Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Underwriting Missing Data and Applicant Nonresponse

Abstract. Underwriting is a decision process conducted on information of unequal completeness, reliability, timeliness, and legal usability. An unanswered application question, a missing laboratory result, an unsuccessful external data search, and a customer who withdraws before verification are distinct events. Each changes the observable information set, but each has a different implication for statistical identification, risk assessment, contractual disclosure, and access to insurance. This monograph investigates how missing applicant information arises, when conventional risk scoring becomes unreliable, how insurers can distinguish uncertainty from evidence of elevated risk, and under what conditions substitute evidence, referral, or an alternative underwriting route can support a defensible decision. It integrates missing-data methodology, insurance economics, behavioral research, actuarial professional standards, underwriting monitoring, and selected comparative legal and supervisory sources. The analysis derives identification bounds, a decision-threshold sensitivity calculation, a value-of-information example, and a stratified verification audit, all using explicitly synthetic data. It establishes that the predictive value of a missingness indicator, the recoverability of the value that is missing, and the permissibility of acting on either are separate questions. The resulting synthesis situates underwriting data policies within a joint assessment of financial risk, evidentiary quality, decision uncertainty, operational burden, and lawful access to coverage. No proprietary insurer-level dataset or new general theory is asserted.

Keywords: Insurance underwriting; missing data; applicant nonresponse; selective disclosure; missing not at random; identification; risk classification; accelerated underwriting; verification; fairness; insurance access; actuarial governance.

Introduction

Underwriting converts a prospective exposure into a decision about eligibility, price, coverage, terms, or additional investigation. The conversion requires information about the exposure and an assessment of the uncertainty that remains after the information has been obtained. The object of this investigation is the second component: incomplete applicant information and the process through which incompleteness affects the underwriting decision. It belongs to the Science of Underwriting and intersects with actuarial modeling, insurance economics, risk management, and insurance law. The inquiry concerns applications and related evidence available before a binding decision, including pre-issue referral. Subsequent claims observations are considered insofar as they provide selective feedback about those decisions.

An application record is a historical artifact rather than a transparent census of the exposure. Its contents reflect which questions were asked, which were applicable, what the applicant understood, what information was accessible, when the system received it, which records it linked, and which facts the insurer was permitted to use. A blank field cannot reveal its own cause. Its interpretation depends on the information-generating process and the definition of the underlying characteristic. A missing property inspection, for example, can reflect a delay in scheduling rather than the physical state of the property. An absent prescription history may reflect a failed match or limited coverage of a data supplier; it does not establish that no prescriptions exist. A skipped question may be logically inapplicable rather than unanswered.

The scientific problem has three levels. The first is descriptive: what is missing, for whom, at which stage, and because of which recorded event? The second is inferential: under which assumptions can the analyst estimate the quantity relevant to the decision? The third is institutional: which decisions and evidentiary uses are lawful, operationally proportionate, and consistent with authorized underwriting policy? A model may satisfy a predictive objective on observed records yet fail when used with all applicants or after a change in the application process. A legally restricted variable may be highly predictive and still unavailable for the proposed decision. A verification intervention may improve the accuracy of a measurement and simultaneously reduce application completion.

The research uses a targeted documentary and methodological review rather than a systematic census of the literature. Primary methodological publications, original research and survey reports, professional standards, official statutory provisions, and supervisory guidance are combined with analytical examples. The examples have no empirical provenance and are identified as synthetic. The documentary comparison covers selected United States, United Kingdom, Australian, European Union, and international supervisory materials. Their respective scope and authority are retained. The review does not estimate the prevalence of nonresponse across insurance markets, establish a universal causal effect, or adjudicate individual contracts. Sources that could be inspected only through abstracts or indexed official excerpts support correspondingly limited claims.

A scientifically useful treatment begins by specifying the decision objective. Consider a candidate for a defined policy, at a defined point in an application process, under a defined set of contractual terms. The expected insured cost, the economic benefit of a viable contract, the cost of obtaining additional information, and the constraints upon the insurer's decisions are related but different quantities. A robust decision requires them to be considered together. The inquiry therefore follows evidence from its production and observation through the estimation of risk, the comparison of feasible actions, the governance of collection and use, and the evaluation of downstream outcomes.

Missingness Mechanisms

The application as a partially observed record

Let X denote a characteristic relevant to a specified underwriting inquiry, Z the available covariates, R an observation indicator equal to one when X is observed and zero when it is not, and Y the insured loss under a defined coverage arrangement and period. For multiple characteristics, R is a vector and each component has its own observation process. This notation distinguishes the unknown value of X from the known fact that X has not been observed. It also distinguishes a missing input from a missing outcome, a distinction that becomes essential when only issued policies generate claim observations.

An information architecture should distinguish at least six states: observed substantive value; structurally inapplicable; requested but unanswered; explicitly declined or unavailable to the applicant; unavailable because of an external data or operational failure; and pending at the decision cutoff. These states may call for different responses. They should not be collapsed into a single numeric value. Zero is an exposure measurement, not a universal replacement for missingness. The reason for nonobservation is itself an observation and needs provenance, including whether the reason was explicitly given or inferred from an event log.

Structural absence arises when a characteristic has no meaningful value for an exposure: a question about a feature of a building that does not exist, for instance. A question not displayed because of conditional logic is another matter. An incorrect branch rule may suppress a question that should have been asked. Distinguishing structural absence from suppression requires the rule version and the preceding answers. Pending evidence needs a timestamp; an external service that returns data after the binding decision must not be treated as though it was available beforehand. Event time, processing time, and the time at which a decision is made should be separately represented.

Nonresponse can be defined at the item or application level. Item nonresponse occurs within a record that otherwise exists. Unit nonresponse includes a person who never completes a proposal or a business that declines to supply an entire requested schedule. Withdrawal after an information request is a transition in the application process, not simply an empty cell. For access analysis, the initial cohort of eligible inquiries and applications may matter as much as the cohort reaching underwriting. The appropriate denominator must be stated before a rate is interpreted.

Statistical mechanisms and their implications

The conventional taxonomy of missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR) refers to a probability model for observation. In a simple single-variable setting, MCAR means the response probability does not depend on the relevant complete-data values. MAR permits response to depend on observed information, but not additionally on the unobserved value after appropriate conditioning. MNAR permits dependence on unobserved values. For general patterns, the condition is defined relative to the observed portion of the complete data and the particular missingness pattern. An observed correlation between response and some covariate rules out a simple MCAR assumption for an appropriately defined model; failure to detect such correlation cannot establish MCAR.

Rubin's (1976) analysis of inference with missing data establishes conditions under which a direct likelihood or Bayesian analysis may ignore the missingness mechanism, including MAR and distinctness of parameters. Ignorability belongs to a specified inference procedure and model; it does not mean that the relevant observations may be discarded without consequence. An insurer cannot infer MAR merely because a supplier describes missing records as accidental. In particular, the fact that applicants with unavailable medical data are observed in a database does not reveal their unobserved medical characteristics.

Under a plausible MAR assumption conditional on Z, the distribution of X among responders can inform the distribution among nonresponders at the same Z, subject to positivity and correct specification. Positivity requires some observed information in the relevant strata; the method has little basis for a category in which nobody responds. If R depends on X after conditioning on Z, observed data alone generally leave aspects of the missing distribution unidentified. The correct inference is therefore conditional on assumptions whose relevance must be examined.

Mohan and Pearl's (2021) graphical treatment of missing data provides a more precise perspective: recoverability depends on the target quantity and the graph governing measurement and missingness. Some quantities may be recoverable under mechanisms conventionally called nonignorable, while others are not. A plausible graph for underwriting may include arrows from a latent health characteristic to a disclosure response, from language or interface difficulty to response, and from both previously observed answers and the response to referral. Conditioning on referral can create or alter associations among its causes. Such structures help explain why a model fitted only to referred applicants cannot automatically be used for all applicants.

Graphical analysis also clarifies interventions. A change in question presentation acts on the response process. An intervention that changes a laboratory value acts on the underlying risk characteristic. A database migration changes measurement availability and may leave the risk distribution unchanged. These are different causal operations even if they produce the same observed missingness rate. The scientific interpretation of a response indicator should survive a plausible description of how it was generated.

Informative missingness and transport

Suppose a scoring model uses a missingness indicator as a predictor. The coefficient or learned split may exploit a real association in its training population. The association can arise from adverse selection, the distribution of access to records, the underwriting workflow, or a mixture of them. Its portability is contingent on the relationship between response and subsequent risk remaining sufficiently stable. When an insurer introduces a new form, an accessible submission channel, or a different data vendor, the distribution of response indicators may change even within the same underlying risk classes.

In addition, the applicant population and the issued-policy population are distinct. Let B denote binding or issuance. When Y is observed only for B = 1, the conditional loss distribution for bound policies cannot generally identify the expected insured loss for applicants who were declined, deferred, or withdrew. The problem is especially acute if the response indicator affects B. A loss model that appears well calibrated after the exclusion of incomplete applications may simply be calibrated for the surviving cohort. The insurer needs an explicit target population and a definition of the counterfactual coverage terms before interpreting loss comparisons for alternative underwriting policies.

Selective Disclosure and Access

Private information and the burden of disclosure

Asymmetric information in insurance comprises both information about risk and information about preferences, constraints, and alternatives. Applicants may possess information that affects their expected loss, but they may also lack knowledge of an exposure characteristic or be unable to demonstrate it. A company purchasing insurance may possess internal engineering documents that it cannot promptly assemble; an individual may lack a complete medical chronology. Intentional concealment, misunderstanding, uncertainty, and practical inability produce overlapping observational consequences. A response indicator alone does not establish intent.

Economic incentives make selective disclosure plausible in settings where an answer can influence eligibility or price. They coexist with the desire to obtain enforceable coverage, the costs of gathering records, perceived privacy risks, expectations of fair treatment, and concern about legal consequences. Finkelstein and McGarry (2006), studying the long-term care insurance market, show through their analysis of private information that information about risk and preferences can yield different associations with purchase and ex post risk. This is evidence against reducing all private information to a single latent risk dimension. Their research does not measure the effect of answering underwriting questions; its contribution here is the economic insight that different motivations can coexist in selection.

A nonresponse policy can affect subsequent applicant behavior. Requiring extensive documentation may discourage low-risk applicants with easy alternatives, as well as high-risk applicants who cannot or will not provide information. A simplified application may attract people whose earlier obstacle was inconvenience; it may also increase the scope for unverified representations. Whether these changes improve or worsen portfolio composition is an empirical question. A complete answer is consequently not an unqualified signal of good risk, and an incomplete answer is not an unqualified signal of bad risk.

Nonresponse rate and nonresponse bias

Let r represent the proportion of the target population responding to an item, and let μ1 and μ0 denote the means of a quantity among responders and nonresponders respectively. The complete population mean equals rμ1 + (1 − r)μ0. The respondent-only mean minus the target mean is therefore (1 − r)(μ1 − μ0). This identity shows that missing frequency and response-group differences govern bias jointly. A high response rate does not eliminate bias if the remaining nonresponders differ greatly; a low response rate need not imply substantial bias for every estimand.

Groves and Peytcheva (2008) synthesize 59 methodological studies in their meta-analysis of nonresponse bias. Their abstract identifies the importance of survey design and the statistic being assessed. Application nonresponse is not identical to survey nonresponse because decisions carry contractual and financial stakes. The transfer of the statistical insight is nevertheless justified by the population-mean identity: any analysis that uses a selected group must account for the prevalence of selection and the possibility of systematic differences.

Rates should be defined on meaningful cohorts. The item-nonresponse rate among applicants shown a question differs from the proportion of all applications without a value for that field. A withdrawal rate among applicants referred for additional verification differs from withdrawal among all eligible applicants. Multiple requests should be tracked both as events and as unique applicants, since an applicant asked for three documents is one person with three opportunities for delay. Unconditioned and conditional figures should accompany each other, with time and product definitions held constant.

A selective disclosure problem is therefore inseparable from access. A mandatory question might improve completion because it prevents submission until an answer is entered; it can also prompt guesses or lead to abandonment before submission. An insurer evaluating only submitted forms would record less item nonresponse while overlooking changed access. Measuring completion at the final stage and tracing earlier departures provides a more complete account of the intervention.

Behavioral evidence

The Society of Actuaries Research Institute's (2024) behavioral underwriting research investigated alternative online health questionnaire designs using paid participants. The report describes question-specific effects and states in its limitations that the participants faced different incentives from real insurance applicants. It also reports no general winning design across every question and does not establish the pricing implications of implementation. This is evidence that questionnaire presentation can change reported disclosure. It is not direct evidence that all additional disclosures are true, that an insurer will obtain equivalent effects in a live application, or that a particular change reduces insured mortality.

A sound application process treats comprehension and accessibility as components of measurement quality. Questions need explicit time windows and definitions where materially relevant, alternatives for uncertainty or unavailable records, and the ability to amend an answer before the final decision. A neutral reminder can reduce accidental omissions; a prompt that presupposes an undisclosed condition can introduce a different type of measurement error. Testing a redesigned form requires attention to verified accuracy where verification is lawful, as well as completion and reported disclosure.

Where an insurer conducts a pilot, random assignment of eligible applications to alternative question formats can identify the effect of assignment on outcomes that are reliably observed for the assigned population. It does not identify the effect among only those who complete the form, because completion can itself be affected by format. The pilot should prespecify outcomes and distinguish effects on reporting, verification, referral, completion, and final coverage. The design should monitor differential burdens across relevant applicant groups and avoid withholding an accommodation or necessary safeguard. Randomization is a tool for measuring process effects, not a license to vary legally required treatment.

Statistical Handling and Sensitivity

Define the estimand before choosing a method

The desired estimate may be a portfolio mean, a conditional expected loss, a probability of crossing a decision threshold, an applicant-level predictive distribution, or the effect of a change in an underwriting procedure. These quantities need not have the same identifying assumptions. The appropriate handling of missingness should be selected after the target and observation process have been specified.

Complete-case analysis estimates quantities within a complete-record subset. Under particular sampling and modeling conditions it can consistently estimate certain parameters, but those conditions differ across estimands. The loss of sample size may substantially increase uncertainty, and selection into the subset may distort associations or exposure weights. Dropping incomplete applications from operational evaluation also makes their access outcomes invisible.

Single deterministic imputation replaces a missing value with a chosen number or category. A fixed replacement can be computationally convenient and may form part of a validated prediction pipeline, provided the pipeline is evaluated on naturally occurring missingness. For parameter estimation, however, treating all imputed values as observed suppresses uncertainty; mean imputation can distort variation and associations. An underwriting system should retain the observation flag, source, and method of replacement rather than storing a constructed value as a verified representation.

Multiple imputation creates several plausible draws of missing information, fits the substantive model to each completed dataset, and combines results for an appropriate estimand. For a scalar parameter with m imputations, the combined estimate is the mean of the m estimates, and the conventional total variance is the average within-imputation variance plus (1 + 1/m) times the between-imputation variance. That calculation expresses both ordinary estimation variation and uncertainty about missing values under the imputation model. It does not repair an implausible MAR assumption or a misspecified imputation model. Outcome information may legitimately be used in retrospective parameter estimation where available, yet the actual prospective prediction for a new applicant may use only information available at its decision time. Mixing these two tasks creates leakage.

Response-probability weighting assigns greater weight to observed units with lower estimated probabilities of response, for a target identified by the weighting assumptions. Very small response probabilities yield unstable estimates; trimming changes the bias-variance tradeoff and may change the effective target. Augmented estimators combine an outcome model with a response model and can be advantageous under corresponding identification and regularity conditions. None establishes the unobserved distribution without assumptions, and neither method resolves a prohibition on using a particular variable.

A selection model describes the joint distribution of substantive values and the response indicator. A pattern-mixture model specifies substantive distributions by observed missingness patterns and requires assumptions that link observed and unobserved portions. Sensitivity analysis makes those assumptions visible. Graphical reasoning, such as that developed in Mohan and Pearl's recoverability analysis, can guide a more targeted specification of what can be learned. For each decision use, the analyst should ask which quantities remain unidentified, which external information might restrict them, and whether the decision changes over the plausible range.

The Actuarial Standards Board's data quality standard (ASOP 23, 2016) provides professional context for assessment and disclosure of material data limitations. Its modeling standard (ASOP 56, 2019) situates model use within purpose, assumptions, validation, and governance. These are US actuarial professional standards, not a global statutory rule or a universal endorsement of any missing-data technique.

Partial identification of a synthetic exposure

Consider 1,000 eligible applications. A binary risk characteristic is observed for 800 applicants; 200 of those 800 are classified as high risk and 600 as low risk. The remaining 200 lack a measurement. Let q be the high-risk fraction among missing cases. The high-risk share among the complete-record subset is 0.25, while the fraction among all applications is 0.20 + 0.20q. Without further restrictions q is between zero and one, and the overall high-risk fraction is bounded by 0.20 and 0.40. The observed complete-case fraction cannot resolve that interval.

Suppose expected insured loss, conditional on the specified characteristic under identical coverage, is 100 monetary units for low risk and 500 for high risk. These are fixed synthetic assumptions and not premium quotations. The conditional expected loss among missing cases is 100 + 400q; the mean across the entire application cohort under the hypothetical common coverage is 180 + 80q.

Assumed q among missing Expected loss among missing Cohort expected loss
0.00 100 180
0.10 140 188
0.25 200 200
0.50 300 220
0.75 400 240
1.00 500 260

An assumption that missing cases have the same high-risk share as responders produces a central estimate of 200. The assumption is compatible with the observed data but not established by them. The same observed application records are compatible with each row of the table. Two analysts using different identifying assumptions could therefore reach markedly different prospective loss estimates without disagreeing about any observed field.

Suppose an authorized decision rule permits a particular coverage route when the estimated loss is no greater than 280 units, holding all other factors fixed. The missing-case threshold is reached when q = (280 − 100)/400 = 0.45. A decision that is viable at q = 0.25 becomes nonviable under that synthetic rule when q exceeds 0.45. A missing-data sensitivity analysis can thus report the tipping point even before choosing a single estimate.

An alternative parameterization uses an odds shift relative to the observed high-risk probability p = 0.25: q = expit(logit(p) + δ), where expit(t) = 1/(1 + exp(−t)). The tipping point corresponds to δ = log((0.45/0.55)/(0.25/0.75)), approximately 0.898. A zero shift encodes equality with the observed risk share; positive shifts express higher odds among the missing cases. The parameter is not a measured effect. Its useful range must come from subject expertise, independent verification, credible external evidence, or intentionally conservative bounds.

The numerical example establishes a distinction between missing-data uncertainty and parameter uncertainty. It assumes the loss values of 100 and 500 are known and isolates the unknown composition q. A live model must additionally address estimation error in class-specific expected losses, uncertainty about exposure and coverage terms, and possible drift. Intervals arising from sampling variability should not be presented as bounds on unidentified quantities. Conversely, an identification interval based on extreme assumptions is not a confidence interval.

Prediction, calibration, and decision uncertainty

Prediction algorithms may implement native missing-value routing, surrogate splits, explicit observation indicators, or distributions over potential values. Allocation to a tree branch can yield accurate predictions for an evaluated observation process. That does not establish a substantive explanation of why the applicant lacks information. The relevant evaluation is prediction quality in representative future cohorts under the intended process, with separate reporting by missingness pattern, channel, and evidence supplier where feasible.

Calibration examines agreement between predictions and outcomes in a well-defined group. A calibration estimate from issued policies with observed outcomes does not independently validate decisions applied to declined applicants. Discrimination measures assess ranking of cases rather than probability accuracy. A system may rank issued risks effectively but assign systematically excessive prices or excessive referrals to a poorly observed group. Neither global discrimination nor global calibration can substitute for an evaluation of decision outcomes.

Evaluation splits should preserve time order and avoid overlap across duplicate applicants, related policies, or repeated records that could leak information. Retrospective data should be reconstructed at the actual decision timestamp. Validation should compare original missingness with simulated masking only as a supplementary stress test: arbitrarily hiding values may reproduce MCAR-like missingness rather than the selective mechanisms an underwriting process actually encounters.

When outputs are uncertain, the appropriate underwriting action depends on the cost of error and the feasible responses. A model can produce a prediction interval or an explicit abstention/referral signal, but an abstention policy still affects access and workload. Its accuracy and its institutional effects need separate assessment. A decision threshold chosen on one validation cohort also needs monitoring after changes in policy, product mix, or question design. Good governance tracks assumptions and observed consequences together.

Referral and Verification Protocols

Information acquisition as an underwriting decision

Verification converts some uncertainty into additional evidence. It should be evaluated as a decision with financial, informational, operational, and access costs. A referral may request a clarification, a document, a new inspection, an authorized data search, a clinical examination where lawful, or a second review. Its expected value depends on the probability of resolving the uncertainty and the extent to which resolution changes a feasible decision. The acquisition of information that cannot lawfully be acted upon, or that cannot change a decision already determined on other grounds, may impose cost without underwriting benefit.

A synthetic two-action example illustrates the calculation. Suppose accepting a high-risk applicant produces an incremental expected decision loss of 400 units and accepting a low-risk applicant produces zero. Declining a low-risk applicant incurs a lost-opportunity cost of 80 units; declining a high-risk applicant incurs zero. With a high-risk probability of 0.25, immediate acceptance has expected decision loss 100 and immediate declination has expected decision loss 60. A perfectly reliable and costless observation would allow selection of the lower-loss action in each state and eliminate both losses. The expected value of perfect information under these assumptions is 60 units. A real test is imperfect, may be unavailable, may delay coverage, and may itself change participation. Its expected value cannot exceed the perfect-information value within the specified simplified model, before considering additional consequences.

More generally, for feasible actions a and available information I, consider the conditional expected loss E[L(a, X) | I]. A proposed observation W has gross expected value equal to the minimum expected loss before W minus the expected minimum conditional loss after W. The net value subtracts acquisition and process costs. The feasible action set incorporates legal and product constraints; the loss function is a decision aid, not a definition of all obligations owed to applicants. Privacy, rights, and mandatory coverage constraints cannot be bypassed by assigning them an arbitrary monetary cost.

Alternative evidence and accessibility

Substitute evidence should be assessed for the construct it measures, its origin, the coverage of its underlying data source, its temporal relevance, and its uncertainty. A recent independent inspection may provide evidence about a property's condition; an address-level database may cover only some properties or use a broader geographic approximation. A financial statement prepared under a recognized accounting framework may clarify a commercial exposure but cannot substitute automatically for a physical safety assessment. A record-linking service that returns no match supplies evidence about the search operation, not necessarily about the absence of the underlying event.

An evidence hierarchy should be specific to the material question and the authorized decision. It should distinguish verified direct measurements, documented representations, vendor-derived estimates, and model-generated imputations. The relative usefulness of each can vary by product and factual circumstance. ASOP 23's assessment of data suitability supports the underlying professional attention to data source and limitations when applicable. A traceable decision record should state what was requested, why it mattered, which sources were available, which limitations remained, and whether the action followed an authorized routine or an exception.

The NAIC's accelerated underwriting overview describes the relationship between streamlined underwriting and traditional pathways when available information is insufficient. This is a practical institutional distinction. Eligibility for a streamlined route and eligibility for insurance coverage are not inherently identical. Where a traditional route remains permitted and available, referral can preserve an opportunity that a simplistic automatic decline would remove. The route must itself be examined for cost, time, and completion outcomes.

The behavioral study discussed earlier (Society of Actuaries Research Institute, 2024) also supports testing question-specific changes rather than presuming that a single questionnaire redesign solves every omission. A useful operational pilot can compare a clarification prompt, an alternative document, and the standard route where the comparison is lawful. The evaluation should include documented accuracy and completion, not only the rate at which a database field becomes nonblank.

Verification samples and mortality monitoring

Accelerated life underwriting illustrates a challenge common to many insurance lines: decisions are made using streamlined evidence, but their quality is often assessed against a different and more intensive verification process. Seeman and Herzog's (2024) mortality slippage monitoring study describes data from multiple accelerated underwriting programs and comparisons using random holdouts and post-issue audits. The published projected mortality-slippage figures are actuarial estimates based on changes in assessed risk class and related assumptions. They should not be read as observed mortality differences caused by the underwriting procedure.

A random holdout selects applicants for more intensive evaluation before issuance. This can reveal discrepancies between a streamlined decision and an assessment based on additional evidence. Where some selected applicants withdraw or fail to complete verification, the realized verification subset may cease to be representative of the randomized cohort. Post-issue auditing can reduce one kind of pre-issue interruption but introduces questions about what can be inspected, when evidence becomes available, and which policies have auditable records. The two methods have distinct denominators and observation processes.

Gen Re's original 2024 accelerated underwriting survey (Potter, 2024), based on 38 responding carriers, reports monitoring practices including random holdouts. It documents the existence of these operational approaches among responding organizations. Survey participation, question-specific denominators, and underlying carrier practices limit inference to the entire market. It is evidence of practice rather than a randomized effectiveness comparison.

Two synthetic calculations demonstrate why sampling and verification completion must be recorded. First, suppose a cohort has 900 routine and 100 referred applicants. An audit independently samples 100 from each stratum. It identifies 10 discrepancies in the routine sample and 30 in the referred sample. The unweighted pooled discrepancy rate is 20%, whereas the population-weighted estimate is 0.9 × 0.10 + 0.1 × 0.30 = 0.12, or 12%. The stratified estimate applies to the original cohort under the stated sampling assumptions and complete ascertainment within each sample. An audit that oversamples difficult cases without correcting weights can exaggerate an aggregate discrepancy rate.

Second, suppose 100 applicants are selected for verification and 80 complete it. Twelve of the 80 show an established discrepancy. The verified-subset rate is 15%. If the other 20 outcomes are entirely unknown, the discrepancy fraction among all selected applicants lies between 12% and 32%, assuming the observed 12 discrepancies and 68 nondiscrepancies are correctly identified. A claim that the random audit proves a 15% whole-cohort discrepancy rate would require additional assumptions about completion or additional evidence.

These examples also distinguish verification discrepancy from ultimate insured loss. A discrepancy can change a risk-class assessment without producing an observed claim; a claim may occur even when application evidence was entirely correct. Discrepancy monitoring is an intermediate measurement-quality assessment. Loss and experience monitoring require distinct exposure definitions, development periods, coverage adjustment, and suitable statistical uncertainty.

Operational controls

Each information request should be linked to an application identifier, purpose, eligibility rule or referral rationale, request date, source, status, and resolution. A separate field should record the reason for pending or absent evidence, without inferring culpability from a technical failure. The process should preserve an accessible channel for correction, record whether an alternative was offered, and distinguish a deliberate applicant decision to abandon from inactivity with an unknown reason.

A monitoring dashboard should track item exposure, response, refusal and unknown reasons, search failures, time to resolution, number of repeat requests, referral rates, conversion, withdrawals, issuance, classification discrepancies, and eventual covered outcomes where observed. Percentages require numerator and denominator definitions and cohort maturity. Reporting both per-request and per-applicant burdens avoids counting frequent requests to the same individual as independent cases.

An escalation process is useful when a material uncertainty survives normal clarification, when contradictory evidence is received, or when a decision would otherwise rely predominantly on a proxy or inferred value. Escalation should identify the decision authority and the evidence needed to resolve the issue. A human review can improve contextual interpretation but can also reproduce bias or introduce inconsistency. Review quality depends on instructions, training, authority, records, and audit rather than the mere presence of a human decision-maker.

Fairness and Jurisdictional Limits

Prediction and permissible use

Missingness may correlate with protected characteristics, unequal record availability, geography, disability, age, financial resources, or other circumstances. Such associations are hypotheses requiring lawful empirical examination; an observed performance difference does not by itself identify a specific cause. Underwriting decisions should evaluate the substantive relevance of the characteristic and the legal permissibility of its use, in addition to model performance.

The Actuarial Standards Board's risk classification standard (ASOP 12, 2005, updated 2011) states professional considerations for relationships to expected outcomes and recognizes that causal relationships are not invariably required for a risk classification characteristic. That proposition does not establish that any predictive proxy is permissible. Applicable laws, the scope of the professional work, reliability, and the suitability of the proposed application remain relevant.

Auditing needs to account for who is represented in the data. A review restricted to completed applications may exclude persons most affected by the collection burden. A review restricted to issued policies may omit declined and withdrawn applicants. It may also lack authorized group information. A defensible assessment should report both decision outcomes and the rates at which their required measurements are observed, stratifying only where permitted. It should preserve uncertainty about unavailable group attributes rather than silently treating inferred categories as verified identities.

Consider two applicant groups with identical underlying risk distributions and different rates of failure to obtain third-party records. A fixed surcharge associated with a failed search would create different average prices even if the underlying risk distribution were identical. That is a hypothetical mechanism, not proof that a real insurer uses such a rule or that any resulting difference necessarily violates a particular law. An insurer would need to investigate the data mechanism, the predictive basis of the rule, the scope of applicable law, and reasonable alternatives. Analysis of the process must extend beyond aggregate model accuracy.

United States

Insurance regulation in the United States involves product-specific rules and state authority, together with applicable federal law. Professional standards and NAIC materials have distinct status from binding statutory and regulatory provisions. A national description of accelerated underwriting is useful for understanding practice, while the permissibility of a particular data use must be evaluated in its jurisdiction and line.

New York's Department of Financial Services (2024) addresses covered uses of artificial intelligence and external consumer data in underwriting and pricing through Insurance Circular Letter No. 7. It describes expectations relating to governance, discrimination, potential proxies, transparency, and documentation. Its discussion of quantitative assessment does not expect insurers to collect additional protected-class data solely for that analysis. Where an audit employs lawful imputation or an appropriate alternative assessment, the inferred audit attributes should remain analytically distinct from applicant-level underwriting inputs. The circular is a New York supervisory document with its own scope, not a nationwide statutory permission or prohibition.

Product differences are fundamental. The federal Marketplace guidance on pre-existing condition coverage states that Marketplace plans cannot refuse coverage or charge more because of pre-existing health conditions. This sharply limits the relevance of certain health questions for individual acceptance or pricing in that setting. A life insurance process involving risk assessment cannot be generalized to Marketplace health insurance. Even when the same underlying variable is available, the lawful decision objective can be different.

United Kingdom and Australia

The United Kingdom's Consumer Insurance Act 2012 establishes a consumer duty to take reasonable care not to make a misrepresentation to the insurer. Section 2 also recognizes that failure to comply with a request to confirm or amend particulars previously supplied can be capable of constituting a misrepresentation. The statutory notion of reasonable care is not equivalent to a statistical missingness category. Whether an omission amounts to misrepresentation depends on statutory scope and the actual facts. The claim here is limited to section 2 as available from the official indexed text; it does not purport to resolve remedies, nonconsumer insurance, or individual cases.

Australia's Insurance Contracts Act 1984 provides an especially informative contrast. Section 20B(5) states, in the consumer insurance duty context, that failure to answer a proposal-form question, or an obviously incomplete or irrelevant answer, does not alone constitute misrepresentation. Under the other disclosure duty, section 21(3) addresses waiver in relation to certain unanswered or obviously incomplete or irrelevant proposal-form answers. These provisions are examined in the official October 8, 2026 compilation. Their application depends on the contract and statutory context. Their underwriting significance is that the legal treatment of an incomplete response cannot be derived simply from the fact of nonobservation. The law may attach different consequences to requests, received answers, and insurer conduct.

Neither comparison makes a missing exposure measurement irrelevant. An insurer may need to evaluate the uncertainty that remains before forming a contract, subject to applicable rules. The distinct question of contractual remedies after an alleged misrepresentation requires the corresponding statutory tests and facts. A data pipeline should therefore not automatically copy a missingness flag into a fraud, misrepresentation, or post-claim remedy category.

European Union and international comparison

The European Union's GDPR establishes principles relevant to collection and use. Its data minimization and automation provisions include Article 5 on processing principles and Article 22 on certain solely automated decisions producing legal or similarly significant effects, subject to specified exceptions and safeguards. The provisions do not establish a blanket prohibition on underwriting automation. They emphasize the importance of assessing the purpose and necessity of collection, the applicable basis for processing, and whether the decision process falls within relevant automated-decision provisions. This discussion is limited to the indexed official provisions and does not assess a particular implementation or national derogation.

EIOPA's 2025 announcement of its AI governance and risk management opinion describes a risk-based approach within the insurance sector and identifies the opinion's relationship to the broader legislative framework, including the exclusion of prohibited and high-risk AI Act systems from that opinion's scope. This statement concerns the published announcement rather than a clause-by-clause analysis of the opinion or the AI Act. It reinforces the distinction between the statistical capability of a system and the governance regime that applies to its intended use.

International supervisory principles address insurers' risk-management structures, including underwriting risks, without constituting a universal individual right to a particular contract or a harmonized missing-data statute. Supervisory expectations, actuarial professional practice, local legal duties, contractual conditions, and internal policies need separate source labels in a comparative account.

Fairness as an operational question

The fairness evaluation of a missing-data policy should compare authorized alternative policies rather than treat a baseline system as self-justifying. The alternatives might involve an additional clarification opportunity, an alternative documentation path, a less burdensome request, a documented abstention and referral, or a robust decision that remains feasible throughout the plausible risk range. Their consequences should be measured on common cohorts and across stages.

Model governance must distinguish data used for decision-making from data used for authorized validation and discrimination assessment. The latter may have separate purposes and restrictions. A group attribute inferred for aggregate audit should not be silently transferred into the decision rule. Where a reliable estimate cannot lawfully or practically be produced, the limitation should be recorded and the institution should consider other permissible evidence.

Transparency also depends on the form of explanation. A statement that an application is incomplete can be accurate yet insufficient to enable resolution if the applicant cannot identify the missing requirement or the suitable alternative. Where applicable, explanations should state the material issue and the next available route in an understandable form, without suggesting an inferred adverse fact was actually observed. A correction channel is scientifically valuable because it supplies feedback about measurement errors as well as supporting procedural accountability.

Comparative Synthesis and Reproducibility

Variations across insurance lines

Life insurance underwriting often combines health representations, medical records, permitted external information, and other characteristics relevant to the product. Accelerated pathways reduce certain requirements for selected applicants. The central missing-data challenge concerns both unavailable evidence and the selective observation of intensive verification results. The monitoring practices described by Seeman and Herzog's accelerated underwriting study illustrate how process design influences the evidence obtained. The distinction between projected mortality and realized experience applies broadly to retrospective performance assessments.

Property underwriting may depend on construction details, occupancy, replacement values, hazard information, and protective measures. Incomplete or stale building information can arise from poor records or rapid renovation. Geography-level estimates can supplement evidence within their verified domain, but spatial estimates and direct inspections have different error structures. When a hazard model uses default building attributes, the uncertainty associated with those defaults should be retained in risk assessment rather than presented as actual construction data.

Commercial liability and cyber insurance may rely on organizational representations about activities, revenues, contracts, controls, and past events. A missing security-control answer might reflect an undefined term, incomplete knowledge by the respondent, an absence of documentation, or the actual absence of the control. A simple binary conversion of the blank into control failure conflates these mechanisms. Where a control is material, precise definitions, authorized alternative evidence, and the applicant's ability to substantiate the response matter.

Reinsurance involves data supplied by cedents rather than individual retail applicants. Incomplete exposure schedules, missing location attributes, and uncertain claim development may materially change aggregate and tail-risk estimates. The principles of provenance, sensitivity, and qualification remain relevant, but the legal relationship, information scope, portfolio units, and feasible contractual mechanisms differ. Missing exposure data in a cedent's file cannot be assumed to have the same behavioral origin as an individual's blank application item.

The four settings illustrate a general scientific distinction: the physical or economic exposure, the evidentiary representation of it, the process that determines the representation's availability, and the underwriting decision based on available evidence. An evaluation is more informative when these are separately specified. This is an analytical synthesis of established methodology and practice, not a newly named theoretical system.

A reproducible research design

A prospective empirical investigation should begin with a fixed eligible applicant cohort, product and jurisdiction, question versions, channels, observation windows, and decision policy versions. An application-event table would contain pseudonymized identifiers and timestamped events, question exposure, observed or absent evidence, recorded reason, referral assignment, verification completion, decision, and application exit. The underlying answer values should be retained only where lawful and necessary, with appropriate access restrictions and retention controls. A separate outcome table would define policy inception, coverage, exposure duration, and observed losses or other authorized outcomes.

The first analysis is descriptive: tabulate the stages of observation and the distribution of missingness by product, channel, question version, and evidence supplier. Estimate confidence intervals for properly defined rates and identify clustering and repeated measures. Separate cases never asked a question from those asked and not responding. Examine missingness co-occurrence and reason codes, reporting unknown causes rather than manufacturing a classification.

The second analysis specifies estimands. An access question may target the probability of reaching an offer among eligible applicants. An accuracy question may target verified discrepancy rates by response pattern. A prospective loss question may target expected covered losses under a particular underwriting policy; if claims are observed only for issued contracts, that target may require assumptions about selection and counterfactual coverage or be only partially identified. An intervention study can estimate assignment effects on observable application outcomes but cannot substitute a selected completion cohort for the original assigned group.

The third analysis compares plausible statistical methods under a stated missingness model, including complete-case estimates as a diagnostic, suitably specified imputation or weighting where identified, and alternative MNAR sensitivities. The analysis should specify the provenance of assumptions and show threshold effects. Models used prospectively should be validated at the actual information cutoff. In an applicable actuarial context, ASOP 56's model validation provisions supply professional support for attention to purpose and model performance; they do not establish that an individual model has been successfully validated.

The fourth analysis evaluates interventions. A lawful randomized reminder or question-format pilot can support intention-to-treat comparisons on outcomes measured across all assigned participants. The protocol should predefine primary measures, handling of attrition, uncertainty, and any subgroup assessments. A stratified audit can improve precision in relatively rare referral patterns if sampling probabilities are recorded and appropriate weighting is used. If verification is selectively incomplete, use bounds and sensitivity alongside verified-subset estimates.

The fifth analysis evaluates institutional constraints and decision consequences. Before access to potentially sensitive information, define collection purpose, legal scope, permissions, and retention. Compare the feasible actions within the actual rules. Record time and cost burdens, applicant correction opportunities, and adverse decision reasons. A technically strong predictor that cannot lawfully be used for a particular decision is not a feasible solution to that decision problem.

Finally, reproducibility requires preserving data definitions, code versions, temporal cutoffs, transformation rules, sampling probabilities, method assumptions, and the complete sequence of model and policy changes. Where underlying private data cannot be shared, reproducibility can be supported by executable code against synthetic datasets, schemas, aggregate quality checks, and controlled verification where authorized. Such measures improve scrutiny without pretending that an external reader has independently reproduced unavailable proprietary data.

Interpreting operational evidence

Suppose an insurer's missingness rate falls following a switch to mandatory answers. One possible interpretation is improved collection; another is a rise in forced guesses; a third is abandonment before submission. Evaluating all three requires questionnaire exposure, submitted response data, applicant exits, and authorized verification. A decrease in blanks alone cannot discriminate among them. Similarly, an improvement in issued-policy loss experience may arise from better information, changed risk mix, stricter selection, changed products, or claim development. A causal claim requires more than a before-and-after comparison.

The governing question is whether an underwriting policy remains justified under the information process that it induces. Once a decision rule changes the incentives to disclose or the probability of referral, it can alter its own future training population. Monitoring should therefore assess the stability of relationships after deployment rather than treating a historical missingness coefficient as a timeless property of risk.

The International Association of Insurance Supervisors' enterprise risk management principles in ICP 16 and related ComFrame material place underwriting policy within wider risk governance. They support attention to the institutional setting in which technical assessment takes place. The appropriate governance arrangements remain sensitive to insurer size, line, domestic implementation, and supervisory scope. This is a framework-level comparison, rather than an attempt to transpose a specific ComFrame requirement to every insurer.

Scope and evidentiary limitations

The documentary research synthesizes inspected primary research reports, publisher abstracts, official professional standards, supervisory texts, and statutory excerpts. It is not a systematic literature review. Rubin, Groves and Peytcheva, and Finkelstein and McGarry are used within the scope of their accessible abstracts; Mohan and Pearl are examined through the open preprint. The Society of Actuaries Research Institute report is examined through extracted text and its visually inspected limitations page; other page-image retrievals were unavailable. Proprietary study microdata and carrier-level responses were not independently accessible. Official United Kingdom and GDPR text retrieval was limited to indexed provisions, and the EIOPA source is the official announcement. These limitations delimit the claims made; specific legal application, quantitative prevalence, and causal identification in live underwriting remain separate research tasks.

All numerical illustrations are synthetic, with specified assumptions and arithmetic that can be replicated directly. No real applicants, insurer experience, loss estimates, or personal records underlie the tables. No external scholarly peer review is represented as completed. The research does not serve as legal, actuarial, or consumer advice for a particular decision.

Conclusion

Missing information is a property of an observation process situated within an underwriting institution. Its impact arises from the relationship between the underlying exposure, the probability of observing relevant evidence, the decision objective, and the permitted uses of the available information. An unanswered question can be predictive under a particular process, yet the associated characteristic may remain unidentified and the proposed action may lack a sufficient lawful basis. These distinctions determine how underwriting should interpret incomplete records.

The statistical investigation establishes why complete-record analysis, deterministic imputation, weighting, multiple imputation, and missingness-aware prediction require different assumptions and serve different purposes. Partial identification and threshold sensitivity are especially important when a plausible nonignorable process cannot be resolved. The synthetic case demonstrates that identical observed records can support materially different prospective expected losses. Decision robustness can be evaluated without presenting a single unidentified estimate as established fact.

The institutional investigation shows that information acquisition is itself an underwriting decision. Verification may reduce uncertainty, but it introduces cost, delay, selective completion, and possible changes in access. Random assignment to an audit or referral does not guarantee representative verified outcomes. A defensible evaluation tracks the original cohort, sampling design, completion process, and downstream outcomes, distinguishing verification discrepancy from realized insured loss.

Comparative legal sources establish the necessity of interpreting nonresponse within specific product and jurisdictional contexts. Consumer disclosure obligations, data processing restrictions, professional risk classification standards, and supervisory governance operate at different levels. Predictive usefulness neither establishes contractual misrepresentation nor determines legal permission. A traceable process preserves the distinction between missing facts, inferred facts, and verified facts.

The principal contribution of this monograph is the integration of these evidentiary, statistical, economic, operational, and legal distinctions into a reproducible scientific analysis of applicant nonresponse. For insurance institutions, the practical consequence is a change in the unit of evaluation: the model and the application process that generates its inputs must be evaluated together. The quality of underwriting is expressed in the quality of decisions under defined uncertainty, with measurable access consequences and a documented evidentiary basis.

References

  • Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. DOI: 10.1093/biomet/63.3.581. Source

  • Mohan, K., & Pearl, J. (2021). Graphical models for processing missing data. Journal of the American Statistical Association, 116, 1023–1037. DOI: 10.1080/01621459.2021.1874961. Open preprint (2018), arXiv:1801.03583. Source

  • Groves, R. M., & Peytcheva, E. (2008). The impact of nonresponse rates on nonresponse bias: A meta-analysis. Public Opinion Quarterly, 72(2), 167–189. DOI: 10.1093/poq/nfn011. Source

  • Finkelstein, A., & McGarry, K. (2006). Multiple dimensions of private information: Evidence from the long-term care insurance market. American Economic Review, 96(4), 938–958. DOI: 10.1257/aer.96.4.938. Source

  • Society of Actuaries Research Institute. (2024). Redesigning the life insurance underwriting journey with behavioral science. Research report. Source

  • Seeman, L., & Herzog, K. (2024, August). Accelerated underwriting: Mortality slippage study and monitoring best practices. Product Matters! Society of Actuaries. Source

  • Potter, R. (2024, November 12). Individual life accelerated underwriting – Highlights of 2024 U.S. survey. Gen Re. Source

  • Actuarial Standards Board. (2005, updated 2011). Actuarial Standard of Practice No. 12: Risk classification (for all practice areas). Source

  • Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data quality. Source

  • Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Source

  • National Association of Insurance Commissioners. (n.d.). Accelerated underwriting. Insurance topics. Accessed November 19, 2026. Source

  • International Association of Insurance Supervisors. (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups. Online tool, ICP 16 and ComFrame. Source

  • 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. Source

  • United Kingdom. (2012). Consumer Insurance (Disclosure and Representations) Act 2012, c. 6, §2. Source

  • Australia. (1984). Insurance Contracts Act 1984 (Cth), compilation October 8, 2026, §§20B and 21. Source

  • European Parliament & Council of the European Union. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88, Arts. 5 and 22. Source

  • European Insurance and Occupational Pensions Authority. (2025, August 6). EIOPA publishes Opinion on AI governance and risk management. Source

  • HealthCare.gov. (n.d.). Coverage for pre-existing conditions. Accessed November 19, 2026. Source