Fair Lending and Insurance: Disparate Impact Analysis

Disparate Impact in Insurance

Disparate impact occurs when a facially neutral practice or rating variable disproportionately affects a protected class, even without discriminatory intent. In insurance, this concept has become increasingly important as pricing models incorporate more variables and use more sophisticated analytical techniques. Variables like credit score, education level, occupation, ZIP code, and homeownership status may serve as proxies for race, ethnicity, or income because they are correlated with these protected characteristics in the population. Regulators and consumer advocates have increasingly scrutinized insurance pricing models for disparate impact, even when the models do not explicitly use protected class variables.

The legal framework for disparate impact in insurance is complex and varies by jurisdiction. Unlike employment and lending, where federal laws clearly prohibit disparate impact discrimination, insurance regulation is primarily state-based, and the applicability of disparate impact standards varies. Some states have enacted specific prohibitions on unfairly discriminatory insurance practices that encompass disparate impact analysis. Others rely on the traditional actuarial standard that rates should not be "unfairly discriminatory," which has historically been interpreted to mean that rate differences should reflect differences in expected losses or expenses. The tension between these interpretations drives much of the current debate about fairness in insurance pricing.

Understanding Proxy Discrimination

Proxy discrimination occurs when a rating variable that is not itself a protected characteristic serves as a statistical proxy for one. The mechanism is straightforward: if a rating variable is correlated with a protected characteristic, using that variable in a pricing model produces rate differentials that align with the protected characteristic, even though the protected characteristic itself is not used as an input. For example, credit-based insurance scores are correlated with race and income in the United States. Using credit scores in auto insurance pricing therefore produces rate differences that correlate with race, even though race is not an explicit rating variable.

The challenge for actuaries is that many variables with proxy potential are also genuinely predictive of loss. Credit scores, for instance, are strongly correlated with insurance loss experience across multiple lines of business, independent of their correlation with protected characteristics. This creates a legitimate actuarial justification for their use. The question becomes whether the actuarial predictiveness of a variable justifies its use when that use also produces disparate impact on protected classes. Different stakeholders, including regulators, consumer advocates, insurers, and actuarial standards bodies, may reach different conclusions on this question.

Analytical Methods for Assessing Disparate Impact

Actuaries use several techniques to assess whether a pricing model produces disparate impact. The adverse impact ratio compares the rate or outcome for a protected group to the rate or outcome for a reference group. In employment discrimination analysis, the "four-fifths rule" considers an adverse impact ratio below 0.80 as evidence of potential discrimination. While this specific threshold is not universally applied in insurance, the concept of comparing outcomes across groups is fundamental to disparate impact analysis.

Regression analysis provides a more nuanced assessment by testing whether disparities persist after controlling for legitimate risk factors. If the protected characteristic remains a statistically significant predictor of the insurance outcome (premium charged, claim denial, etc.) after including all legitimate rating variables in the model, this suggests that the pricing model may not fully account for the risk differences that drive the observed disparities. Conversely, if the protected characteristic loses significance after controlling for legitimate factors, the observed disparities may be attributable to genuine risk differences rather than discrimination.

Proxy discrimination testing evaluates the impact of specific variables by comparing model outcomes with and without the suspect variable. If removing a variable substantially reduces the disparate impact on a protected class without significantly degrading the model's predictive accuracy, this suggests that the variable is functioning primarily as a proxy rather than as an independent predictor of risk. The Shapley value decomposition, borrowed from cooperative game theory, can quantify how much of the model's disparate impact is attributable to each individual variable, helping actuaries identify the most problematic proxies.

Fairness Metrics and Their Trade-offs

Defining "fairness" in a quantitative context is surprisingly difficult because different fairness metrics can conflict with each other. Several commonly discussed fairness criteria include demographic parity (equal outcomes across groups), equalized odds (equal true positive and false positive rates across groups), calibration (predicted probabilities match observed frequencies within each group), and individual fairness (similar individuals receive similar treatment regardless of group membership).

A fundamental result in the fairness literature, sometimes called the "impossibility theorem of fairness," demonstrates that it is generally impossible to satisfy all fairness criteria simultaneously except in trivial cases. For example, if base rates (the true underlying risk levels) differ between groups, a well-calibrated model will necessarily produce different average predictions for each group, violating demographic parity. Actuaries must therefore choose which fairness criteria to prioritize, and this choice involves value judgments that extend beyond purely technical considerations.

In insurance, calibration is often considered the most natural fairness criterion because it aligns with the actuarial principle that premiums should reflect expected losses. A calibrated model charges each policyholder a premium that accurately predicts their expected cost, regardless of group membership. However, calibration alone does not prevent disparate impact if the underlying risk factors used in the model are themselves correlated with protected characteristics.

Bias Mitigation Techniques

Several techniques can reduce disparate impact in insurance pricing models while preserving predictive accuracy. Variable exclusion removes the suspect proxy variable from the model entirely. This is the simplest approach but may sacrifice predictive accuracy, particularly if the excluded variable captures legitimate risk information not available from other variables. Pre-processing techniques transform the input data to reduce its correlation with protected characteristics before model fitting. For example, the residualization approach regresses the suspect variable on the protected characteristic and uses only the residual (the portion uncorrelated with the protected characteristic) as the model input.

In-processing techniques modify the model fitting algorithm to incorporate fairness constraints directly. Constrained optimization adds a penalty term to the objective function that penalizes disparate impact, producing a model that balances predictive accuracy against fairness. The trade-off between accuracy and fairness can be controlled by adjusting the weight of the penalty term. Post-processing techniques adjust model outputs after fitting to reduce disparate impact. For example, group-specific thresholds or calibration adjustments can equalize outcomes across groups while maintaining the model's overall predictive performance.

Regulatory Landscape

The regulatory approach to disparate impact in insurance is evolving. Colorado's SB 21-169, enacted in 2021, was among the first state laws to explicitly address algorithmic discrimination in insurance, requiring insurers to test their models for unfairly discriminatory outcomes. Other states have proposed or enacted similar legislation. The NAIC has formed working groups to develop model laws and guidance on the use of big data and algorithms in insurance, including fairness testing requirements.

At the federal level, the Fair Housing Act's disparate impact standard applies to homeowners insurance, and several federal agencies have expressed interest in how AI and algorithms affect insurance fairness. The evolving regulatory landscape means that actuaries must stay informed about jurisdictional requirements and be prepared to conduct and document fairness analyses as part of their standard pricing workflow.

Study Relevance and Professional Responsibility

Disparate impact analysis appears in actuarial exam syllabi covering ratemaking, data analytics, and professionalism. ASOP No. 56 (Modeling) requires actuaries to consider the appropriateness of their models, which increasingly includes fairness considerations. Candidates should understand the concepts of disparate impact and proxy discrimination, the analytical methods for assessing and mitigating bias, and the trade-offs between different fairness metrics. As the profession's role in ensuring fair and equitable insurance pricing grows, competency in disparate impact analysis is becoming essential for actuarial practice across all lines of business.