VPRAI (or Comparable Pretrial Risk-Assessment Instrument) Racial Gap Replication

confidence: high · 3 sources

What the evidence says

Among defendants the VPRAI expects to reoffend or fail to appear, the tool recommends detention for 86.0% of Black defendants but only 49.3% of white defendants with the same expected risk (Hochstedler Webb 2023). This gap traces to a sensitivity/specificity asymmetry: the VPRAI's specificity is high and stable (99.6% — it correctly recommends release almost every time release is warranted), while its sensitivity is both lower overall (67.5%) and racially variable, running 86.0% for Black defendants and 49.3% for white defendants among those expected to fail. The same pattern — a detention recommendation that tracks race more than it tracks the underlying risk being measured — appears even among defendants expected to have neither a new offense nor a failure to appear, where Black defendants are still more likely to receive a VPRAI detention recommendation, though the source does not give a side-by-side percentage for that subgroup.

Judicial decisions add a second layer of disparity on top of the tool's own output. Among defendants not expected to fail and who received a VPRAI release recommendation, judges release white defendants 60.3% of the time versus 49.0% of the time for Black defendants — a gap that cannot be attributed to the tool, since both groups received the same recommendation. Judges also detain Black defendants at higher rates than white defendants regardless of VPRAI recommendation or of whether the defendant was, in fact, a pretrial failure. The compounded result of tool-level and judicial-level disparity is a wrongful detention rate of 51.4% for Black defendants versus 39.7% for white defendants. Consistent with both layers, Black defendants are misclassified more often than white defendants by both the VPRAI itself and, to a greater degree, by judicial officers.

The pattern is not limited to the VPRAI. For Kentucky's PSA instrument, the failure-to-appear (FTA) scale is more predictive for white defendants (ROC = 0.655) than for Black defendants (ROC = 0.612), and the same source infers a higher false positive rate for Black defendants under PSA scoring, though this is not a directly measured comparison in the study. COMPAS shows a comparable asymmetry: an independent audit (Angwin et al., cited in Bolinger 2021) found COMPAS disproportionately misclassifies Black defendants as high-risk and disproportionately misclassifies white defendants as low-risk, i.e., a higher false-positive rate for Black defendants and a higher false-negative rate for white defendants.

Calibration does not resolve this. A theoretical result cited by Faddoul (2020) establishes that because base rates of reoffense/FTA differ across race and gender, calibration and error-rate balance cannot both hold at once — calibrating a tool does not remove its predictive bias. Empirically, the PSA is described as "relatively well calibrated" between white and Black defendants on two of its three risk scores (NCA and NVCA), but not on the FTA score, where Black defendants receive worse calibration — consistent with the ROC gap reported for that same scale.

The sources here do not measure what happens downstream of these detention and release disparities — actual recidivism or FTA outcomes for released versus detained defendants by race are not reported — and no claim below gives sample sizes or replication across jurisdictions beyond the single study each finding is drawn from.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

Does the VPRAI detention-recommendation gap (86.0% vs. 49.3%) replicate in jurisdictions outside the one studied by Hochstedler Webb (2023)?
Does the inferred higher false-positive rate for Black defendants under PSA hold when measured directly, rather than inferred from differential ROC?
What pretrial-failure or recidivism outcomes actually follow from the detention and release disparities documented here — none of the sources track outcomes for released vs. detained defendants by race?
Is the PSA's FTA-specific calibration gap confirmed by validation studies outside Kentucky?

Findings

VPRAI detention recommendations increase for Black defendants relative to white defendants among those expected to reoffend or fail to appear

confidence — direction: highconfidence — magnitude: highconsensus: thin — single study, direction leans toward higher detention recommendations for Black defendants

mechanism: the tool's sensitivity (true positive rate) is racially variable while its specificity stays high and race-stable
effect: 86.0% detention-recommendation rate for Black defendants vs. 49.3% for white defendants among those expected to reoffend/FTA; overall sensitivity 67.5%, specificity 99.6%
conditions / moderators: holds among defendants expected to reoffend or fail to appear; also observed (direction only, not quantified) among defendants expected to have neither outcome

Kimberly A. Hochstedler Webb (2023), arXiv · doi:10.48550/arxiv.2309.08599v2 · https://arxiv.org/pdf/2309.08599v2 (retrieved 2026-08-14)

Judges decrease release rates for Black defendants relative to white defendants beyond VPRAI's own recommendation

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single study, direction leans toward lower release rates for Black defendants

mechanism: judicial discretion diverges from the tool's recommendation in a racially patterned direction
effect: among defendants not expected to fail with a VPRAI release recommendation, 60.3% of white defendants are released vs. 49.0% of Black defendants; Black defendants are also detained at higher rates regardless of VPRAI recommendation or true failure status (direction only for this second comparison)
conditions / moderators: observed both when the tool recommends release and independent of the tool's recommendation

Kimberly A. Hochstedler Webb (2023), arXiv · doi:10.48550/arxiv.2309.08599v2 · https://arxiv.org/pdf/2309.08599v2 (retrieved 2026-08-14)

Wrongful detention increases for Black defendants relative to white defendants

confidence — direction: highconfidence — magnitude: highconsensus: thin — single study, direction leans toward higher wrongful detention for Black defendants

mechanism: combined effect of tool-level and judicial-level disparities described above
effect: 51.4% wrongful detention rate for Black defendants vs. 39.7% for white defendants
conditions / moderators: not established

Kimberly A. Hochstedler Webb (2023), arXiv · doi:10.48550/arxiv.2309.08599v2 · https://arxiv.org/pdf/2309.08599v2 (retrieved 2026-08-14)

Misclassification increases for Black defendants relative to white defendants, in both VPRAI and judicial decisions

confidence — direction: highconfidence — magnitude: highconsensus: thin — single study, direction leans toward higher misclassification for Black defendants

mechanism: not established beyond the tool sensitivity/specificity gap and judicial deviation described above
effect: direction only; magnitude reported as context-dependent, with judicial misclassification exceeding VPRAI misclassification for Black defendants
conditions / moderators: not established

Kimberly A. Hochstedler Webb (2023), arXiv · doi:10.48550/arxiv.2309.08599v2 · https://arxiv.org/pdf/2309.08599v2 (retrieved 2026-08-14)

PSA's FTA scale predictive accuracy decreases for Black defendants relative to white defendants

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single study, direction leans toward lower FTA predictive accuracy for Black defendants

mechanism: not established; reported as a calibration gap specific to the FTA score (PSA calibrated on its other two scores)
effect: ROC = 0.655 for white defendants vs. 0.612 for Black defendants on the FTA scale; higher false positive rate for Black defendants (inferred, not a direct measurement); FTA score shows worse calibration for Black defendants while the NCA and NVCA scores are relatively well calibrated across race
conditions / moderators: applies specifically to the FTA scale, not to NCA or NVCA scores

Marc Faddoul (2020), arXiv · doi:10.48550/arxiv.2005.07299v1 · https://arxiv.org/pdf/2005.07299v1 (retrieved 2026-08-14)

Calibration produces no measured effect on racial predictive bias when base rates differ across groups

confidence — direction: highconfidence — magnitude: highconsensus: thin — single theoretical source

mechanism: mathematical incompatibility between calibration and error-rate balance when base rates of reoffense/FTA differ across race and gender
effect: no measured effect — calibration efforts do not remove predictive bias under differing base rates
conditions / moderators: applies when base rates differ between the groups being compared

Marc Faddoul (2020), arXiv · doi:10.48550/arxiv.2005.07299v1 · https://arxiv.org/pdf/2005.07299v1 (retrieved 2026-08-14)

COMPAS increases high-risk misclassification for Black defendants and increases low-risk misclassification for white defendants

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single cited audit (Angwin et al. 2016), direction leans toward higher false-positive rate for Black defendants

mechanism: not established in the source beyond the audit finding itself
effect: higher false-positive rate for Black defendants (misclassified as high-risk); higher false-negative rate for white defendants (misclassified as low-risk)
conditions / moderators: not established

Renée Jorgensen Bolinger (2021), Canadian Journal of Philosophy · doi:10.1017/can.2021.28 · https://www.cambridge.org/core/services/aop-cambridge-core/content/view/3EA248379E4082CFFECEAB74750999EF/S004550912100028Xa.pdf/div-class-title-algorithms-and-the-individual-in-criminal-law-div.pdf (retrieved 2026-08-15)