Predictive bias in pretrial risk assessment: Application of the Public Safety Assessment in a Native American...

Native Americans are vastly overrepresented in U.S. jails and people in rural communities face unique barriers (e.g., limited public transportation and services) that may impact how well pretrial risk assessments predict outcomes. Yet, these populations are understudied in the literature examining the predictive validity and, more importantly, the potential predictive bias of pretrial risk assessments. We sought to address these gaps.

We had three aims: (a) examine the validity of Public Safety Assessment (PSA) scores in predicting pretrial outcomes in a county with a high degree of rurality, (b) compare predictive validity and test for predictive bias among Native American and White people, and (c) compare predictive validity and test for predictive bias among men and women.

Our sample comprised 4,570 closed cases involving people released on personal recognizance bonds over a 3.5-year period. About two thirds were Native American and men. The PSA was completed and outcome data were collected as part of routine pretrial practice.

In slightly more than one third of cases, people failed to appear or were rearrested during the pretrial period. In the full sample, PSA scores demonstrated poor validity in predicting failure to appear but fair validity in predicting new arrest. Further analyses revealed predictive bias as a function of both race and sex in the prediction of failure to appear. In contrast, we did not find evidence of bias in the prediction of new criminal arrest, although predictive validity was slightly better for White people and men.

Our findings raise concerns regarding the use of PSA scores to inform pretrial decisions related to risk for failure to appear in rural communities and among Native American people. They also highlight concerns regarding reliance on static factors as well as the need for research on the validity of pretrial risk assessments in these populations. (PsycInfo Database Record (c) 2025 APA, all rights reserved).