Definition
Measures the degree to which AI-driven non-renewal recommendations disproportionately affect policyholders in demographic proxy groups, revealing whether predictive non-renewal models produce disparate impact.
The Measurement Gap
Carriers increasingly use AI models to identify policies for non-renewal. These models predict which policyholders are likely to generate future losses exceeding profitability thresholds and recommend non-renewal before those losses materialize. The models use property characteristics, geographic factors, claims history, and coverage utilization patterns to make these predictions.
The problem is that many of these predictive features correlate with protected class membership. Property age, property value, geographic location, and neighborhood characteristics are powerful loss predictors, but they also correlate with the racial and economic composition of the community. A model that non-renews based on these features may produce non-renewal rates that differ significantly across demographic groups, even though the model never uses protected class information directly.
Legacy non-renewal metrics report non-renewal volumes, reasons, and retention rates. They do not measure whether non-renewal rates vary across demographic proxy groups after controlling for actuarially justified risk factors. A carrier can report that all non-renewal decisions follow consistent underwriting guidelines while producing outcomes that would fail a regulatory disparate impact analysis.
NRDII closes this gap by comparing non-renewal rates across demographic proxy groups, controlling for actuarially justified risk factors. The residual disparity after controlling for legitimate risk factors is the disparate impact signal. This makes the fairness dimension of non-renewal decisions visible and measurable before a regulator discovers it.
Decision Supported:
Is the predictive non-renewal model producing recommendations that would fail a disparate impact analysis, and which model features are the primary drivers of the disparity?
Formula
Plain Language: Compare non-renewal rates across demographic proxy groups (defined by census tract demographics), controlling for actuarially justified risk factors. If non-renewal rates vary systematically after controlling for legitimate risk factors, the residual disparity is the disparate impact signal.
Formula:
NRDII_g = (NonRenewal_rate_g - NonRenewal_rate_expected_g) / NonRenewal_rate_expected_g x 100
NRDII = max|NRDII_g| across all groups g
Where NonRenewal_rate_expected_g = expected non-renewal rate for group g based on actuarially justified risk factors only.
Variables:
- NonRenewal_rate_g: actual non-renewal rate for policies in demographic proxy group g (percentage of policies non-renewed)
- NonRenewal_rate_expected_g: non-renewal rate predicted by a model using only actuarially justified risk factors (loss history, coverage changes, claim frequency) without property/geographic proxies (percentage)
- g: demographic proxy group (census tract income quintile, racial/ethnic composition quartile)
Units:
Percentage deviation from expected. Report both the maximum absolute deviation and group-specific values.
Data Requirements:
| Data Element | Source System | Classification |
|---|---|---|
| Non-renewal records with reason codes | Policy administration system / underwriting system | NATIVE |
| AI model non-renewal recommendation and score | Predictive non-renewal model output | NATIVE |
| Policy addresses / geocoding | Policy administration system | NATIVE |
| Census tract demographics | US Census Bureau ACS data | INTEGRATION |
| Actuarially justified risk factors (loss history, claims, coverage) | Policy and claims systems | NATIVE |
Proposed formula, pending validation pilots.