Predictive Factors
SHAP feature importance and unadjusted rearrest rate differences for the NIJ Year 1 static model.
Top 10 SHAP Features (Y1)
ImportanceMean absolute SHAP values quantify each feature's average contribution to individual predictions. Gang affiliation status is the single strongest predictor, followed by prior-violation history and age group indicators.
SHAP Feature Importance (Y1 Static XGBoost)
Mean |SHAP value| across the test set, sorted descending
Unadjusted Rearrest Rate Deltas
Base rate: 29.4%Each bar shows the difference in Y1 rearrest rate between the subgroup and the overall base rate (29.4%). Positive values indicate a higher-than-average rearrest rate; negative values indicate a protective association. These are unadjusted—they do not control for other covariates.
Rearrest Rate Delta from Base Rate (Y1)
Percentage-point difference from the 29.4% overall rearrest rate
Unadjusted differences. Actual causal effects may differ after controlling for confounders.
Age & Rearrest
GradientRearrest rates decline monotonically with age. The youngest cohort (18–22) has more than double the rearrest rate of the oldest (48+), consistent with well-established age-crime curve findings in criminology.
Y1 Rearrest Rate by Age Group
Percentage of each age group rearrested within one year
Cross-Horizon Comparison
Y1 / Y2 / Y3Top predictive factors shift across forecast horizons. Year 2 and Year 3 models are trained on conditional populations—individuals who were not rearrested at the prior horizon—so the population composition changes. Factors that dominate early prediction (gang affiliation, young age) may decrease in relative importance as the remaining population becomes lower-risk.
| Feature | Y1 Rank | Y2 Rank | Y3 Rank | Trend |
|---|---|---|---|---|
| Gang Affiliated | 1 | 1–2 | 2–3 | Remains important; slight decline |
| Prior Violations (none) | 2 | 2–3 | 1–2 | Stable or rising |
| Age 23–27 | 3 | 4–5 | 5–7 | Declines as young cohort filters out |
| Supervision Risk Score | 4 | 3–4 | 3–4 | Relatively stable |
| Age 48+ | 5 | 5–6 | 4–5 | Protective effect persists |
| Property Arrests 5+ | 7 | 6–8 | 6–8 | Stable mid-tier predictor |
Note: Y2 and Y3 are conditional populations (those not rearrested at prior horizons). The population changes at each horizon, so rank shifts reflect both changing feature importance and changing cohort composition.