How to Search for Less Discriminatory Alternatives in Models

Current as of October 30, 2026. This guide draws on Regulation B and the 2026 CFPB rule amending it, Executive Order 14281, OCC, FDIC and HUD texts, and New York and Colorado insurance texts, as published on that date. Which of them apply depends on the firm; it is not legal advice.

A search for a less discriminatory alternative (LDA) asks whether a different variable, model or decision policy would meet the same business need with smaller disparities across protected groups. Whether a firm must run one depends on the product, the regulator and the state, and in the US those positions have moved; the box below records them with their dates. The method does not depend on them: a fixed performance bar, sound disparity measures, a wide set of candidates, a fair comparison and a decision record.

This guide sets out that method as model-development and model-risk practice. Unless a step quotes a rule, it is a recommendation, labeled (interpretation); the techniques described are options, not requirements found in the texts reviewed. Testing mechanics for insurers are covered in our guide to testing insurance models for unfair discrimination.

On scope: two input limits apply whatever the position on disparate impact. Regulation B, which applies to creditors, says "Except as provided in the Act and this part, a creditor shall not take a prohibited basis into account in any system of evaluating the creditworthiness of applicants", and New York says an insurer should not use external data or AI systems for underwriting or pricing unless it can establish that they do not use and are not "based in any way on any class protected pursuant to Insurance Law Article 26". A search measures protected-group outcomes; it does not add protected characteristics to the model (interpretation).

Step 1: Define the business need and the performance bar before searching

New York frames the test around an alternative "that would reasonably meet the insurer's legitimate business needs". None of the texts reviewed defines how much predictive performance an alternative may give up, so the firm has to set the bar (interpretation).

Write down, before any candidate is built (interpretation):

• The business objective: what the model decides, for which product and population, and what it must achieve, such as loss, default or approval targets.
• The performance metric: rank ordering, calibration or expected loss, measured on a named holdout sample and time window.
• The tolerance: how far below the baseline a candidate may fall and still meet the need, with the reasoning, approved by the model owner and an independent reviewer.

Fixing the bar first stops the search being tuned to a preferred answer, and makes a rejection explainable later (interpretation).

Step 2: Choose the disparity metrics and how group membership is estimated

New York says "insurers are encouraged to use multiple statistical metrics in evaluating data and model outputs to ensure a comprehensive understanding and assessment." Its list includes adverse impact ratios, denial odds ratios, marginal effects, standardized mean differences and significance tests, and "Drivers of Disparity", defined as "Identifying variables in AIS that cause differences in outcomes for protected classes relative to control groups."

On group membership, New York's Step 1 covers protected classes whose membership is available or "may be reasonably inferred using accepted statistical methodologies"; its guidance on quantitative assessment adds "There is no expectation that insurers collect additional data from, or about, individuals to perform exemplary analysis."

Pick a primary disparity metric that matches the decision (approval rates for accept or decline, price differences for pricing), plus one or two secondary metrics (interpretation). Document the inference method, its known error and how that error could bias the measured gap, and report intervals rather than single numbers (interpretation). Keep inferred labels out of the scoring path (interpretation).

Step 3: Generate candidates with feature, model, threshold and policy changes

New York's Step 3 refers to alternative "variable(s) or methodology"; the texts reviewed do not prescribe a search method. Options (interpretation):

• Feature substitution and removal: start from the drivers of disparity, then drop or replace each, singly and in groups, with a less correlated variable that carries the same risk signal, and retrain.
• Regularization and constrained training: penalize reliance on features that drive disparities, or add a disparity term to the training objective, and trace the trade-off curve. Methods that use protected-group data during training raise a separate disparate-treatment question; take legal advice first.
• Threshold and policy alternatives: cutoffs, tiers, segment-specific rules, exceptions and how inputs such as income are counted. The decision policy around a model can itself create or widen a disparity.
• Model multiplicity: there may be many models of near-equal accuracy. Search that set, varying seeds, hyperparameters, feature sets and algorithms, for the one with the smallest disparity.

Scale the search to the model's materiality and record what was tried, including dead ends (interpretation).

Step 4: Compare candidates against the bar on the same data

Score every candidate and the baseline on the same holdout sample, with the same performance and disparity metrics and the same inference method (interpretation). Plot performance against disparity; candidates that are worse on both can be set aside, and those inside the tolerance with a smaller disparity are the alternatives to consider (interpretation).

Check more than the headline: stability across time periods and segments, effects on other protected groups, calibration and operational fit (interpretation).

Step 5: Decide and document why an alternative was or was not adopted

For New York insurers: "If a less discriminatory alternative exists, the insurer should modify its use of ECDIS or AIS accordingly, and should evaluate the modified use of ECDIS or AIS beginning with Step 1." If none exists, "the insurer should conduct ongoing model risk management consistent with Section III of this guidance, and repeat Step 3 at least annually." New York also says "An insurer should appropriately document the processes and reasoning behind its testing methodologies and analysis for unfair or unlawful discrimination commensurate with the insurer's use of ECDIS and AIS and the complexity and materiality of such ECDIS and AIS." Colorado's Regulation 10-1-1 requires covered insurers to document quantitative testing "including the methodology, assumptions, results, and steps taken to address unfairly discriminatory outcomes".

A decision record should hold the bar and its approval, the metrics and inference method, every candidate with its results, the choice, who made it, and, for any candidate inside the tolerance with a smaller disparity that was not adopted, the specific reason (interpretation).

Step 6: Govern the search and keep the records

New York places the work inside the insurer's governance, with senior management responsible for "ensuring effective challenge and independent risk assessment". On vendors: "An insurer may not rely solely on a third-party's claim of non-discrimination or a proprietary third-party process to determine compliance with anti-discrimination laws."

Put the search in the model risk policy: which models are in scope and why, who runs it, who reviews it independently, who approves the decision, and how the records are retained and, with counsel, protected (interpretation). Ask vendors for the inputs and access needed to search, or for their own search results with enough detail to review (interpretation).

Step 7: Re-run the search when the model or its data changes

For New York insurers, "Unfair or unlawful discrimination testing and analysis should be administered prior to putting AIS into production and on a regular cadence thereafter, as well as whenever material updates or changes are made to either the ECDIS or AIS." Colorado's Regulation 10-1-1 requires covered insurers to keep a "Documented description of ongoing monitoring regarding the performance of algorithms and predictive models that use ECDIS including accounting for model drift."

Treat redevelopment, new or changed data sources, vendor version changes, policy and cutoff changes and drift in the disparity metric as triggers, alongside a scheduled cycle set by the model's tier (interpretation). Reuse the original bar unless the business need has changed, and record why if it has (interpretation).

Checklist

Items restate the recommendations above (interpretation).

• The legal position for each model is checked against the box above at each review
• Business objective, performance metric and tolerance are fixed and approved before candidates are built
• Disparity metrics and the group-inference method, with its error, are documented
• Protected characteristics are used for measurement only, with legal sign-off on any training use
• Candidates cover feature substitution, constrained training, threshold and policy changes and near-equal models
• All candidates are compared with the baseline on the same data and metrics
• Every decision is recorded, including the reason any less disparate candidate inside the tolerance was rejected
• The search sits in the model risk policy with independent review, vendor coverage and retention
• The search re-runs before production, on material change and on a schedule set by tier

How we sourced this guide: every factual statement in it is drawn from published rules, regulatory guidance and other public documents, and we keep a full set of sources and citations for it. If you would like them, or want to question a point, get in touch through our contact page and we will share them.