reports a VPRAI validation study of predictive accuracy, such as its own AUC or calibration. No claim measures how the input variables change actual release, detention, or appearance outcomes in practice.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

• What are the VPRAI's own predictive accuracy and calibration, overall and by race, in the populations where it is deployed?
• How much of the 2.76 to 10.72 percentage-point score gap survives when true crime history is measured directly instead of estimated from national surveys?
• How large is the decrease in Black defendant detention from removing race-proxy variables, and what does it do to failure-to-appear and rearrest rates?
• Do variables such as financial stress or parental neglect enter the VPRAI, and what weight do they carry?