Cohort bias in predictive risk assessments of future criminal justice system involvement

Abstract

Risk assessment instruments (RAIs) are widely used to aid high-stakes decision-making in criminal justice settings and other areas such as health care and child welfare. These tools, whether using machine learning or simpler algorithms, typically assume a time-invariant relationship between predictors and outcome. Because societies are themselves changing and not just individuals, this assumption may be violated in many behavioral settings, generating what we call cohort bias. Analyzing criminal histories in a cohort-sequential longitudinal study of children, we demonstrate that regardless of model type or predictor sets, a tool trained to predict the likelihood of arrest between the ages of 17 and 24 y on older birth cohorts systematically overpredicts the likelihood of arrest for younger birth cohorts over the period 1995 to 2020. Cohort bias is found for both relative and absolute risks, and it persists for all racial groups and within groups at highest risk for arrest. The results imply that cohort bias is an underappreciated mechanism generating inequality in contacts with the criminal legal system that is distinct from racial bias. Cohort bias is a challenge not only for predictive instruments with respect to crime and justice, but also for RAIs more broadly.

Keywords: bias; cohort; criminal justice; risk assessment; social change.

Fig. 1.

ROC curves showing performance of logistic regression models. Models were trained on the older cohort. Performance on the older cohort is shown in red, while performance on the younger cohort is shown in blue. Results are shown for both (A) an unregularized logistic regression using the classic risk-factor feature set and (B) a lasso logistic regression using the full feature set.

Fig. 2.

Calibration plots for models trained on older cohort. Calibration plots for the classic risk-factor unregularized logistic regression model (Top) and the full lasso logistic regression model (Bottom). Models were trained on the older cohort, and performance on the older cohort is shown on the Left, while performance on the younger cohort is shown on the Right. The regression line falls below the ideal 45° line indicating persistent overprediction of the likelihood of arrest for the younger cohort in both models.

Fig. 3.

Comparison of rankings between full lasso logistic regression models trained on different cohorts. Plot of younger cohort, risk rankings produced by the full lasso logistic regression model trained on the older cohort (x axis) and the same model trained on the younger cohort (y axis). Both the plot and the Spearman’s correlation value of 0.76 show that while the two sets of rankings are correlated, there are substantial ranking differences between the model produced by the older cohort and that produced by the younger cohort.

Fig. 4.

Calibration plots for classic risk-factor logistic regression model trained on older cohort by race. Performance on the older cohort of the classic risk-factor logistic regression model trained on the older cohort is shown in red, and performance on the younger cohort is shown in blue. Systematic overprediction of the likelihood of arrest is observed across racial groups.

Fig. 5.

Calibration plots for models including an adolescent arrest history predictor and trained on older cohort. Calibration plots for two models which include a binary variable indicating arrest at age 17 or 18 y: the classic risk-factor logistic regression model (Top) and the full lasso regularized logistic regression model (Bottom). The left column shows the performance of these models on the older cohort on which they were trained. The right column shows the performance of the same models on the younger cohort. When arrest history is included as a predictor, models still overpredict arrest likelihood for the younger cohort.

Fig. 6.

Calibration plots for models trained on older cohort applied to individuals who are two SDs below the average on self-control. The Top row shows the calibration of the classic risk-factor logistic regression model with performance for the older cohort on the Left and performance for the younger cohort on the Right. The Bottom row shows the performance of full lasso logistic regression model. The deviation of the regression line from the diagonal in the right column indicates that when considering only individuals with low self-control, cohort bias is observed.