The Color of Risk: Unpacking the Implications of Actuarial Risk Prediction at Sentencing

Restricted (Penn State Only)
Laskorounskaia, Julia A
Graduate Program:
Doctor of Philosophy
Document Type:
Date of Defense:
June 15, 2017
Committee Members:
  • Jeffrey Todd Ulmer, Dissertation Advisor
  • Jeffrey Todd Ulmer, Committee Chair
  • Gary Zajac, Committee Member
  • Eric P Baumer, Committee Member
  • Christopher Jon Zorn, Outside Member
  • sentencing
  • judges
  • recidivism
  • pennsylvania
  • race
  • risk assessment
This dissertation examines the practice of using actuarial risk prediction to inform sentencing decisions and assesses its implications for racial disparity in sentencing. I use a large sample of serious offenders (N=7,935), convicted and released in Pennsylvania, to construct and validate a risk assessment instrument and analyze the relationship between race, risk score, and recidivism. In the first part of the analyses, I introduce an approach to develop a static risk assessment instrument that judges can use at sentencing. Using the development sub-sample, a series of bivariate and multivariate logistic regression analyses are used to identify significant covariate patterns among convicted individual with a new arrest or parole revocation within three-years of release. I identify eight parsimonious risk factors associated with recidivism and use a modified Burgess scoring method to construct an 11-point additive risk scale. Scores are collapsed into Low, Medium, High, and Very High risk levels to maximize interpretability. Finally, the instrument is validated on the validation sub-sample. The area under the curve values (AUC) for the risk score are .72 for the full validation sample and .70 for both the White and Black offender sub-samples — showing a medium to strong ability to discriminate between recidivist and non-recidivism. In the second part of the analyses, I conduct disparate impact and instrument bias analyses to determine the instrument’s predictive utility across race. I use a series of bivariate and multivariate logistic regression models to examine the extent to which race is associated with recidivism, net of other factors, and to test for instrument fairness and interaction effects between risk factors and race. Results show that Black offenders receive higher average risk scores on the instrument than White offenders, mostly due to differences in their criminal histories. However, the instrument slightly under-predicts the recidivism rate for Black offenders, and over-predicts the recidivism rate for White offenders, because race is significantly correlated with recidivism, but not included in the scoring model. I situate the consideration of risk in the focal concerns perspective of sentencing and discuss the policy implication of using actuarial risk assessment tools to structure sentencing decisions.