Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Insurance Underwriting Reject Inference and Selective Labels

Abstract. Insurance underwriting creates the population from which future insurance experience is learned. Applicants who are declined, deferred, offered modified contracts, or who do not take up coverage generally fail to generate comparable insured outcomes at the original insurer. An underwriting model evaluated on the resulting portfolio can therefore exhibit excellent apparent performance while leaving the consequences of alternative decisions unidentified. This monograph examines the selective-label problem as an insurance-specific problem of statistical identification, contract design, observation, and decision evaluation. It distinguishes underlying hazard, realized exposure, covered loss, reported claims, and economic contribution, and evaluates complete-case estimation, propensity weighting, joint selection models, imputation, partial identification, off-policy evaluation, and additional evidence collection. Published methodological studies and insurance industry evidence support a conditional rather than universal assessment of reject-inference techniques. Formal identities and synthetic examples establish how observationally equivalent applicant populations imply different estimates and policy choices. The analysis also addresses legal and ethical constraints on validation, including the distribution of coverage opportunities and sensitive-data processing. Its principal finding is that uncertainty arising from underwriting selection must remain visible throughout model validation and decision governance. An inferred outcome is a model-dependent estimate, not an observation. The study provides an analytical synthesis and reproducible validation design; it does not estimate an actual insurer's reject population or introduce a new named theory.

Keywords: Insurance underwriting; reject inference; selective labels; missing data; sample selection; identification; partial identification; actuarial modeling; off-policy evaluation; insurance access; model governance; exposure; contract heterogeneity.

Introduction and conceptual foundations

Underwriting is the process by which an insurer evaluates a proposed risk and determines whether and under what conditions it will assume that risk. Its decisions govern both the distribution of coverage and the subsequent production of insurance data. The empirical problem begins at application, before a policy exists. A declined application can include a recorded risk score, supporting documentation, a time stamp, and a decision reason, while its subsequent insured loss under the proposed contract remains unavailable. The apparent completeness of the administrative record conceals this absence.

Reject inference denotes attempts to learn the distribution of outcomes that would have been recorded for rejected applicants under a specified acceptance and exposure regime. Selective labels is the broader learning problem in which previous decisions determine which observations have outcome information. The distinction is significant because a system can approve a proposal while leaving a label unobserved, and because labels can be selectively absent even after inception. A decision to approve, a customer decision to purchase, continued exposure, loss occurrence, reporting, and final loss valuation constitute different observation mechanisms.

Rubin's (1976) analysis of missing data establishes conditions under which statistical inference can disregard a missingness mechanism. Its relevance to underwriting lies in the relationship between recorded information and the absent outcome. Lakkaraju et al. (2017) independently focus on the evaluation of decision rules where the historic decision controls which outcomes can be observed. Their selective-label study includes a direct insurance application. For insurers, the central methodological issue is to determine whether the observed outcome relationship remains meaningful beyond the contracts that generated it.

Three questions organize this study. First, when does the historical acceptance mechanism distort estimates of the risk quality of the applicant population? Second, which assumptions permit reject inference, and what should be reported when they cannot be verified? Third, what forms of additional observation and validation can improve evidence without unfairly altering the provision of insurance?

The object is prospective inference from application-level information. Portfolio reserving, rating regulation, reinsurance pricing, claims adjudication, and the economics of adverse selection enter where they materially affect the observation process, rather than being independently reconstructed. The analysis covers voluntary private underwriting in a general international perspective. No implication is made that every insurance market lawfully permits declining, individual risk classification, or experimental offers.

The units of observation and the object of prediction

Let an application be a request associated with one applicant or insured exposure, a proposed coverage specification, an underwriting date, and a decision process. A distinct application identifier connects the quote and underwriting states to policy inception and future exposure. A person can submit multiple applications; a company can insure multiple locations; a policy can cover multiple assets. An analysis that treats all these rows as statistically independent may deliver misleadingly narrow uncertainty intervals.

Risk, claim, and loss are separate scientific objects. A physical or biological event can occur without insurance. An insurer's obligation arises from coverage, exclusions, deductibles, limits, and other applicable contract provisions. A claim is a reported or otherwise identified demand that enters an administrative process. The final financial outcome may incorporate adjustment expenses, recovery, and payment development. An insurer may know a third-party event without knowing the corresponding contractual loss it would have incurred.

Define Y(c,t)Y(c,t) as the outcome that would arise if contract specification cc were in force for a stipulated exposure and horizon tt. It may be an indicator of a covered event, an ultimate amount subject to contract terms, or a carefully defined discounted cash-flow measure. The notation presupposes a meaningful intervention and a sufficiently precise population. It does not require that issuing such a contract be legal or feasible for every historical applicant; the target population should be restricted before defining the estimand.

When the outcome is profit, revenue and expenditure must be defined symmetrically for competing offers. A profitable approved application is not equivalent to a low-hazard applicant. Higher premiums, exclusions, reinsurance, and inspection requirements can produce materially different results from the same underlying exposure. An evaluation based on financial contribution must therefore include terms and take-up, or expressly fix them.

Evidence and methods

This is a source-based analytical monograph. Searches targeted primary methodological research on missingness and selective labels, insurance-specific observations, and official actuarial and supervisory materials. The evidence includes directly inspected portions of author manuscripts and official publications, a narrowly interpreted journal abstract, and an industry study. The bibliography distinguishes the examined version and publication identity where relevant. The literature selection is purposive; no claim of a systematic review, exhaustive database search, or meta-analysis is made.

The scientific procedure has four components: define the target outcome and denominator; establish identification conditions through elementary probability identities; compare method families against those conditions; and translate the results into an insurance validation design. Synthetic scenarios illustrate mathematical consequences. They are independent constructions with no claim to represent observed rates.

Direct insurance evidence

Lakkaraju et al. (2017), section 5.2 of their evaluation study, describe about 50,000 insurance requests submitted to a commercial provider and handled by 74 decision-makers. The available financial outcomes concern approved requests. This is explicit insurance-domain evidence of selective labeling, but the dataset does not establish a universal numerical relationship between rejection and loss. The study investigates comparison of model and human decisions under particular assignment and agreement assumptions. Commercial profit and loss also differs from pure premium, mortality, or claim frequency.

In life underwriting, Seeman and Herzog (2024) report random audit monitoring across 30 accelerated underwriting programs with over 33,000 audited lives. Their calculation translates classification differences into expected mortality effects using actuarial assumptions. The evidence is informative about medical evidence requirements, audit sampling, and the distinction between classification discrepancy and directly realized mortality. It does not reveal the counterfactual claim history of every declined applicant.

Transfer from credit scoring

Ehrhardt et al. (2021) examine credit reject-inference methods by comparing their statistical assumptions and experiments on synthetic and previously financed observations. Their results offer a warning against treating any method as generally superior. Credit repayment is conditional on a loan and its terms, which makes the analogy useful. Insurance nevertheless differs in its loss-frequency and severity structure, duration of exposure, contractual triggers, and risk transfer.

The appropriate reading of credit research is methodological. A result about a credit scorecard cannot be converted into an asserted percentage improvement in insurance underwriting. The commercial insurance evidence and accelerated underwriting studies complement rather than replicate it.

Selection mechanisms and missing labels

The application-to-experience process

Let XX denote information recorded and available before decision, UU denote relevant information absent from the analytical dataset, DD denote the underwriting disposition, BB denote binding of coverage, and RR denote the presence of an outcome that is comparable under the research definition. Let CC denote actual terms, EE denote duration or amount of exposure, and TT denote the valuation date. Decisions may include accept on standard terms, accept with modifications, request further information, defer, decline, and refer to another decision-maker.

For a standardized target contract cc, full observation requires more than an accepted disposition. It requires that a suitable policy was bound, exposure was established, the observation horizon was achieved or appropriately adjusted, and reporting and valuation were sufficient for the outcome definition. In shorthand,

R=I(D permits c,B=1,C=c,outcome observable at T).R=I(D\text{ permits }c, B=1, C=c, \text{outcome observable at }T).

This indicator is a conceptual record of conditions, not an assertion that exposure and reporting are always binary. In practice there may be right censoring, delayed reporting, partial data access, and different effective periods. Those mechanisms require distinct recording and analysis.

The chronological order matters. A cancellation can be affected by the customer's private information, and an earlier proposal may be reopened. A pending application at a data extract date should not be recorded as a final rejection. A withdrawal during medical examination and an outright refusal should not share an outcome code. A nonbinding offer at a high premium may reveal price sensitivity more strongly than physical risk.

The NAIC's accelerated underwriting overview describes pathways that use available information to waive traditional requirements for some life applicants, while others proceed to additional assessment. A binary accept/reject encoding removes this intermediate state. Routing decisions create selective observation of medical evidence even before the coverage decision, and these observations may then influence the offer itself.

Missingness and contract intervention

A useful simplified causal description has arrows from XX and UU to the historical decision and to potential loss; from the decision to the terms offered; from terms and private information to take-up; and from contract and exposure to the realized covered outcome. Observation can be influenced by the event, reporting costs, and record access. Conditioning on policy issuance selects jointly on the historical risk controls and customer behavior.

Consider two applicants with equal observed information. One has additional unrecorded evidence of substantial hazard; the other does not. If an underwriter accurately acts on the hidden evidence, the observed outcome distribution among bound policies becomes favorable relative to the total applicant population within that recorded risk class. The same pattern may arise if a historical underwriting rule mistakenly associates an irrelevant trait with risk and excludes applicants arbitrarily. Accepted-only loss data cannot by themselves determine which of these explanations is correct.

A contract can also alter events. Requirements for fire suppression or cyber controls can reduce frequency. A deductible changes the payment distribution and can alter reporting or prevention incentives. Health coverage can affect use of medical services. In such settings, the hypothetical loss under one contract is not the observed loss under another. Conditioning on the actual outcome while treating coverage as a passive observation instrument conflates selection and intervention.

Different layers of missingness

The following mechanisms warrant different treatment:

Mechanism Missing scientific information Appropriate study direction
Decline or deferral Outcome under a specified offered contract Identification assumptions, bounds, suitable new evidence
Modified offer Outcome under original terms Contract standardization or causal comparison of terms
Non-take-up Outcome under actually bound exposure Offer/election analysis, explicit estimand
Early lapse or cancellation Subsequent experience Censoring analysis, competing risks, exposure measurement
Reporting or settlement delay Mature claim or ultimate loss Development and valuation, reporting analysis
Unavailable third-party record External event or proxy Linkage quality, observation and measurement models

The target is different in each row. A model that imputes one undifferentiated missing label risks generating values for incompatible outcomes.

In accelerated underwriting, random referral to additional risk assessment can illuminate an otherwise unobserved classification. Seeman and Herzog's (2024) monitoring study also describes withdrawals and differences in the availability of medical records. These mechanisms mean that even selection into an audit and completion of that audit must be distinguished.

Economic selection and the learning feedback loop

Adverse selection concerns the distribution of risks associated with participation and pricing when relevant information is asymmetrical. Statistical selection concerns the observation of outcomes in an analytical sample. Both may arise in the same insurer, but the existence of one does not prove the other. A random underwriting restriction can create selective labels without applicant-side private information. Conversely, price-induced consumer selection may shift the portfolio even if every bound policy has complete claim data.

Repeated model updates can reinforce historic decision boundaries. A class with limited representation receives few new observations. Subsequent validation may then reward a model that performs well in the better-known region while silently extrapolating in the excluded region. This is a structural limitation of the data-generating process. It is not resolved by simply increasing the number of approved exposures within that same region.

Identification assumptions

The decomposition of an applicant risk mean

For a fixed feasible outcome definition, abbreviate Y(c,t)Y(c,t) as YY. Write p(x)=P(R=1∣X=x)p(x)=P(R=1\mid X=x) and mr(x)=E[Y∣X=x,R=r]m_r(x)=E[Y\mid X=x,R=r]. Consistency means that the observed comparable outcome for R=1R=1 corresponds to this target potential outcome. Conditional expectation gives

m(x)=E[Y∣X=x]=p(x)m1(x)+(1−p(x))m0(x).m(x)=E[Y\mid X=x]=p(x)m_1(x)+(1-p(x))m_0(x).

The data identify p(x)p(x) when the complete applicant frame is retained, and identify m1(x)m_1(x) on observed support, subject to sampling uncertainty. They do not generally identify m0(x)m_0(x). The complete-case conditional bias is

m1(x)−m(x)=(1−p(x))(m1(x)−m0(x)).m_1(x)-m(x)=(1-p(x))(m_1(x)-m_0(x)).

Thus the sign is conditional on the unseen contrast. When applicants excluded at a given xx have higher expected loss, the estimate from the observed subset is downward biased. A high overall acceptance rate can coexist with substantial bias in a subgroup having low observation probability. A zero conditional bias does not itself mean that the marginal portfolio mean represents applicants, because the distribution of XX may differ.

Selection based entirely on recorded variables can permit estimation of conditional risk even when accepted and rejected distributions of XX differ markedly. With a correctly specified conditional outcome function and adequate support, selected records may be sufficient for the conditional relationship. Integration over the applicant distribution then changes the population estimand. If the regression is misspecified, the selected-sample optimum can differ from the target-population optimum even under conditional independence.

Ignorability and overlap

In a simplified missing-outcome design, an identifying condition is conditional mean independence:

E[Y∣X=x,R=0]=E[Y∣X=x,R=1].E[Y\mid X=x,R=0]=E[Y\mid X=x,R=1].

The stronger distributional condition Y⊥R∣XY\perp R\mid X is useful for identifying other risk functionals. Rubin (1976) discusses ignorability conditions for incomplete-data inference. Conditional independence in this monograph is an explicit identifying assumption, not a conclusion from the original abstract or from a good model fit.

Positivity requires p(x)>0p(x)>0 on the target support. Deterministic eligibility rules violate it for some applicants. Strictly positive but extremely small probabilities can make estimation practically unreliable because few observed cases carry large extrapolation weights. Neither a flexible propensity model nor a highly accurate classifier of acceptance can supply outcomes in an empty region.

Conditional independence also requires that recorded predictors are available at the relevant decision point and adequately describe variables jointly influencing selection and loss. Features learned after inception, after medical referral, or from subsequent claim investigations can invalidate this premise for a prospective model. Hidden information may come from manual judgment, broker communications, inspection reports, prior losses, or applicants' private knowledge. A selection model estimated on the available covariates cannot test away an omitted factor shared by the selection and outcome processes.

Joint models and identifying restrictions

Heckman's (1979) sample-selection analysis motivates the following illustration. A continuous latent-selection specification can posit an outcome equation Y∗=XTβ+ϵY^*=X^T\beta+\epsilon and an observation indicator R=I(ZTγ+η>0)R=I(Z^T\gamma+\eta>0), with errors jointly normal conditional on the exogenous regressors, zero conditional means, and a standardized selection error. When continuous Y∗Y^* is observed for R=1R=1 under the stated specification,

E[Y∗∣X,Z,R=1]=XTβ+ρσϵϕ(ZTγ)Φ(ZTγ).E[Y^*\mid X,Z,R=1]=X^T\beta+\rho\sigma_\epsilon\frac{\phi(Z^T\gamma)}{\Phi(Z^T\gamma)}.

Here ϕ\phi and Φ\Phi denote the standard normal density and distribution functions; ρ\rho is the error correlation. The expression follows by taking the conditional expectation of the outcome error given a truncated normal selection error. Its use in insurance is an illustration of the classical two-equation selection approach, not a recommendation to impose normality on insurance losses.

A credible exclusion variable in ZZ can improve identification when it influences observation but does not directly change the target outcome after accounting for relevant predictors. Variables that look administratively convenient can be invalid. A regional office may influence both acceptance and exposure; a broker may influence customer selection and contract terms; underwriter workload may influence assessment quality. A routing instrument requires its own institutional investigation.

Binary claim outcomes call for a suitable binary response model, and nonnegative cost outcomes may require separate frequency and severity structure or a distributional alternative. A linear normal correction is not a probability model for claim incidence. Parametric identification from functional form alone is fragile: observed accepted outcomes can fit several competing joint models, each with substantially different implications for rejects.

Counterfactual outcomes and support

When a policy change defines a different action or coverage, define Y(a)Y(a) for each feasible action aa, and assume the recorded outcome for an actual action corresponds to that potential outcome. Identification of E[Y(a)]E[Y(a)] from observational data requires action-specific conditional exchangeability, positivity, and a stable target definition, rather than simply missing-at-random labels. Interference between accounts can matter in a portfolio because capital use, reinsurance exhaustion, or aggregations alter outcomes and decision constraints. These dependencies limit the interpretation of isolated individual effects.

A denied application at the original insurer has zero contractual claim expenditure for that insurer, but that realized no-contract expenditure is not the hypothetical covered claim expenditure under acceptance. Distinguishing the two values prevents an administratively convenient zero from becoming a false training label.

Bias diagnostics and sensitivity analysis

Diagnosing selection before fitting models

A defensible study begins with a complete application frame and reconciliation of transitions. At minimum, tabulate counts, application periods, channel, product, proposed terms, decision, reason for nonbinding, actual terms, exposure, and label maturity. Reconcile the final disposition and ascertain whether each application should be in the target frame. Examine accepted, declined, deferred, and withdrawn classes separately.

Compare distributions of predecision covariates using standardized differences and quantile plots. For a continuous variable with group means xˉ1\bar x_1 and xˉ0\bar x_0 and sample variances s12s_1^2 and s02s_0^2, the standardized mean difference is

xˉ1−xˉ0(s12+s02)/2.\frac{\bar x_1-\bar x_0}{\sqrt{(s_1^2+s_0^2)/2}}.

Zero pooled variance makes this statistic undefined. A low difference in measured covariates does not demonstrate exchangeability on unobserved variables. High differences show that naive portfolio results are unlikely to describe the complete frame without adjustment, but cannot identify the direction of conditional bias.

A propensity model of RR can be used diagnostically to display coverage of the covariate distribution, inspect near-zero support, and distinguish missing labels due to contract terms from those due to follow-up. Record the effective sample size for nonnegative weights wiw_i:

neff=(∑iwi)2∑iwi2.n_{eff}=\frac{(\sum_iw_i)^2}{\sum_iw_i^2}.

The ratio can be dramatically smaller than the nominal observation count. Weight clipping exchanges variance reduction for bias and potentially changes the scientific target; any restriction should be explicit.

Partial identification

Manski (1990) demonstrates nonparametric bounds using observed outcomes and outcome limits. The following specialization is direct. If the outcome satisfies l(x)≤Y≤u(x)l(x)\le Y\le u(x) for every targeted application, then

p(x)m1(x)+(1−p(x))l(x)≤m(x)≤p(x)m1(x)+(1−p(x))u(x).p(x)m_1(x)+(1-p(x))l(x)\le m(x)\le p(x)m_1(x)+(1-p(x))u(x).

The interval width is (1−p(x))(u(x)−l(x))(1-p(x))(u(x)-l(x)). No unobserved covariance, selection intercept, or individual reject labels need be estimated to establish this interval. Each endpoint is achievable within the assumption set by assigning all missing outcomes the corresponding endpoint. If further constraints connect applications or outcomes, the range may narrow and the restrictions must be stated. Integrate both endpoints against the actual target distribution of XX to bound the population mean.

A binary event has natural bounds zero and one. A fixed, enforceable aggregate limit can bound indemnity over a specified contract period if the outcome excludes other expenses and extra-contractual liabilities. An occurrence limit is not necessarily an annual aggregate cap. Unbounded severity and uncapped cost cannot receive an arbitrary physical upper bound merely for convenience. Tail modeling may produce a conditional scenario estimate but does not supply a nonparametric finite upper bound.

An identification interval concerns the population quantities compatible with observations and assumptions. A confidence interval concerns sampling variability. Both are necessary in finite samples. Sampling uncertainty around estimated endpoint values can be reported by an appropriate joint confidence procedure or resampling design respecting clustered applications and time. A narrow bootstrap interval around an imputed mean does not measure the unidentified contrast.

Synthetic example of observational equivalence

Suppose 1,000 applications are eligible for a common hypothetical coverage and horizon. Exactly 600 are observed under that coverage, with 60 realized binary events. The remaining 400 have no comparable outcomes. At the observed data level, p=0.6p=0.6 and m1=0.1m_1=0.1. The lower possible population rate is 60/1,000, or 6%; the upper rate is (60+400)/1,000, or 46%.

Unobserved event rate Observed events Hypothetical missing events Full-frame event rate
0% 60 0 6%
10% 60 40 10%
20% 60 80 14%
40% 60 160 22%
100% 60 400 46%

Every row produces the same observed data. No algorithm trained only on those rows can distinguish them without additional assumptions or information. The second row corresponds to equal mean risk between observed and missing applications in this unstratified example. The fourth row would imply a 12 percentage-point difference between the true full-frame rate and an accepted-only estimate of 10%.

For an unobserved subgroup of 400400 and a fixed indemnity of $20,000$ per event, if issuance entails full comparable exposure and all event indemnities equal the stipulated amount, mean covered loss per missing-policy exposure at a 10% event rate would be $2,000;at40; at 40% it would be \8,000. A candidate price of \4,000$ before expenses and capital charges therefore changes from apparently favorable to unfavorable under different observationally compatible conditions. This is a synthetic illustration of identification risk, not an insurance price recommendation.

Pattern-mixture sensitivity

A practical sensitivity analysis specifies the missing group's outcome distribution relative to the observed group's. For binary outcomes, consider

m0(x;δ)=logit⁡−1(logit⁡(m1(x))+δ(x)).m_0(x;\delta)=\operatorname{logit}^{-1}(\operatorname{logit}(m_1(x))+\delta(x)).

A zero log-odds shift recovers the equal-probability assumption where $0<m_1(x)<1$. Positive shifts assert greater event odds among missing applicants. At the boundary values zero and one, the formula needs additional modeling or smoothing rather than applying an undefined logit. The sensitivity parameter is not estimated from the same observed outcome sample solely by virtue of fitting this equation.

In the synthetic example, an observed probability of 0.1 corresponds to odds 1/9. Doubling those odds yields $2/11,about18.18, about 18.18% for the missing group, and a full-frame rate of approximately 13.27%. Halving them gives \1/19$, about 5.26%, and a full-frame rate of approximately 8.11%. A range containing both directions is useful where rejection reasons include administrative, capacity, and risk considerations.

For costs, a positive mean ratio can be defined where observed mean cost is positive, but equal ratios can hide large severity changes. Evaluate frequency shifts, severity shifts, and tail dependence separately, and name the target exposure measure. A sensitivity report should show whether the decision would change across plausible values, not merely publish a favored scenario.

Model evaluation and cautious alternatives

Complete-case models and outcome regression

A transparent observed-portfolio baseline is indispensable. Fit a model to comparable observed labels with variables fixed at the prediction time. An issue-stage mortality model may use evidence legitimately available at issue; an application-stage routing model cannot rely on evidence gathered only through the route it is itself choosing. Define the intended use before selecting features.

For frequency, a Poisson mean with log exposure offset may be appropriate under its assumptions; an overdispersed or zero-inflated model may be justified for different processes. For severity, the target can be conditional amount given a covered event, with treatment of censoring and policy limits. Expected pure loss per standardized exposure combines event frequency and severity under compatible definitions. A contract portfolio's unconditional mean depends on the distribution of exposure and records, not merely a probability of any claim.

Outcome-model extrapolation under missing-at-random assumptions is a conditional estimation approach. Predictions outside observed support should be marked as extrapolated and, when relevant, excluded from approval for operational use. Predictive accuracy among historical policies is useful evidence about those policies, not universal evidence about rejects.

Propensity weighting and doubly robust estimation

Suppose consistency, conditional mean independence, and positivity hold for the fixed target. The full-frame mean can be expressed as

μ=E[RYp(X)].\mu=E\left[\frac{RY}{p(X)}\right].

This follows by taking the expectation conditional on XX. In a finite dataset of nn applications, the inverse-probability-weighted mean estimates the target using observed outcomes weighted by inverse observation probabilities. When missing outcomes are represented in software, sum only over observed outcomes; do not multiply undefined values or store false observed zeros. Weighting can correct differences in the measured predictor distribution but cannot repair hidden selection conditional on those predictors.

An augmented inverse-probability form is

μ^=1n∑i=1n[m^(Xi)+Rip^(Xi)(Yi−m^(Xi))],\hat\mu=\frac1n\sum_{i=1}^n\left[\hat m(X_i)+\frac{R_i}{\hat p(X_i)}(Y_i-\hat m(X_i))\right],

where the correction term is zero for Ri=0R_i=0. Conditional on a fitted outcome model and propensity, the expected correction removes outcome-model bias if the propensity is correct; if the outcome model is correct, the correction has mean zero under the identification assumptions. Consistency and uncertainty for estimated nuisance functions require suitable regularity, sample splitting or other appropriate inference methods. Double robustness does not remove the requirement of identifying support, and it is not a guarantee of small finite-sample error.

Dudík et al. (2011) develop doubly robust policy evaluation for settings where feedback follows the selected action. Insurance applications must account for multiple contract offers and customer take-up when importing that method. The fixed-outcome estimator above and the action-value estimator below address different estimands.

Label augmentation, imputation, and generative models

Hard pseudo-labeling predicts missing outcomes and converts the prediction into a class. It makes a convenient training record but discards uncertainty; a threshold choice can amplify the historical score's assumptions. In soft pseudo-labeling, fractional weights retain a predicted event probability rather than a forced binary classification. A flexible learning algorithm can nevertheless reproduce its own initial conditional expectations in the unlabeled portion, creating apparent confidence without fresh evidence.

Ehrhardt et al. (2021) show in their method comparison that some fuzzy augmentation procedures reproduce the fitted baseline under the specified model, whereas parcelling introduces externally chosen assumptions about risk among rejects. Their empirical comparisons are context-dependent. An underwriting adaptation must preserve this distinction rather than interpreting augmentation as direct validation.

Multiple imputation propagates uncertainty within a specified imputation model and can aid estimation when assumptions are credible. Under nonignorable missingness, repeated completed datasets conditional on a single chosen sensitivity parameter do not span the unidentified parameter. An outer sensitivity analysis over alternative outcome or selection mechanisms is still necessary. Generated claims histories must be tagged as synthetic and kept separate from observed policy experience.

Generative modeling uses a joint distribution over outcomes, selection, and covariates. Unlabeled predictor distributions can inform model parameters if the assumed structure ties them to outcomes. Multiple plausible mixture decompositions can produce the same observed law. Improved held-out likelihood for the observed variables is not proof that the unseen conditional outcome law is correctly recovered. Validate auxiliary distributional predictions and make the identification assumptions explicit.

Comparing methods on their evidentiary basis

Method family Main source of identification Principal limitation
Complete-case conditional regression Conditional comparability, correct model and support Hidden selection and extrapolation
Reweighting Recorded observation probabilities and overlap Extreme weights, hidden selection
Augmented weighting Same design assumptions, either nuisance model correct under regularity Weak support, both-model error
Joint selection models Distributional and structural restrictions Weak or invalid exclusions, model dependence
Pseudo-labeling Predictions or judgment supplied by modeler Self-confirmation, unobservable validation target
Partial identification Outcome limits and explicit restrictions Wide intervals where evidence is thin
New audit or follow-up Sampling design and distinct new evidence Cost, incomplete response, proxy mismeasurement

The table is a methodological comparison. It makes no unconditional ranking of predictive performance. The appropriate choice is dictated by the target, availability of external evidence, and the consequences of acting on unsupported extrapolation.

Comparing historical and proposed policies

Let an underwriting policy π(a∣x)\pi(a\mid x) assign probabilities to feasible actions, b(a∣x)b(a\mid x) be the historical action probability, and G(a)G(a) denote a defined reward, including contract margins or other decision objectives. In a simplified one-step setting with comparable observed reward GiG_i for actual action AiA_i, the value is

V(π)=EX[∑aπ(a∣X)E[G(a)∣X]].V(\pi)=E_X\left[\sum_a\pi(a\mid X)E[G(a)\mid X]\right].

With conditional exchangeability and overlap for the chosen action support, importance weighting uses π(Ai∣Xi)/b(Ai∣Xi)\pi(A_i\mid X_i)/b(A_i\mid X_i) times the observed reward. An augmented version uses outcome estimates for each feasible action plus an importance-weighted correction for the actual action. This follows the general contextual decision setting of Dudík et al. (2011), while insurance introduces distinctive timing and measurement constraints.

A decline reward might include acquisition costs and no contractual loss; the accepted reward may include premium, expenses, claims and the effect of cancellations. The actual reward for an unaccepted applicant must not be substituted for the unobserved reward under an acceptance action. New offer prices affect take-up. A policy that changes terms and eligibility changes the mix of risks and may alter future application demand. The one-step value is at most a model of a specified period and environment; long-run market equilibrium requires additional assumptions.

Policy comparisons should use the same access and risk definitions. An apparently more profitable policy may achieve that result by offering less coverage or rejecting more cases. The acceptance rate, binding rate, expected contribution, available coverage, tail risk, and uncertainty need to be analyzed separately. A capital capacity constraint can make two individually favorable decisions jointly infeasible when exposures are concentrated.

Evaluator heterogeneity and contracted comparison

Lakkaraju et al. (2017) propose contraction-based evaluation using differing human acceptance tendencies under assignment and agreement conditions. The method exploits the information from a relatively permissive decision-maker while limiting comparison to appropriate acceptance rates and controlling the effect of cases without labels. Its validity depends on the conditions underlying the comparison, not merely the existence of underwriters with different thresholds.

An insurance implementation should assess assignment rules, office and broker specialization, authority levels, expected contract terms, and time. Random allocation of cases within a narrowly defined homogeneous queue may be plausible; allocation by industry expertise or hazard level is not. Tests of recorded covariate balance can reveal problems but cannot prove equality of unobserved information. A candidate policy that admits many cases previously rejected even by the more permissive evaluator remains weakly supported. The result is a restricted comparison with a stated domain, not a reconstructed reject population.

Validation that reflects deployment

Separate application dates from valuation dates. Build time-based development and test cohorts with sufficient outcome maturity. Keep all records of a person, corporate group or insured property together when splitting data. Avoid training on later versions of applications that reveal information unavailable to the original decision rule. For a new workflow, retain the historical rule and versioned candidate model, with decisions and action support recorded.

Use proper scoring metrics for their applicable population: log loss and Brier score for binary probability predictions; calibration by risk strata and expected versus actual experience; exposure-adjusted frequency and severity error; out-of-time loss ratio or margin as appropriate. A discrimination metric such as AUC depends on the case mix and class-specific distributions. Accepted-only AUC is therefore conditional on selection. Calculate it only using observed labels in the declared evaluation population, and report uncertainty and sample size. Transfer to the entire frame requires a justified design.

Technical backtesting against imputed labels primarily measures alignment with the imputation model. It cannot substitute for independent observations. A comparison of policy value should state feasible actions, common support, reward maturity, and the sensitivity of incremental outcomes in unsupported regions.

Rights, privacy and practical validation

Additional information without manufacturing ground truth

The preferred evidence source is direct comparable experience collected under a lawful design. Where that cannot be obtained, additional medical or property assessment, externally recorded events, and post-issue review provide narrower evidence. A death record can support mortality follow-up when linkage and authority are sound. A third-party property loss record may describe physical damage without proving coverage under the insurer's proposed terms. An observed event remains a proxy for an insured claim when contract and exposure differ.

Design a stratified random audit among applicants lawfully eligible for assessment, with known inclusion probabilities. Record sample-selection time and completion status; track refusal, withdrawal and failed record retrieval. Compare completed versus selected cases and adjust or bound nonresponse under explicitly stated assumptions. Separate targeted anomaly investigations from probability-sampled monitoring. Targeted reviews are useful for discovering defects, but their outcomes cannot be mechanically treated as representative.

A hypothetical audit of 200 applicants with inclusion probability one half in one stratum and one tenth in another would require different design weights for estimating stratum-combined totals. If response probabilities vary, the product of selection and response probabilities may be needed for complete ascertainment under a justified response model. A sample assignment is not the same as an outcome observation.

For life underwriting, an additional medical assessment can produce an alternative risk-class judgment rather than realized mortality. The distinction is central to Seeman and Herzog's (2024) mortality-slippage methodology. Random holdouts and retrospective reviews should therefore be understood as validations of portions of the process. When subsequent experience matures, compare proxy-based estimates against directly observed loss rates under compatible exposure definitions.

Ethical scope of prospective designs

Experimentation may generate knowledge by altering decisions or data collection, but an insurer's obligation to customers and compliance with binding coverage rules take priority over research convenience. Randomization can be confined to genuinely equivalent, lawful routes or additional assessment that does not reduce promised coverage. An offer pilot designed to admit applicants within a preapproved range requires eligibility, consent or other appropriate legal conditions, capital limits, monitoring, and procedures for adverse events. No person should lose an entitlement or receive inferior treatment merely to generate labels.

A field experiment requires a prespecified estimand. Randomizing an offer measures the consequences of offer assignment including potential non-acceptance; it does not automatically identify the effect of binding coverage among all applicants. If offer probabilities are validly randomized and issuance differs, intention-to-treat analysis remains meaningful for the offer-level decision. Effects of actual coverage may require additional assumptions and comparison groups. Record protocol deviations and retain adverse and null results.

Shadow-mode testing is an operationally conservative first step. A model can score all applicants without making actual decisions and reveal disagreement with existing rules. The missing outcomes for historically declined applicants remain missing. Shadow evaluation identifies decision disagreement and data quality, but cannot confirm the proposed rule's realized loss results under expansion.

United States professional practice and regulatory context

The Actuarial Standards Board's adopted ASOP No. 12 provides a US professional practice reference for risk classification and material limitations; the posted revision originally took effect May 1, 2006. Applying its reasoning to selective labels calls for attention to the relevance of the experience supporting classification. The existence of a historic approval pattern is insufficient evidence, by itself, that a risk characteristic is predictive or permissible.

The ASB's (2019) ASOP No. 56 applies to covered actuarial modeling work commenced on or after October 1, 2020. Its sections on intended model purpose, data, assumptions, testing, and validation align with recording the eligible population, selection effects, and differences between observed and simulated outcomes. Professional standards operate within their defined applicability; they do not settle each state's substantive insurance law.

The NAIC describes adoption of accelerated underwriting regulatory guidance on August 14, 2024 in its insurance topic record. This guidance addresses external data and predictive analytics in life underwriting. The study draws a general validation implication: documented inputs, classification grounds, and assessment of unfair discrimination are material to reviewing missing-label mechanisms. NAIC guidance and model materials must be distinguished from binding law adopted or applied by a particular state. This monograph does not assert uniform enactment.

European Union legal context and international supervision

Regulation (EU) 2016/679, applicable since May 25, 2018, creates legal constraints on the use of personal information. The General Data Protection Regulation establishes processing principles in Article 5 and special-category conditions in Article 9; Article 22 concerns decisions based solely on automated processing that produce legal or similarly significant effects, subject to specified exceptions and safeguards. Article 89 addresses safeguards for research uses. An actuarial desire for improved labels does not establish a lawful basis or an applicable exception for processing health information.

EIOPA's August 6, 2025 AI governance announcement identifies AI uses for risk assessment and pricing in life and health insurance as high-risk under the EU AI Act and explains that the accompanying sectoral Opinion excludes high-risk or prohibited uses to avoid overlap. This is a bounded account of the authority's published position; this study does not provide a consolidated analysis of application dates, subsequent legislative amendments, or particular classification exceptions. A deployment assessment must consult the applicable legal text and jurisdictional requirements at the time of use.

Internationally, the IAIS (2025) AI supervision application paper emphasizes traceability, documentation of performance, data provenance, and risk-based governance. An application paper supplies supervisory interpretive guidance rather than itself constituting domestic legislation. Its relevance is particularly clear when records combine direct observations, external estimates, and synthetic labels: the origin and purpose of each field should remain reconstructible.

Fairness when outcomes are missing

The distribution of access can be measured using dispositions and offers across the complete application frame, where collection and analysis of relevant group variables is lawful. Equal future insured outcome rates, false-negative rates, or calibration among all applicants generally cannot be determined from accepted outcomes alone.

For example, suppose two groups have observed event rates of 10%, with observation fractions of 80% and 40%. Without additional assumptions their full-frame binary risk intervals are respectively 8% to 28% and 4% to 64%. The same observed calibration does not establish equal population risk, and the wider interval represents an important difference in the available evidence. A fairness auditor should report observed access disparities, the distribution of label coverage, support limitations, and outcome-based bounds separately. Legal conclusions require the relevant statutory framework and evidence; a statistical disparity is not itself a complete legal determination.

Training a model to match a previous decline decision can measure consistency with historical judgment but cannot establish risk validity. Historical judgments may encode justified expertise, operational limitations, or unjustified disparity. A fair process requires actual evidence, review of adverse decisions, and an opportunity for correction where required. Imputed losses must not be communicated as facts about an individual.

Analytical synthesis, evidence gaps and reproducibility

The hierarchy of supported conclusions

The scientific difficulty of underwriting reject inference is not primarily computational. Its first constraint is the definition of the potential outcome; its second is the relationship between selection and that outcome; its third is the limited support for alternative actions. Increasing model complexity addresses estimation within an assumed structure. It does not generate independent evidence that the structure is valid.

One implication is that decision archives and exposure archives should be treated as distinct but connected objects. The first records choices and their institutional basis. The second records the experience generated by those choices and coverage terms. A complete archive of all underwriting decisions gives the denominator for an access analysis and the probability of observation. A complete archive of issued-policy claims provides direct insured experience. Their joint value does not include unattained counterfactual outcomes unless the research design supplies additional identification.

A second implication concerns the relationship between predictive risk and offered action. Expected loss under a fixed contract and expected profit from offering a changing contract are different estimands. A commercial insurer may classify prospects correctly in hazard terms yet compare decisions incorrectly because the price, deductible, or limit affects the composition and result of the bound portfolio. An insurer optimizing expected margin may nevertheless lack evidence about applicants outside the historical support.

A third implication concerns the representation of uncertainty. A conventional predictive interval reflects uncertainty of a model for future events under its fitted assumptions. A sensitivity or partial-identification interval reflects uncertainty about which data-generating explanation is compatible with the incomplete evidence. In actuarial decision review, the latter can be much wider than the former. Combining the two without distinction produces a number with an unclear scientific interpretation.

A fourth implication is that uncertainty has an operational geography. Some classes may have stable and extensive experience, some may be observed through occasional manual exceptions, and others may be entirely outside historic authority. Report support boundaries and prohibit stronger claims beyond them. Where the proposed decision is insensitive to a broad credible range of outcomes, evidence may be sufficient for that narrowly defined decision even without point identification. Where the plausible range changes the decision or materially affects access, further investigation is justified.

These arguments are analytical applications of established statistical concepts to insurance institutions. They do not purport to establish priority for a new mathematical estimator or theory.

Application across insurance lines

In life insurance, the target can be mortality over a specified duration after a hypothetical issue date. An external death indicator does not identify the death-benefit obligation where exclusions and policy persistence differ. Missing follow-up and differential linkage must be separately evaluated.

In property insurance, site characteristics, insured value, deductible, protection measures, and geographically shared hazards affect the risk and its observation. A declined site may still experience a well-documented catastrophe, which can be useful evidence of hazard. Policy-specific coverage and conditional damages determine the actual indemnity. Geographic and event clustering complicates precision estimates.

In commercial casualty, long reporting and development periods mean that historical cohorts differ in valuation maturity. Claim absence at a common extraction date cannot be treated as absence of eventual liability. A decision model for a new industry class may extrapolate beyond the relevant exposure and legal environment. Restrict comparisons by jurisdiction, coverage trigger, and contract terms.

In cyber insurance, requirements for protective controls can change between application and binding; threats and technology also evolve quickly. Self-reported incidents, vendor telemetry, and claims reflect separate observation processes. A rejected organization's incident may be observable outside insurance but subject to different reporting incentives. The candidate model must be evaluated under the actual assumed security conditions.

These are applications of the same identification problem under different contracts and observation horizons, not empirical assertions about relative market-wide prevalence.

A reproducible research protocol

An institution conducting a reject-inference study should maintain a versioned data dictionary and an immutable or appropriately traceable analytical snapshot within lawful retention constraints. Specify the target population, action, terms, outcome horizon and valuation date. The application frame should include disposition histories and timestamps, rather than a retrospective binary decision alone.

Construct explicit observation flags with reasons. Reconcile row counts across application, offer, policy, exposure, and claims tables. Document the criteria for a comparable outcome and how exposure is standardized. Keep raw observation, external proxy, inferred label, and synthetic example separate. For records linked across systems, document match rates, validation findings, unresolved errors, and the consequences for the target estimand.

Fit and evaluate at least an interpretable observed-only baseline. Inspect support, sample weighting, missingness, time trends and groups. Test candidate models only against the correct target. Specify identifying assumptions and their business rationale, then report counterexamples and sensitivity results sufficient to show where recommendations change. If a model depends on an exclusion variable, identify the actual assignment mechanism and possible violations.

Evaluate operational decision outcomes and customer consequences together with financial metrics. Predefine stop, revise, and additional-evidence criteria. Assign responsibility for documentation and meaningful review. This guidance is consistent with ASB (2019) model risk guidance and the IAIS (2025) traceability expectations, within their distinct professional and supervisory scopes.

Remaining research needs

Published insurance evidence demonstrates the phenomenon and several ways of gaining partial additional information. It does not establish the missing label distribution for a representative global applicant pool. Access to insurer-level raw data, standardized outcome definitions and linkage across coverage decisions would support stronger conclusions. This monograph has no access to proprietary rejected-applicant outcomes, underlying corporate insurer records, or the Munich Re seriatim monitoring data. It does not replicate the cited studies or estimate causal effects from their aggregate reports.

Needed research includes multi-insurer longitudinal applicant cohorts with lawful access and clearly harmonized contract definitions; randomized or plausibly exogenous variation within suitable decision processes; the comparison of label-coverage gaps under actual uptake, exposure and censoring; and independently evaluated sensitivity procedures for severity tails and aggregate insurance outcomes. Such work must report the cost, limits, governance, and uncertainty of evidence creation as well as empirical accuracy.

External data can narrow an identification interval only insofar as a justified measurement relationship connects it to the target. Evidence that a third party knows an applicant's later event is not, on its own, evidence that the original insurer knows its own hypothetical net claim or profit. The distinction persists across every proposed technology.

Conclusion

Insurance underwriting reject inference concerns the outcomes that historical underwriting and customer decisions prevented from being observed under a specified insurance regime. The completeness of the applicant record cannot solve the absence of comparable exposure and outcomes. The effect is especially consequential when a new model expands, contracts, or modifies the existing selection boundary.

Acceptance-conditioned experience can estimate a well-defined conditional relationship under suitable identification and support conditions. Outside those conditions it establishes a more limited answer. Weighting and augmented estimation can adjust observed-data imbalance, while joint models, imputation and sensitivity scenarios formalize assumptions about unobserved outcomes. Partial identification makes the remaining uncertainty measurable without pretending to recover individual truth. Additional audits and carefully governed prospective designs can improve the evidence, subject to legal and ethical conditions.

The scientific adequacy of reject inference depends on the fit between the target outcome, the contract and exposure mechanism, available evidence, identification assumptions and intended decision. Evaluations should report label provenance, application-level support, uncertainty about excluded risks, and consequences for access and financial performance. Where the evidence remains inadequate, the proper scientific result is a bounded conclusion or an explicit gap. This is a substantive finding, not an invitation to substitute invented data for insurance experience.

References

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

  • Lakkaraju, H., Kleinberg, J., Leskovec, J., Ludwig, J., & Mullainathan, S. (2017). The selective labels problem: Evaluating algorithmic predictions in the presence of unobservables. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 275–284. DOI: 10.1145/3097983.3098066. Selective labels paper

  • Ehrhardt, A., Biernacki, C., Vandewalle, V., Heinrich, P., & Beben, S. (2021). Reject inference methods in credit scoring. Journal of Applied Statistics, 48(13–15), 2734–2754. DOI: 10.1080/02664763.2021.1929090. Reject inference methods

  • Manski, C. F. (1990). Nonparametric bounds on treatment effects. American Economic Review, 80(2), 319–323. Nonparametric treatment-effect bounds

  • Dudík, M., Langford, J., & Li, L. (2011). Doubly robust policy evaluation and learning. Proceedings of the 28th International Conference on Machine Learning, 1097–1104. Doubly robust policy evaluation

  • Actuarial Standards Board. (2005). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). Effective May 1, 2006; posted version includes subsequent cross-reference changes. ASOP No. 12

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

  • National Association of Insurance Commissioners. (n.d.). Accelerated Underwriting. Insurance Topics. Retrieved November 19, 2026. Accelerated underwriting overview

  • Seeman, L., & Herzog, K. (2024, August). Accelerated Underwriting: Mortality Slippage Study and Monitoring Best Practices. Product Matters!, Society of Actuaries. Mortality slippage study

  • International Association of Insurance Supervisors. (2025, July). Application Paper on the supervision of artificial intelligence. AI supervision application paper

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

  • European Parliament & Council of the European Union. (2016). Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88. Applies from May 25, 2018. GDPR official text

  • Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica, 47(1), 153–161. DOI: 10.2307/1912352. Sample selection bias