RESEARCH DOSSIER
Life Insurance Underwriting Fairness and Proxy Discrimination
Abstract. Life insurance underwriting transforms heterogeneous evidence about human mortality into decisions about eligibility, coverage amount, contractual terms, and price. The information that improves estimation can also reproduce unequal opportunity, encode protected characteristics through apparently neutral variables, and convert social disadvantage into differences in access to financial protection. This monograph investigates the intersection of actuarial classification, proxy discrimination, selection, and legal constraints in individual life insurance. It distinguishes statistical association, legitimate prediction, causal interpretation, prohibited use of information, unequal impact, and the separate question of what constitutes an equitable insurance allocation. It examines the history and contemporary architecture of medical and automated underwriting, mechanisms of direct and indirect discrimination, the relationship between mortality risk and price, and the relevant regulatory structures in the United States, the European Union, the United Kingdom, Canada, and Australia as of November 16, 2026. The methodological contribution is an integrated, source-grounded comparison of actuarial evidence, causal interpretation, decision-stage auditing, and institutional accountability. A transparent hypothetical cohort demonstrates why aggregate group disparities, conditional acceptance rates, and risk-standardized comparisons may yield different conclusions while still leaving questions of social causation and legal permissibility open. The central finding is that fair life underwriting requires the joint assessment of lawful information use, credible mortality evidence, proportionate decision rules, distributional consequences, explainability, and meaningful correction mechanisms. Predictive performance alone cannot resolve these distinct questions.
Keywords: life insurance; underwriting; actuarial fairness; proxy discrimination; risk classification; protected characteristics; mortality selection; algorithmic bias; disparate impact; genetic information; counterfactual fairness; insurance regulation; consumer protection; causal inference.
Introduction
A life insurer promises a specified benefit upon an insured death, generally conditional on the terms, duration, exclusions, and premium obligations of the contract. Individual underwriting determines whether the promise is offered and, if offered, on what terms. An applicant can therefore experience disadvantage through several distinct mechanisms: refusal of an application; a higher premium; a reduced amount of cover; a rider or exclusion; a requirement for additional medical evidence; diversion from an accelerated path into a longer examination; or an uncorrected error in an external record. These outcomes differ in economic magnitude, observability, and legal treatment.
The problem of fairness becomes especially complex when information is predictive and socially patterned. Age, smoking status, measured medical indicators, diagnosed disease, occupational exposures, credit-related variables, geography, purchasing behavior, and digitally inferred lifestyle indicators can have different relations to mortality and different permissions under law. A variable can be well correlated with mortality and still be unsuitable for a particular decision if it is unreliable, inaccessible to correction, impermissible in the jurisdiction, or an avoidable substitute for prohibited information. Conversely, a group-level association with a protected trait does not by itself establish an unlawful act or the presence of a causal proxy mechanism. Scientific assessment requires carefully specifying what was measured, which decision it affected, the population to which an estimate applies, and the applicable legal rule.
The contemporary problem extends beyond model-based pricing. Automated underwriting may change who reaches a human underwriter, which applicants receive a request for laboratory tests, how rapidly an offer is issued, and who abandons an application before any formal rejection. The NAIC account of accelerated underwriting emphasizes the role of alternative information and changing evidence requirements. The New York 2024 circular letter treats algorithmic systems, external consumer data, governance, adverse effects, and explanations as connected matters of market conduct.
This study adopts a documentary and analytical method. Legislation, supervisory publications, actuarial professional standards, and established scholarship are read according to their respective authority and effective dates. Formal expressions and the numerical example are explanatory constructions, not estimates from proprietary insurer portfolios. International comparison identifies functional similarities without assuming that a rule in one jurisdiction applies elsewhere. A source mentioned in the narrative is linked at the point where it supports the claim; full bibliographic records appear in References. The account is current to November 16, 2026.
The Scientific Object, History, and Documentary Foundations
Underwriting as a sequence of decisions
Life underwriting is usefully studied as a sequence rather than a single accept-or-decline outcome. Its stages include application solicitation and eligibility screening; collection of disclosures and third-party information; verification and medical evidence; estimation of the risk of death over an insured period; assignment to a risk class; decision on amount and price; communication of the result; and any review, correction, or appeal. A model used to decide who qualifies for accelerated review can influence access even when the final premium formula does not contain that model's score.
This distinction matters empirically. A study limited to issued policies excludes persons deterred by application design, rejected before full assessment, or redirected into a burdensome process. It measures outcomes among selected survivors of a decision pipeline. Such conditioning may create misleading estimates of comparative access and may distort measured relationships among risk factors, group membership, and price. Reliable analysis accordingly needs an applicant-flow denominator, not merely an in-force portfolio.
Traditional life underwriting relied on medical histories, examinations, occupations, age, family history, and evolving mortality experience. Industrialized life assurance and the growth of actuarial statistics made risk classification increasingly systematic. Modern predictive models add digital records, external scores, and multivariate interactions to those older evidence categories. The actuarial problem of insured selection remains recognizable: the mortality experience of accepted policies is a product of population risk, applicant behavior, eligibility criteria, and underwriting decisions rather than a simple random sample of the population.
The original economics of private insurance explains part of the pressure to classify. Rothschild and Stiglitz (1976), in their analysis of competitive insurance equilibrium, demonstrate why asymmetric information can change the availability and structure of insurance contracts. Their framework is a theory of market selection and incentives; it does not establish that every predictive input is lawful, reliable, or ethically justified. That latter inquiry requires separate evidence and normative and legal analysis.
What constitutes an underwriting fairness claim
At least five materially different claims appear in debates over underwriting. A descriptive claim states that a measurable disparity exists. A predictive claim states that an input improves out-of-sample estimation of relevant mortality. A causal claim states that an input or decision rule changes outcomes through a particular mechanism. A legal claim states that the decision complies with or violates an operative legal requirement. An evaluative claim states that the resulting distribution of access, price, and protection is justified. Their truth conditions are distinct.
For example, a neighborhood variable may improve predictive fit in historical data. The improvement alone does not show that neighborhood causes mortality, that its effect will persist, or that it may lawfully be used. The ability to correct an outdated address may be another independent issue. An insurer could meet a narrow predictive benchmark while failing a documentary test of data provenance or a regulatory requirement to consider a less discriminatory alternative.
The professional ASOP No. 12 on risk classification addresses actuarial design and evaluation across practice areas. It discusses the relationship between characteristics, experience, practicality, and the framework in which a classification system operates. Its guidance should be interpreted alongside, rather than in place of, binding legislation and supervisory rules. Whether causation is necessary for legal permissibility depends on the applicable rule; actuarial association and legal justification are different inquiries.
Evidence hierarchy and the limits of measurement
A primary legal source establishes what the text of an operative rule says at a particular date. Supervisory guidance communicates enforcement expectations within a regulator's remit. A professional standard governs relevant professional practice subject to its own adoption and applicability. Peer-reviewed scientific literature supplies general methods and empirical findings, frequently under assumptions that do not transfer directly to life underwriting. Insurer validation reports and application-level records, if available, can test actual deployment. The evidentiary significance of each source should be retained rather than collapsed into one category of authority.
Mortality evidence may come from national population statistics, insured-life experience studies, reinsurance and insurer portfolios, and clinical investigations. Differences in cohort age, calendar period, policy duration, country, underwriting standard, and socioeconomic composition affect comparisons. A public life table cannot by itself validate the incremental predictive value of a private digital underwriting score. Likewise, data from automobile or criminal justice prediction can illuminate mathematical fairness trade-offs without establishing the distribution of insured mortality.
The essential unit of analysis is therefore not the isolated predictor. It is a documented decision process linking specified information, a specified applicant population, an identified risk outcome, and a particular contractual consequence. This unit supports meaningful comparison across models, insurers, and legal systems.
Protected Characteristics, Proxy Variables, and Discrimination Mechanisms
Direct use, indirect pathways, and proxies
Direct discrimination in underwriting involves making a decision on a protected ground in circumstances prohibited by the applicable law. An indirect or disparate-impact concern arises when a formally neutral rule produces an adverse group pattern and the governing legal system subjects that pattern to further examination. Proxy discrimination is the use of another characteristic or combination of characteristics that stands in for a protected trait, potentially reproducing a prohibited distinction. The concepts overlap, but their factual predicates and legal consequences are not identical.
Prince and Schwarcz (2020), in Proxy Discrimination in the Age of Artificial Intelligence and Big Data, explain how high-dimensional predictive systems can recover protected information through other inputs and why existing antidiscrimination frameworks face difficulties with such practices. Their account is particularly useful for distinguishing an input's predictive utility from the possibility that its utility arises through protected-class information. It is a legal and conceptual analysis, not a finding that every insurance score is an unlawful proxy.
Four examples illustrate different mechanisms. A postal code may correlate with income, environmental exposure, care access, and racialized residential patterns. A credit-related variable may encode financial distress and past exclusion from credit markets. A purchasing or web-browsing profile may reveal socioeconomic status or health-related interests. A family-history response may provide clinical information while also implicating relatives' privacy and legally regulated genetic information. Each variable requires its own provenance and causal assessment; a single label of “alternative data” conceals significant differences.
A proxy assessment starts with a target variable P representing a protected characteristic and an observed input X. Detecting association between P and X is a warning signal, not a full explanation. If X is associated with mortality Y through a legitimate exposure pathway, or through health-care deprivation, or through a data-generation artifact, the legal and ethical questions differ. An additional question is whether P remains predictable from the combination of permitted inputs after the most obvious proxy has been removed. Complex algorithms can reconstruct an excluded characteristic from several weak correlates.
Measurement and causal pathways
A useful causal description distinguishes protected status P, background conditions S, clinical risk C, observed proxy X, underwriting decision D, and mortality outcome Y. Possible pathways include P to S to C to Y, S to X, C to X, and X to D. These arrows are hypotheses about mechanisms, not empirical facts supplied by the diagram. The same observed correlation between X and P may result from several structures, and the mechanism must be supported by domain knowledge, temporal order, and data.
Counterfactual analysis asks whether a decision would differ under a hypothetical change to a protected characteristic while holding appropriately specified background conditions fixed. Kusner and colleagues (2017) formalize an influential counterfactual fairness framework. In insurance, such analysis depends on controversial and empirical choices about which descendants of protected status may appropriately vary. Changing a variable such as geographic residence while holding the social determinants of residence fixed can describe a mathematically convenient intervention with uncertain real-world interpretation. Causal assumptions should be explicit.
There is also a temporal problem. Historical underwriting decisions can shape future data. If access to cover, pricing, or medical examinations differs by group, insured-life outcomes and medical disclosure patterns may be recorded unevenly. Training only on issued policies means the model learns from a selected population produced by past decisions. Apparent improvements in discrimination or calibration after retraining may be consequences of who remained observable.
Differential access through data quality and application design
Disparate impact can arise without an explicit risk-rating differential. Digital identity verification may work less reliably for applicants whose records are sparse or inconsistent. Automated systems may refer certain applicants for medical testing more often. Longer decision times can increase noncompletion, particularly for persons who cannot easily attend examinations or obtain records. Documentation that records only final declinations will overlook these mechanisms.
Data accuracy can also vary across groups. A commercial file may contain outdated address information, mistaken household association, or inferred behavioral information that an applicant has never supplied. The resulting score can appear precise while resting on weak or inapplicable evidence. A fair review therefore addresses provenance, permission, completeness, missingness, corrections, and model performance together.
Barocas, Hardt, and Narayanan (2023), in Fairness and Machine Learning, examine the limits of purely technical fairness remedies and the importance of institutional context. In life underwriting, that context includes the terms of a mortality-contingent contract and the ability of applicants to discover and contest an adverse decision. A model audit detached from the distribution process is incomplete.
Genetic information as a distinct boundary
Predictive genetic test results, diagnosed conditions, family history, and medical symptoms require separate treatment. They may have different predictive properties, different privacy implications, and different legal permissions. The United States National Human Genome Research Institute guidance explains that the federal Genetic Information Nondiscrimination Act of 2008 restricts specified uses in health insurance and employment but does not extend the same protections to life, disability, or long-term care insurance. Other federal and state protections may apply; GINA must not be described as a blanket national prohibition on life underwriting uses.
Genetic law also changes over time. Under the United Kingdom's 2026 code, signatory insurers ordinarily do not use predictive genetic test results, with a specified exception for Huntington's disease above a life-insurance aggregate limit. Australia enacted a statutory prohibition, effective October 8, 2026, against soliciting or using protected genetic testing information for specified life-insurance decisions. Treating these rules as interchangeable would conceal both their legal force and the distinctions among genetic tests, diagnoses, and coverage products.
Actuarial Justification, Selection, and the Economics of Classification
Mortality prediction and the insured event
For an insured person i, let q_i(t) denote the probability of death over a specified interval t under a defined underwriting state, age, policy duration, and observation population. For a simple one-period benefit B paid on death, an undiscounted expected claim cost may be written as B multiplied by q_i. Real pricing adds discounting, expenses, capital costs, policy options, profit and risk margins, competing events, and uncertainty. In particular, a fitted probability is not a complete premium and a group mortality contrast is not a justification for every underwriting decision.
The actuarial objective must be specified. Calibration tests whether predicted death probabilities agree with observed experience in a defined population and time horizon. Discrimination tests whether a score separates higher- from lower-risk cases. Stability tests sensitivity to sample, model specification, and changing conditions. Decision utility asks whether changes in rating or referral are proportionate to credible risk differences. Fairness adds questions about variable legality, selection, group impact, alternatives, and contestability.
Risk classification can support sustainable pooling by aligning expected liabilities and premiums, but the process may simultaneously determine which risks enter the pool. A high-risk applicant who cannot purchase ordinary coverage may seek guaranteed-issue insurance, accept waiting periods, reduce the amount insured, or remain uninsured. Therefore, the relevant policyholder outcome is often a sequence of coverage and welfare changes rather than the offered premium alone.
Statistical association and causal justification
A predictor may have substantial out-of-sample predictive power because it measures exposure, captures a downstream manifestation of disadvantage, or reflects how a dataset was assembled. These explanations cannot be distinguished from the coefficient of the predictor alone. Actuarial justification should therefore disclose at least the endpoint, time horizon, cohort, amount of incremental value, reliability, stability, and lawful permissible purpose.
Consider an external lifestyle score that predicts mortality after controlling for age and basic health variables. The value of the score for a decision depends on whether its improvement persists in independent insured cohorts; whether missing scores receive unequal treatment; whether the score reflects financial, demographic, or clinical pathways; and whether the insurer can explain and correct the underlying information. An impressive area under a receiver operating characteristic curve is insufficient evidence of a justifiable price or eligibility differential.
This analysis separates three meanings of “similarly situated.” Applicants may be similarly situated on observed clinical measurements, on validated expected mortality, or under a legally prescribed comparison. These populations need not coincide. If two applicants have equal valid estimated mortality but receive different access because of a nonclinical digital proxy, the differential demands particular scrutiny. If estimated mortality differs partly because of unequal environmental exposures, an actuarial difference may remain even when the equity and legal questions are unsettled.
Classification, solidarity, and selection
Uniform premiums across groups can produce cross-subsidization when expected risks differ. More granular pricing can reduce cross-subsidies, but may also render protection unaffordable to people with elevated risks that originate partly in social inequality. The efficiency and equity effects depend on product design, mandated pooling, substitution options, public protection, and the behavior of consumers and insurers. Neither perfect individualized pricing nor universal pooling is an empirically inevitable outcome of life insurance science.
The Rothschild–Stiglitz selection problem helps explain why unregulated competition may not yield a simple welfare optimum. The presence of asymmetric information can affect which contracts survive competition and which applicants purchase them. However, the welfare effect of prohibiting an input cannot be inferred mechanically from its predictive correlation. The relevant assessment considers the possible reallocation of premiums, changes in take-up, the availability of substitute evidence, and the possibility of exclusion or misclassification.
A classification rule can also generate a dynamic feedback effect. Suppose an insurer removes a proxy but introduces a burdensome laboratory requirement for applicants who previously would have completed an accelerated path. The nominal model becomes less reliant on the proxy, yet access may decline among the people most affected by the referral burden. Evaluating the fairness of the final premium without recording the added requirement would miss the substitution effect.
Legal permission as a condition of actuarial use
Actuarial credibility, by itself, is not a license to use any variable. In the United States, state laws and market-conduct rules vary; New York's 2019 life underwriting guidance requires attention to the actuarial validity and potential discrimination of external data sources. Its later 2024 guidance sets out documented steps to evaluate disproportionate adverse effects, explanations, and less discriminatory alternatives. This is an example of a regulator requiring evidence that extends beyond conventional aggregate prediction quality.
The relationship between legal constraints and actuarial method is reciprocal. Law specifies permissible information and required treatment; within that feasible domain, actuarial methods estimate uncertainty and the consequences of alternative rules. A scientifically credible policy must therefore state the admissible decision space before optimizing prediction. Attempting to repair an impermissible variable by merely improving model fit confuses a constraint with a performance objective.
Comparative Law, Supervisory Standards, and Institutional Governance
United States: state supervision and federal boundaries
Insurance underwriting is regulated primarily by the states, subject to other applicable law. A nationwide model bulletin or professional paper does not automatically impose identical binding requirements in every jurisdiction. State-specific legal and supervisory analysis remains necessary for any particular product.
New York provides an unusually explicit sequence of regulatory instruments. The 2019 Circular Letter No. 1 addresses external data and algorithms in life insurance underwriting and emphasizes actuarial relevance, unlawful discrimination, and disclosure of reasons. Circular Letter No. 7 (2024) expands supervision of artificial intelligence systems and external consumer information across insurance business. Its three-stage analysis moves from identified adverse effects to legitimate explanations and a search for less discriminatory alternatives. It clarifies that quantitative group testing applies where protected-class data are available or can reasonably be imputed and that insurers are not expected to collect additional personal data solely for those tests. It also addresses third-party controls, record retention, complaints, explanations, and regular model monitoring. These are supervisory expectations under cited New York law, not a universal international fairness formula.
Colorado illustrates a related but independent state approach. Its 2025 regulatory update identifies amended Regulation 10-1-1 governing the use of external consumer data, algorithms, and predictive models by specified insurers, with an effective date of October 15, 2025. The applicable lines and details must be checked against the operative Colorado rule before treating a provision as specific to a particular life product. State variation can change both documentation obligations and the legal significance of an apparently neutral characteristic.
Federal genetics law presents a separate boundary. GINA provides specified protections in health insurance and employment; its omission of life underwriting from those protections does not remove other applicable state rules or authorize all uses of genetic information. A comparative account must identify the actual legal instrument governing the decision instead of treating a single federal statute as the entire United States insurance rulebook.
European Union: equality, data protection, and staged AI regulation
European life underwriting is influenced by several legal layers. The General Data Protection Regulation governs the processing of personal data, with particularly stringent conditions for health and genetic data under Article 9 and safeguards concerning solely automated individual decisions under Article 22 where its conditions are met. Article 22 is not a categorical ban on every algorithm used in underwriting: its application depends on the decision's legal or similarly significant effect, whether it is based solely on automated processing, the permitted exceptions, and accompanying safeguards. A human signature on a workflow is inadequate evidence of meaningful human involvement if review is purely formal.
The Court of Justice's Test-Achats ruling in Case C-236/09 led to the end of using sex as a factor producing differences in individual insurance premiums and benefits for new contracts within the affected EU framework from December 21, 2012. The Court's Test-Achats decision record is the controlling starting point for the historical legal change. Its significance is analytically important: even a statistically predictive classification can be restricted by equality law.
The AI Act, consolidated July 2026, identifies AI systems intended for risk assessment and pricing in life and health insurance for natural persons as high-risk in Annex III. Its temporal application is staggered. Under the July 2026 consolidated Article 113, the principal Annex III high-risk obligations apply from December 2, 2027; accordingly they must not be represented as fully applicable in November 2026. Other provisions have different start dates, and existing data protection and nondiscrimination duties retain their own legal force.
The EIOPA 2025 Opinion on AI Governance discusses sectoral supervisory expectations concerning governance, proportionality, explainability, data management, and fairness. Its scope and legal status differ from a directly applicable legislative provision; it should be read alongside insurance regulation and the AI Act without merging them into one obligation.
United Kingdom: statutory distinctions and an industry genetics code
The United Kingdom's Equality Act 2010, Schedule 3, contains insurance-related exceptions subject to their conditions. The current text maintains a disability-related insurance provision tied to relevant, reliable information and reasonableness; it also contains provisions concerning age in relevant circumstances. The original Schedule 3 paragraph 22 exception relating to sex was omitted effective December 21, 2012. Citing the enacted 2010 version as though that deleted exception were still current would produce a materially false legal conclusion.
The Code on Genetic Testing and Insurance, updated in March 2026, is a commitment of participating insurers rather than a general statutory insurance ban. Its predictive-testing exception currently concerns Huntington's disease where the total applied-for life cover exceeds £500,000. The code distinguishes predictive and diagnostic testing and contains other specific financial limits for critical illness and income protection. A faithful analysis identifies the signatory-based character, the test covered, the threshold, and the insurance line.
This comparison also clarifies the use of “actuarial justification” in law. A legal system may allow some differentiated treatment when evidence is reliable and the response reasonable while absolutely restricting other protected characteristics. The same estimated mortality association may thus be treated differently because the relevant legal permission differs.
Canada: statutory genetic nondiscrimination
Canada's Genetic Non-Discrimination Act, enacted in 2017, prohibits requiring an individual to undergo a genetic test as a condition of providing goods or services or entering or continuing a contract. Other provisions regulate disclosure of genetic-test results and consent, subject to statutory details and exceptions. The law's protected object is specifically genetic testing and its results; it should not be casually equated with every medical or family-history question.
This legal architecture places an information-acquisition constraint before the question of predictive validity. A test result with real actuarial signal may remain inaccessible to the underwriter when use would conflict with the statute. For cross-country modelling, that means the feature space itself differs according to law. Empirical models estimated in one jurisdiction cannot be transferred uncritically to another with different permissible predictors.
Australia: an October 2026 statutory transition
Australia's Insurance Contracts Act 1984, current compilation incorporates amendments effective October 8, 2026, governing protected genetic testing information and life insurance. The relevant Division 5 provisions define genetic testing and constrain solicitation and use of that information for covered decisions. Genomics Australia's October 2026 explanation states that the prohibition applies to new policies beginning, or existing policies varied, on or after that date, and covers relevant forms of life insurance including death, disability, trauma, and income protection; it does not extend to health insurance as a product category in that account.
This is a significant chronological distinction. Analyses written before October 2026 may describe an industry moratorium or a planned legislative transition, whereas a November 2026 comparison must analyze the operative statutory provisions. An evaluation that pools applications before and after the change should label the legal regimes separately and avoid attributing all observed trends to the statutory change.
International governance and operational responsibility
The IAIS Insurance Core Principles establish an international supervisory framework in which fair treatment of customers, governance, risk management, and conduct duties are relevant to underwriting. They are supervisory benchmarks adopted and implemented through national systems, not a single international insurance statute.
Operational responsibility spans the full decision chain. Data suppliers should document definitions, provenance, permitted use, quality, and correction channels. Model developers should document training populations, variables, performance, failure modes, and revalidation. Underwriters should explain how scores affect actual eligibility, amount, referral, and price. Compliance and legal teams should map permitted inputs and customer rights to jurisdictions. Actuarial functions should test mortality validity, credibility, and uncertainty. Boards and senior management should review material residual risks; complaints and independent audit should supply evidence that controls work in practice.
The NIST AI Risk Management Framework provides a voluntary, cross-sector organizing reference for governance, mapping, measuring, and management of AI risk. In insurance it is most useful as a disciplined process framework, not as a substitute for statutory compliance. The significance of governance is demonstrated by measurable records: who approved a variable, when a performance drift was detected, which applicants were adversely affected, and what corrections followed.
Empirical Testing, Sensitivity Analysis, and Model Validation
Choosing the denominator and outcome
A useful fairness audit begins by defining the population and outcome before selecting a metric. The population may be all visitors who express interest, all completed applications, all applications reaching underwriting, or all individuals offered contracts. The outcome might be acceptance, price per unit of comparable coverage, ability to obtain a requested sum assured, time to decision, medical-referral rate, or lapse and abandonment. A result estimated among issued contracts cannot be interpreted as the access rate among all potential applicants.
For a defined group g, let A_g be the number of applicants offered the requested contract and N_g the number who entered the specified underwriting stage. The observed acceptance rate is A_g divided by N_g. A group difference is the subtraction of rates; an acceptance ratio divides one group rate by another. These simple numbers are sensitive to product mix, data exclusions, missing protected-group identifiers, and application pathways. A raw group ratio should be accompanied by denominators, uncertainty, and the comparison population.
Protected-group data may be missing because they were never collected, are prohibited for operational use, or are unreliable. Proxy-based or imputed group membership can help a narrowly authorized audit, but introduces classification error and privacy concerns of its own. Report sensitivity to plausible imputation error and distinguish audit-only use from underwriting decisions.
A transparent hypothetical application cohort
Consider a purely illustrative cohort of 2,000 completed applications: 1,000 from Group A and 1,000 from Group B. Assume that every applicant is classified into one of two validated risk strata, Low and High, using legally permissible evidence. The underwriting rule offers a contract to 90% of Low applicants and 40% of High applicants in each group. These numbers are constructed for demonstration and are not empirical estimates of life-insurer behavior.
| Group | Low applicants | Low accepted | High applicants | High accepted | Total accepted | Overall acceptance |
|---|---|---|---|---|---|---|
| A | 800 | 720 | 200 | 80 | 800 | 80% |
| B | 400 | 360 | 600 | 240 | 600 | 60% |
| Combined | 1,200 | 1,080 | 800 | 320 | 1,400 | 70% |
The crude acceptance difference is 20 percentage points and B's acceptance ratio relative to A is 0.75. Yet the conditional acceptance rates are the same within each displayed stratum: 90% for Low and 40% for High. If both groups are standardized to an identical reference mix of 50% Low and 50% High, each standardized acceptance rate becomes 0.5 × 0.90 + 0.5 × 0.40 = 0.65, or 65%.
This arithmetic has a narrow interpretation. The aggregate disparity in this constructed dataset is fully explained by the stipulated differences in risk-stratum composition; no within-stratum difference was built into the rule. It does not prove the rule lawful or fair. The stratifying variable might itself be an unlawful proxy, its measurement could be unequal, the risk tiers might have limited mortality validity, and the distribution of risk could reflect historical inequities. The appropriate next analysis examines how the tiers were constructed and whether a less discriminatory, legally permissible alternative preserves sufficient risk control.
An alternate rule illustrates sensitivity. If the High-stratum acceptance rate for Group B rose from 40% to 50%, while all other figures stayed fixed, 300 rather than 240 of its High applicants would be accepted, and the B overall acceptance rate would rise from 60% to 66%. The crude A–B gap would decrease to 14 percentage points. This change, however, would introduce a within-stratum group differential unless the rule was based on a legally relevant additional factor. A smaller aggregate gap is therefore not sufficient evidence of procedural or legal fairness.
Prediction, calibration, and classification metrics
Calibration concerns the correspondence of estimated mortality probabilities to subsequently observed deaths. For predicted probability q and observed death indicator Y, one can compare mean predictions with the observed death rate within carefully defined score bands and time horizons. In a life portfolio, censoring, duration, policy termination, and changes in health or risk exposure complicate this exercise. Backtesting must account for how long an applicant was actually at risk and which deaths were observable.
An aggregate model can be calibrated overall and miscalibrated across groups, ages, policy durations, or high-risk tails. A model may also be calibrated within groups while producing unequal error rates under a given classification threshold. Chouldechova (2017), studying a different prediction domain, documents the mathematical tension among predictive parity and error-rate fairness. Kleinberg, Mullainathan, and Raghavan (2017) analyze related inherent trade-offs in risk scoring. These results clarify mathematical constraints; they cannot be transplanted as empirical findings about life insurance applicants.
Fairness metrics should be chosen according to the actual decision's purpose and harm. Equal acceptance rates may be relevant to access but insensitive to price differences among accepted applicants. Equal false-positive rates may be relevant when high-risk classification triggers a costly or invasive procedure. Calibration within groups may be relevant to the reliability of estimated mortality. Counterfactual checks can expose proxy pathways but depend on contestable causal assumptions. No single metric discharges every legal, actuarial, and consumer-protection duty.
Testing the incremental value of external data
A credible validation design compares a documented baseline underwriting model with a candidate augmented model using the same independent outcome period. It reports incremental discrimination and calibration, uncertainty intervals, missingness, and group-specific performance. Features should be linked to their actual source, collection date, update frequency, and permitted purpose. Apparent improvement from a data source that was updated after the underwriting decision can be temporal leakage rather than genuine predictability.
Temporal validation is central. Randomly splitting records from the same historical years can overstate performance when policyholders, families, vendors, or changing economic conditions connect observations. A stronger design reserves later application cohorts and, where possible, independent insurers or regions for testing. The analysis should distinguish mortality observation for issued policies from the unknown mortality that would have occurred for rejected applicants, because selective outcome labels constrain inference.
Practical uncertainty checks include alternate missing-data models, plausible error in group imputation, alternative mortality horizons, model recalibration, small-subgroup confidence intervals, regional transport tests, and drift in data availability. For a highly sparse subgroup, a large percentage disparity can arise from a few applications; the report should disclose numerators and denominators without dismissing a consistent pattern of small harms.
Distinguishing statistical tests from causal evidence
A descriptive difference-in-rates test establishes statistical evidence about a difference under assumptions about sampling and dependence. A regression conditional on measured variables describes conditional associations. Neither establishes how the insurer's rule caused an observed social disparity. Causal identification additionally requires a credible intervention or natural experiment, adequate measurement of confounders, and defensible assumptions about selection into application and policy issuance.
A policy change, such as the introduction or withdrawal of a digital score, may provide an opportunity for before-and-after or difference-in-differences analysis. Such designs need comparable cohorts, a credible comparison group, stable measurement, and scrutiny of simultaneous product or market changes. A regression discontinuity around a documented referral threshold can be informative near that threshold but may have little external validity elsewhere. A controlled rollout, if lawful and ethically designed, can offer stronger evidence of process effects while still requiring monitoring for customer harm.
The Australia October 2026 genetic testing transition is an example of a legally identifiable intervention date that may enable future study. It would be premature to claim a measured effect on access, premiums, or adverse selection without observed post-change data, suitable controls, and a specified time horizon. A change in law establishes a research opportunity; it does not by itself establish the empirical result.
Evidence-to-decision and redress testing
An operational audit should follow each tested adverse outcome back to its data and decision rule. Can the insurer identify which input was material? Is the underlying record accurate and currently applicable? Can the applicant inspect or correct it under the relevant law? Was there a meaningful route to human reconsideration? Did a third-party vendor cooperate? Was the decision altered after correction? These questions are verifiable through case logs and sampled files.
The New York 2024 guidance requires attention to written reasons for specified accelerated-underwriting referrals and a process to review the accuracy of external data in relevant circumstances. It is therefore insufficient to document a fairness score while giving affected applicants an uninformative response. Consumer recourse supplies a test of whether explainability and accountability have practical effect.
A mature evaluation reports both predictive outputs and actual customer consequences: proportion referred for further evidence, completion after referral, median decision time, proportion of adverse decisions corrected, number and nature of complaints, and changes in cover purchased. These measures should be segmented by relevant populations where lawfully possible and should record the costs and benefits of alternative underwriting rules.
Integrated Analytical Synthesis and Research Priorities
An integrated interpretation of fairness
The sources support a coherent interpretation: fairness in life underwriting is a property of a governed decision process, evaluated against several independently justified requirements. Risk classification must be based on credible evidence about an appropriate insured event; input collection and use must comply with law; decisions must be tested for relevant group effects; the distribution of access must be observed over the application pipeline; and material errors must be open to correction. These findings synthesize existing actuarial, legal, and fairness literature. They do not assert a newly established theory or proprietary formal model.
The central tension is created by the relationship between individualized risk estimation and collective protection. Life insurance transfers mortality-contingent financial loss through a portfolio whose composition is altered by the very underwriting rules used to estimate it. Classification can reduce informational asymmetry while encoding historical inequalities and affecting access to protection. This makes the selection mechanism itself a necessary object of fairness research.
An insurer that measures only mortality accuracy may miss unjustified eligibility exclusions. An auditor that measures only demographic parity may miss material differences in legitimate insured risk and the economic effects of adverse selection. A regulator that examines only explicit protected features may miss reconstructed proxies. A model developer that tests only issued policies may miss applicants excluded before the outcome was observable. Each failure arises from omitting a distinct part of the same underwriting system.
A defensible sequence of scientific questions
A rigorous study can proceed in an evidentiary sequence without imposing a universal fairness formula. First identify the contract, applicant population, coverage period, decision stage, and governing jurisdiction. Then document each variable's meaning, source, intended use, lawful permission, and quality. Assess its incremental predictive value against mortality or the specific risk endpoint, including uncertainty and transportability. Measure disparities in eligibility, price, amount, speed, and referral with explicit denominators. Investigate plausible proxy pathways and alternative explanations. Finally, compare feasible substitute variables, decision thresholds, or product designs and document their effects on both actuarial performance and customer outcomes.
The sequence gives special weight to variable exclusion when use is prohibited by law. A forbidden data item is removed because the applicable rule controls, not because its predictive value is low. Conversely, a legally permissible and predictive feature may still demand correction or replacement if it is measured unreliably, creates avoidable adverse effects, or fails an operative reasonableness or necessity requirement.
Decision records should preserve the distinction between scientific findings and governance choices. A finding that a score is miscalibrated in a defined cohort is empirical. A decision to discontinue the score is managerial or regulatory. A finding that an alternative rule reduces disparity without materially degrading mortality discrimination is comparative. An assertion that a particular balance of fairness and selection is justified requires the relevant legal and institutional rationale. Transparent separation permits future reassessment when evidence, law, or social conditions change.
Cross-border comparability and legal heterogeneity
International comparison establishes that genetic information is a particularly clear test of the limits of actuarial reasoning. US federal GINA, Canada's statutory regime, the UK's signatory code, and Australia's October 2026 legislation do not impose the same rules. The EU adds its own combination of data protection, equality constraints, and staged AI regulation. A pooled global model trained under one information environment may therefore contain variables unavailable or unsuitable elsewhere.
This heterogeneity requires more than changing a country code in a model. Researchers should record the date of decision, underwriting location, contract type, availability of protected-group information for audit, applicable legal permissions, model version, and whether an input came from a vendor. Comparable fairness indicators must use consistent definitions and explicitly identify exceptions. Global averages that erase legal and product differences risk suggesting a precision they do not possess.
The historical evolution of the legal frameworks also changes longitudinal inference. EU sex-based premium restrictions, New York's expansion of external-data oversight, and Australia's new genetic provisions each mark different interventions with distinct populations and dates. Their consequences should be investigated empirically rather than assumed from the mere existence of a rule.
Priority empirical questions
Several research questions follow directly from the identified evidence gaps. What are the causal effects of accelerated underwriting on completion and access by group, beyond issued-policy mortality? How accurately do external consumer features predict future mortality out of sample after ordinary medical variables are included? Can a reduced feature set preserve credible mortality assessment while lowering adverse effects and improving explainability? Which forms of proxy imputation are sufficiently accurate for lawful audit use without introducing new privacy harms? How do applicants respond to referral delays, explanatory notices, and opportunities to correct a vendor record? What changes in insurance take-up and coverage adequacy follow restrictions on genetic-test information?
Each question requires data not supplied by this documentary study. Answers should be sought through carefully designed cohorts, transparent decision logs, legally permissible subgroup analysis, model transport studies, and credible natural experiments. A single insurer's experience cannot automatically establish an international regularity.
Limits of the completed investigation
This monograph uses statutes, supervisory texts, professional material, and published scholarship. It did not obtain proprietary underwriting files, individual genetic records, vendor scores, insurer-level application funnels, or mortality microdata. The hypothetical cohort is an arithmetic demonstration. It cannot support prevalence estimates of discrimination or insurer-specific causal claims. Some academic materials are accessible through publisher abstracts and bibliographic records rather than full licensed text; the claims drawn from them are correspondingly limited to established methodological or bibliographic points.
Legal comparisons are snapshots as of November 16, 2026, with exact operative distinctions where verified. Later amendments, national implementation of international principles, sector-specific exceptions, policy wording, and court interpretations can change the result for a specific decision. These limits delimit the evidence while preserving the substantive conclusions supported by the sources examined.
Conclusion
Life insurance underwriting fairness depends on the interaction of mortality evidence, access to financial protection, permitted information, and the governance of decision processes. Proxy discrimination is scientifically significant because apparently neutral inputs can recover protected information, encode unequal social conditions, and alter outcomes at several points before a policy is issued. Its detection requires analysis of the variables, the applicant population, the decision pathway, and the relevant legal definition; a correlation alone is an incomplete diagnosis.
Actuarial validity and fairness should be assessed together through explicitly separated questions. Credible mortality estimation establishes what a predictor contributes to risk assessment. Law establishes what information and distinctions may be used. Group comparisons show where outcomes diverge. Causal inquiry identifies plausible mechanisms. Alternative-rule analysis shows whether the decision can achieve its legitimate insurance purpose with fewer adverse consequences. Explanation and recourse establish whether an affected applicant can understand and correct the result.
The comparative evidence demonstrates enduring jurisdictional differences. New York's insurance guidance specifies proxy and adverse-effects review, European law combines data protection and equality requirements with a staggered AI Act regime, the United Kingdom maintains a distinct statutory and code-based structure, Canada regulates genetic testing through federal legislation, and Australia has implemented a new statutory prohibition effective in October 2026. These differences are part of the scientific object of study and must remain visible in any international model or historical analysis.
The resulting conclusion is operationally testable: a defensible underwriting process makes its risk evidence, permitted inputs, decision effects, uncertainty, alternatives, and correction procedures observable. The quality of that evidence determines how confidently fairness can be assessed. Sound insurance science therefore evaluates the whole underwriting pathway and the actual protection available to applicants, while preserving the distinct legal and actuarial grounds of each conclusion.
References
American Academy of Actuaries / Actuarial Standards Board. (2011, current standard listing). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). Official ASOP No. 12.
Australian Government, Federal Register of Legislation. (1984, compilation current October 8, 2026). Insurance Contracts Act 1984, including Division 5, Use of Protected Genetic Information in Relation to Contracts of Life Insurance. Federal Register of Legislation.
Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press. Open book edition.
Canada. (2017, consolidated text). Genetic Non-Discrimination Act, S.C. 2017, c. 3, section 3. Justice Laws Website.
Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. Publisher DOI.
Colorado Division of Insurance. (2025, August 20). Amended Regulation 10-1-1 effective October 15, 2025: External consumer data and information sources, algorithms, and predictive models. Official announcement.
Court of Justice of the European Union. (2011, March 1). Association belge des Consommateurs Test-Achats ASBL and Others v Conseil des ministres, Case C-236/09. CJEU case record.
European Insurance and Occupational Pensions Authority. (2025, August 6). Opinion on Artificial Intelligence Governance and Risk Management. EIOPA publication notice.
European Parliament and Council. (2016). Regulation (EU) 2016/679: General Data Protection Regulation, particularly Articles 9 and 22. EUR-Lex GDPR text.
European Parliament and Council. (2024, consolidated July 27, 2026). Regulation (EU) 2024/1689: Artificial Intelligence Act, including Annex III and Article 113. Consolidated AI Act.
Genomics Australia. (2026, October 1). Genetic Testing and Life Insurance: New Statutory Protections from October 8, 2026. Government explanatory guidance.
International Association of Insurance Supervisors. (2024, revised framework). Insurance Core Principles and Common Framework, particularly ICP 19 on conduct of business. IAIS online tool.
Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. 8th Innovations in Theoretical Computer Science Conference, Article 43. Dagstuhl proceedings.
Kusner, M. J., Loftus, J., Russell, C., & Silva, R. (2017). Counterfactual fairness. Advances in Neural Information Processing Systems, 30. NeurIPS proceedings.
National Association of Insurance Commissioners. (2026, topic updated). Accelerated Underwriting. NAIC topic resource.
National Human Genome Research Institute. (current public guidance). Genetic Discrimination: Scope of the Genetic Information Nondiscrimination Act. NHGRI guidance.
New York State Department of Financial Services. (2019, January 18). Insurance Circular Letter No. 1: Use of External Consumer Data and Information Sources in Underwriting for Life Insurance. DFS Circular Letter No. 1.
New York State Department of Financial Services. (2024, July 11). Insurance Circular Letter No. 7: Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing. DFS Circular Letter No. 7.
Prince, A. E. R., & Schwarcz, D. (2020). Proxy discrimination in the age of artificial intelligence and big data. Iowa Law Review, 105, 1257–1318. Iowa Law Review article.
Rothschild, M., & Stiglitz, J. E. (1976). Equilibrium in competitive insurance markets: An essay on the economics of imperfect information. Quarterly Journal of Economics, 90(4), 629–649. Journal article.
Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. National Institute of Standards and Technology. NIST AI RMF 1.0.
United Kingdom. (2010, as amended through 2026). Equality Act 2010, Schedule 3, Part 5 (including insurance exceptions and omission of paragraph 22). Current Schedule 3.
United Kingdom Department of Health and Social Care and Association of British Insurers. (2026, March). Code on Genetic Testing and Insurance. UK government code.