Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Selection Bias and Exposure Representativeness in Insurance Data

Abstract. Insurance data are generated through an organized sequence of market participation, eligibility, underwriting, contract acceptance, exposure accrual, renewal, incident occurrence, reporting, settlement, recording, and sometimes voluntary provision of behavioral information. Every transition can change the relationship between the observations available to a researcher and the population for which an insurer seeks predictions or estimates. This monograph develops a comprehensive analysis of selection bias and exposure representativeness in insurance statistics, distinguishing population composition from exposure measurement, selection into coverage from selection into an observed dataset, and predictive transport from economic adverse selection. The analysis synthesizes survey-sampling theory, missing-data methods, sample-selection econometrics, covariate-shift learning, insurance-specific telematics studies, claims development research, and actuarial data-quality guidance. Formal identification conditions are derived for poststratification, inclusion-probability weighting, and target-risk evaluation. Three transparent constructed examples quantify how selection changes claim-frequency estimates and how correction depends on measured selection determinants and positivity. A practical audit sequence connects estimand definition, population-frame reconciliation, selection-path documentation, diagnostics, correction, uncertainty, and governance. The principal conclusion is that representativeness is a relationship between a dataset, an exposure measure, a target estimand, and a period of use. Statistical performance measured within a selected portfolio alone cannot establish predictive validity for a broader insurance population.

Keywords: Selection bias; exposure representativeness; insurance data; earned exposure; underwriting selection; renewal bias; claims reporting; telematics; covariate shift; poststratification; inverse probability weighting; calibration; missing data; transportability; actuarial modeling; model governance.

Introduction

The database ordinarily available to an insurance analyst is a record of completed processes. An applicant must appear in a distribution channel, supply enough information to be considered, satisfy an insurer's acceptance criteria, receive an acceptable offer, purchase coverage, generate covered time or another exposure unit, and be retained in the administrative system. A claim must occur and become recognizable under contractual terms; it may then require notification, adjustment, and development before its information is used in model training. The resulting records possess organizational integrity while containing several distinct opportunities for statistical selection.

A scientific assessment begins by defining the quantity of interest independently of the dataset. In the language of Lundberg, Johnson, and Stewart (2021), an estimand specifies the population and quantity a statistical analysis aims to learn. For an insurer, possible estimands include the mean covered-claim frequency of currently insured vehicle-years, the frequency expected among all eligible applicants under a proposed contract, ultimate losses per earned exposure for the forthcoming underwriting year, and the predictive loss for policyholders offered voluntary telematics. These quantities are different even when the same predictive model is fitted to each.

Selection bias here denotes distortion of the desired statistical quantity when the observed distribution differs from its target through mechanisms associated with relevant outcomes or covariates. Representativeness denotes the adequacy of the observed exposure distribution for that specified target and purpose. A dataset can be representative of renewed policies and unrepresentative of initially insured policies; both statements can be true. A portfolio can also be complete relative to the insurer's current in-force book while offering limited evidence about rejected applicants or competing carriers. Thus the population label has to accompany every claim about completeness.

The methodological foundations are longstanding. Horvitz and Thompson (1952) established the logic of estimation under unequal sampling probabilities. Heckman (1979) investigated the specification consequences of nonrandom sample selection. The missing-data framework of Little and Rubin (2019) explains why the observation process must enter inference. More recent work on covariate-shift weighting examines prediction when the distribution of covariates changes across populations. This monograph applies these established results to insurance exposure and defines their boundaries explicitly. It presents original analytical comparisons and constructed calculations; it claims neither a new named theory nor new empirical identification from proprietary insurer data.

Portfolio and Exposure Sampling

Observation units, risk units, and target frames

A policy record, a policyholder, an insured object, a vehicle-year, a life-year, a covered property value, and a claim are alternative units. They must be matched to the question being estimated. If a motor portfolio covers one vehicle for three months and another for twelve, the two contracts contribute different amounts of covered time. A claim-frequency estimate using policy counts as a denominator has a different interpretation from claims per vehicle-year. The Casualty Actuarial Society's GLM guidance, by Goldburd and colleagues (2020), distinguishes count models with log-exposure offsets from frequency models using exposure weights under relevant distributional assumptions.

Let i index policy-period segments. Denote earned exposure by e_i > 0, recorded claims by n_i, covariates available at the exposure period by x_i, and a population-membership indicator by t_i. A basic covered-claim frequency for target T is

λT=∑i∈TE[Ni∣IT]∑i∈Tei,\lambda_T=\frac{\sum_{i\in T}\mathbb{E}[N_i\mid\mathcal{I}_T]}{\sum_{i\in T} e_i},

where the information set and covered-claim definition are fixed before estimation. The conventional empirical ratio uses actual observed claims in the numerator. A model for counts might posit N_i conditional on x_i and e_i as Poisson with mean e_i \lambda(x_i), recognizing that overdispersion, dependence, reporting delays, and heterogeneity often require richer specifications.

The target need not consist of contracts historically observed. When predicting prospective eligible applicants, the analyst requires an exposure distribution for a hypothetical insured population and an explicit definition of coverage and price. Exposure among people who never purchased insurance may be conceptually definable as driving distance or property-time, but an insurable-claim outcome under a counterfactual contract may not be observed. The lack of outcomes cannot be repaired by rescaling policy-year counts. An additional assumption connects risk behavior and claims under the proposed coverage to historical experience.

The record-generation process can be described as a sequence: eligible universe → reachable distribution population → applicant population → offered population → bound contracts → earned exposure → retained exposure → reported claims → developed outcomes → analysis extract. Each arrow corresponds to a denominator transition. These transitions should be recorded with calendar dates, eligibility decisions, channel identifiers, insurance product, contract terms, and loss-definition changes.

A constructed composition example

Consider a deliberately artificial target comprising 10,000 equal policy-years: 7,000 low-risk years with expected claim frequency 0.05 and 3,000 high-risk years with expected frequency 0.20. The target expectation is 350 + 600 = 950 claims, or 0.095 claims per year. Suppose the available sample consists of 900 low-risk and 100 high-risk policy-years with the same risk-specific frequencies. Its expected claim count is 45 + 20 = 65, and its sample frequency is 0.065. The observed rate understates the target by 0.030 claims per year, or approximately 31.6% of the target rate.

The discrepancy arises entirely from composition in this construction. It does not demonstrate erroneous recordkeeping or a defect in a fitted regression. Applying known target class shares yields 0.70(0.05) + 0.30(0.20) = 0.095. Equivalent expansion weights are 7,000/900 for each sampled low-risk year and 3,000/100 for each sampled high-risk year. The example illustrates the central distinction: a model can estimate class-specific frequency perfectly while an unweighted portfolio mean still fails for another target population.

The constructed counts are expectations, not observed insurer claims, and they assume that class-specific risks are stable and measured correctly across populations. If class labels omit important predictors of participation and outcome, reweighting by the two observed classes cannot ensure recovery of the desired target.

Other denominators and sources of compositional change

For property insurance, insured values, location-risk years, insured replacement costs, policy limits, and deductibles play different roles. Premium-volume weights are not interchangeable with hazard-exposure weights because premium reflects price, selection, insured value, loading, and contract coverage. For life insurance, counts of lives, sums assured, policy durations, and cause-specific risk sets require a clear denominator. For health insurance, member-months and service utilization are not equivalent when benefit design changes demand. Cross-country insurance-market totals introduce variations in coverage definitions, enrollment systems, and data completeness; direct pooling can hide denominator incompatibility.

Temporal composition matters as much as cross-sectional composition. A book written before a regulatory change, flood, telematics rollout, or economic shock may be statistically complete for its own period and a poor approximation of the forthcoming risk environment. The appropriate selection diagnosis separates population drift, exposure-definition drift, and shifts in conditional outcome mechanisms before applying corrective weights.

Underwriting Selection

Conditioning on acceptance and purchase

Underwriting affects which risks become observable. Let A indicate an application, U an acceptable underwriting decision, B a bound policy, and S an observation in the training data. The common dataset conditions on A=1, U=1, and B=1, potentially followed by retention and administrative selection. Risk factors such as prior claims, occupancy, driving patterns, or health information can influence both acceptance and expected loss. Price and product design can influence the decision to purchase. A regression fitted solely to accepted business estimates relationships conditional on this selection path.

Economic adverse selection concerns differences in insurance purchase or contract choice associated with information asymmetry and risk. Statistical sample selection concerns the distribution and observation mechanism of records. They may interact without being equivalent. The observed book may be statistically selected by distribution-channel reach and insurer acceptance even where no private-information adverse selection is present. Conversely, private information may affect purchase while a particular within-book predictive estimator remains suitable for its precise insured-population target.

The Heckman selection analysis demonstrates how nonrandom observation can enter an econometric relationship as a specification problem. Its familiar two-stage parametric correction requires a credible selection equation, functional-form assumptions, and suitable information about selection determinants; a proposed exclusion restriction should influence observation without independently determining the conditional outcome in the modeled relationship. Insurance settings rarely warrant the automatic use of this estimator solely because an underwriting gate exists. A decision tree or neural network cannot identify unobserved rejected risks merely by increasing model flexibility.

A causal diagram further clarifies the issue. Let X contain observed risk and access factors, U an unobserved determinant of loss, P the offered price, and S the inclusion decision. If both X and U influence S, conditioning on S can change associations among their causes; when selection is a common effect of relevant causes of exposure and outcome, spurious associations may emerge. Berkson's original discussion of selected hospital observations is a classic warning against transferring associations from an institutionally selected sample to a broader population. It is an analogy about selection structure, not empirical evidence about an insurance portfolio.

Positivity and the underwriting boundary

Suppose every applicant with a particular prohibited risk characteristic is rejected, so the probability of observed coverage in that stratum is zero. Data from accepted policies provide no direct observations of covered outcomes there. No finite weight can recover an outcome that has no support in the dataset. This is a positivity failure, and its practical meaning is important: a statistically attractive extrapolation can be governed entirely by model assumptions.

Even positive but extremely small selection probabilities create unstable expansion factors. An underrepresented segment carrying a small portion of observed exposure can dominate estimates when weighted to its intended target share. Diagnostics therefore need both selection probabilities and the actual supported covariate ranges, including joint combinations of variables. Two samples with matching marginal age and territory distributions may differ sharply in the joint distribution of age, territory, occupation, property construction, and deductible.

The appropriate audit asks whether the target is the insurer's accepted book, its quoted applicant population, all licensed vehicles in a market, or a future offering under revised underwriting. The first target may be estimable from an accurately retained internal portfolio. The later targets generally need data about applications, quotes, market exposure, third-party populations, or explicit identification assumptions. Rejected applicants' claims under hypothetical contracts cannot be observed retrospectively as ordinary insurance claims.

Record availability and model interpretation

Feature selection can compound underwriting selection. Variables measured only after policy issue—for example an inspection triggered by a high premium or a claim investigation—are missing by design for other risks. Treating their missing values as if they were random attributes can contaminate both training and validation. If an analyst also includes a future policy outcome in a predictor, leakage can appear as unrealistically good prediction. Thus collection time, intended decision time, and eligibility for feature observation belong in the data dictionary.

Risk classification methods may remain useful despite limited portability. Within-book predictions can answer an insurer's current pricing or reserving question if exposure, outcome maturation, and distribution are aligned. A sound conclusion specifies this domain of validity. It does not universalize the current book into the insurance market as a whole.

Renewal and Claims Selection

Renewal as longitudinal selection

Policies enter and leave insurance books for many reasons: price, switching options, claims experience, perceived service, liquidity, business closure, lapse, nonrenewal, and changes in eligibility. The renewing portfolio is therefore a dynamically selected subset of previous contracts. Retention is itself a relevant business outcome and can be statistically dependent on future claim cost through common observed and unobserved determinants.

Consider a second constructed example. A starting population contains 600 low-risk and 400 high-risk equal-exposure policies. Their expected claim frequencies are 0.05 and 0.20, yielding an initial expected frequency of (30 + 80)/1,000 = 0.110. If 90% of low-risk policies but only 50% of high-risk policies renew, the subsequent in-force population contains 540 low-risk and 200 high-risk equal exposures. Its expected frequency becomes (27 + 40)/740 ≈ 0.09054. This is a valid expectation for the constructed renewal book and a biased description of the original cohort's expectation if it is substituted for that estimand.

The assumed renewal probabilities here are illustrative and could reverse in another market; the direction of the bias is empirical, not universal. Life-insurance lapse analysis also requires cohort, duration, product, and contract-benefit distinctions. A study of life-insurance lapse determinants by Reck, Schupp, and Reuß (2023) offers empirical modeling of termination behavior in a European portfolio, but its predictive relationships are particular to the insurer, countries, product definitions, and available features.

For longitudinal analysis, define C_it as continued observation of risk i at interval t. An inverse-probability-of-censoring factor can have the form

wit=∏k=1t1P(Cik=1∣Ci,k−1=1,Hik),w_{it}=\prod_{k=1}^{t}\frac{1}{P(C_{ik}=1\mid C_{i,k-1}=1,H_{ik})},

where H_ik is measured history available before interval k. The weighted estimator targets the original risk set under sequential conditional exchangeability, correctly specified continuation probabilities, positivity, and consistent outcome/exposure definitions. These are substantive conditions. A price increase influenced by a prior claim creates time-varying confounding and makes a naive model that ignores history unreliable. The review by Seaman and White (2013) describes both the possibilities and the practical sensitivity of inverse-probability methods.

Reported claims, closed claims, and maturity

An observed claim extract is rarely the same object as all ultimately incurred covered events. Claims may be incurred but unreported at valuation; reported but insufficiently developed; open with uncertain ultimate severity; or processed only after administrative delay. Inclusion into a closed-claim-only training set can be associated with settlement duration and severity. The sample of closed claims can overrepresent rapidly settled claims, especially when the observation window is short.

Research by Bücher and Rosenstock (2023) on micro-level outstanding claim counts models accident-to-reporting delays and claims whose existence is not yet known. Work on processing backlogs and development patterns further separates incident occurrence, reporting, and processing. These mechanisms mean that claim-file inclusion and claim-outcome maturity are separate selection variables. A portfolio denominator of earned policy-years does not by itself correct a numerator that excludes immature or unreported claims.

A clean analysis fixes an accident or occurrence cohort, a valuation as-of date, a defined reporting/development horizon, and the claim status rule. For prospective predictions, validation uses a chronologically later cohort with sufficient outcome maturity and an explicitly frozen information cutoff. If a later claim payment is used to label an earlier insurance period, all features in the training row must still be restricted to those observable at the actual decision date.

Why complete-case analysis may fail

Complete-case selection can remove policies with missing device information, claims with absent reserve histories, or contracts with interrupted administrative records. If completeness depends on a cause of losses, the complete-case subset can be compositionally distorted. Missing at random, conditional on measured history, can justify certain likelihood, imputation, or weighting procedures; missing not at random calls for additional sensitivity assumptions. The comprehensive Little–Rubin missing-data treatment makes explicit the connection between missingness mechanisms and inferential claims.

Imputation of a missing numerical field, inverse weighting for absent records, and prediction of unreported claims solve different problems. One reconstructs a variable value, one adjusts inclusion, and one estimates an outcome not yet observable at valuation. A scientific workflow should not relabel all three as the same operation.

Telematics Participation

Voluntary observation and behavior

Telematics illustrates how a newly measured feature can create a new selection mechanism. Drivers who join usage-based programs may differ in risk, privacy preferences, expected discounts, familiarity with devices, geography, and driving patterns. The program may also change driving behavior while monitored, so a difference between participants and nonparticipants can reflect both selection into monitoring and an effect of monitoring. The analyst must distinguish an observed association from either a population-wide risk distribution or a causal behavioral response.

In a primary actuarial study, Peiris, Jeong, Kim, and Lee (2024) examine integration of traditional and telematics data using calibration weights and explicitly consider favorable selection in telematics availability. Their empirical demonstration uses a synthetic telematics dataset as described by the authors; it supports methodological feasibility under its design rather than an insurer-independent universal improvement percentage.

The empirical Duval, Boucher, and Pigeon study examines the amount of driving data useful for claim classification in a particular North American insurer dataset. The reported approximately three months or 4,000 km sufficiency finding is tied to its classification objective, data, and study setting, and should not be generalized automatically to every insurer or to severity prediction. Jin and Vasserman's monitoring study examines opt-in behavior and incentives in a U.S. auto-insurance program; its publicly listed January 2026 working-paper revision provides empirical evidence that selection and behavior may both matter.

An additional study of driving behavior and coverage choice reports no statistically significant evidence for the specific coverage-selection relationship it tests, while identifying driving variables associated with risk and modeling selection into a pay-as-you-drive contract. This finding is important precisely because the direction and presence of selection effects must be established by study design rather than presumed. Favorable selection into telematics is a plausible and sometimes documented mechanism, not an unavoidable property of every program.

A constructed voluntary-enrollment example

Return to the artificial target with 70% low-risk and 30% high-risk exposure. Suppose 20% of low-risk but only 5% of high-risk years enroll in monitoring. Before any monitoring-induced behavior change, the enrolled population is proportional to 0.70 × 0.20 = 0.14 low-risk and 0.30 × 0.05 = 0.015 high-risk units. Its normalized high-risk share is 0.015/0.155 ≈ 9.68%. Under unchanged within-group frequencies of 0.05 and 0.20, expected enrolled frequency is approximately 0.06452, against 0.095 for the full eligible population. The apparent discrepancy is a product of self-selection in the constructed example.

An analyst who has both participants' outcomes and complete traditional-covariate information for nonparticipants may estimate enrollment propensities and calibrate observed outcomes to a target, subject to measured selection and overlap. If monitoring changes driving, an outcome measured while monitored is different from a counterfactual unmonitored outcome. Standard reweighting does not remove this treatment difference. Identifying a program effect requires an appropriate comparative design and assumptions about selection, behavior, and time.

Device missingness and multiple clocks

Telematics data introduce several observation clocks: policy coverage, device activation, trip recording, claim occurrence, upload, and model extraction. A driver might have a full year of insured exposure but only two months of valid telematics coverage. Treating twelve months of claims as predicted by two months of unchanging driving behavior assumes temporal stability. A Society of Actuaries telematics report documents precisely such a mismatch between longer policy-claims histories and shorter periods of available telematics measurements, noting an implicit representativeness assumption about driving style.

Device installation failures, delayed activation, opt-out, connectivity gaps, and consent changes may be associated with location, income, occupation, trip frequency, or driving behavior. Completeness indicators should be retained as study variables, together with time coverage and measurement quality. Restricting a model to drivers with abundant recorded trips can inadvertently turn the target population into frequent drivers who continuously provide data.

A portable telematics model therefore needs a clearly stated target: enrolled drivers under monitoring, currently eligible customers, or all drivers under a proposed mandatory device regime. These are different observational and counterfactual settings. Cross-market transfer also requires attention to road network, vehicle fleet, trip measurement, seasonal weather, policy terms, and privacy constraints.

Reweighting and Calibration

Identification of a target expectation

Let S=1 indicate membership in the observed modeling sample, T denote a specified target exposure population, X be recorded preselection covariates, and Y a well-defined outcome normalized to the appropriate exposure. Under conditional outcome stability, E[Y|X=x,S=1]=E_T[Y|X=x], and target support wherever target density is positive, the target mean is identified as

μT=∫E[Y∣X=x,S=1] dFT(x).\mu_T=\int E[Y\mid X=x,S=1]\,dF_T(x).

Equivalently, when the observed and target X-distributions admit densities p_S and p_T, one may use the density ratio r(x)=p_T(x)/p_S(x) to write \mu_T=E_S[r(X)Y]. Under the stated normalization and common support, E_S[r(X)]=1. This identity depends on conditional outcome stability; equality of measured covariate marginals alone is not enough. Shimodaira (2000) gives an influential treatment of such weighting in predictive likelihood inference.

In actual insurance records, total exposure often varies across rows. If the target's covariate distribution is measured by earned exposure, a weighted frequency estimate has the form

λ^T=∑i:Si=1wiNi∑i:Si=1wiei.\widehat{\lambda}_T=\frac{\sum_{i:S_i=1}w_iN_i}{\sum_{i:S_i=1}w_ie_i}.

The weights must expand exposure to the target distribution, not simply policy counts. When estimating loss cost, claim amounts, limits, deductibles, and inflation adjustment must have corresponding definitions. Incorrectly choosing the unit of weighting can produce a statistically balanced policy sample that remains unbalanced in the units relevant to the risk estimate.

Selection-probability weights and calibration

When inclusion probabilities \pi_i are known for a probability sample from a finite target frame, Horvitz–Thompson estimation constructs weighted totals from inclusion reciprocals. When selection is modeled from measured variables, inverse-probability weighting estimates inclusion under assumptions about the response mechanism. The two approaches have related algebra but distinct evidentiary foundations: design probabilities derive from an actual sampling scheme, whereas estimated participation probabilities require a trustworthy observation model.

Calibration weighting exploits known population totals. In the Deville–Särndal calibration formulation, survey weights are adjusted subject to constraints linking weighted sample auxiliary variables with known target totals. For an insurance portfolio, reliable totals might include exposure by territory, vehicle class, contract start quarter, or distribution channel. Calibration can reduce visible imbalance but cannot guarantee correction for outcomes related to omitted predictors or unmeasured selection determinants.

Propensity-score reasoning is associated with Rosenbaum and Rubin (1983). Their original setting is treatment assignment and causal effects, not automatically the transfer of insurance prediction models. A modeled probability of sample inclusion can help balance observed covariates, but causal-treatment interpretations require their own assumptions. An insurer should state whether a propensity score models enrollment, renewal, missingness, or a treatment; these event indicators cannot be interchanged.

Variance and effective sample size

Reweighting alters precision. In the first constructed example, weights are 7,000/900 ≈ 7.7778 for 900 low-risk sampled units and 3,000/100 = 30 for 100 high-risk units. The weighted total is 10,000. A simple Kish-style effective sample size diagnostic is

neff=(∑iwi)2∑iwi2=10,0002900(7,000/900)2+100(30)2≈692.31.n_{\mathrm{eff}}=\frac{(\sum_iw_i)^2}{\sum_iw_i^2} =\frac{10{,}000^2}{900(7{,}000/900)^2+100(30)^2} \approx692.31.

This quantity is a descriptive weight-dispersion proxy, not the actual information content under clustered claims, outcome heterogeneity, or an estimated weight model. It demonstrates the cost of correcting strong imbalance: the nominal sample has 1,000 records, yet the highly weighted small stratum reduces this particular precision proxy. Any actual confidence interval should account for the sampling or portfolio design, weight estimation, clustering, and outcome variance.

A stable weighting protocol examines propensity distributions by group, joint-support diagnostics, extreme weights, weighted covariate balance, total target-exposure reconciliation, and sensitivity to trimming. Trimming reduces variance by changing the effective target or imposing modeling bias; its effect should be reported rather than silently absorbed. Bootstrap or other uncertainty procedures must repeat estimation of the weights where appropriate.

Target-aligned validation

Randomly splitting a selected dataset produces two samples governed by roughly the same selection process. An excellent random-holdout score therefore measures performance under that sample distribution. Under covariate shift, Sugiyama, Krauledat, and Müller (2007) show why ordinary cross-validation may misrepresent target risk and develop importance-weighted cross-validation under explicit assumptions.

For insurance models, a target-aligned validation plan specifies outcome maturity and exposure, uses chronological splits when predicting future years, keeps policies or insureds grouped to prevent leakage across splits, and evaluates target-weighted loss along with group-level calibration. Validation should report which strata have no support in training, how weights were estimated, and whether outcome relationships remained stable. Reweighting held-out data does not magically reveal the outcomes of never-observed applicants.

Standardized residuals and calibration errors should be reported on meaningful exposure bases. A class with apparently good unweighted calibration may contribute little to the observed portfolio but materially to the intended target. An aggregate predictive score cannot substitute for this class-specific assessment.

Population Transport and Limits

Statistical and causal transport

Transportability concerns which knowledge learned in one setting remains informative in another. Cole and Stuart (2010) demonstrate inverse-selection weighting for generalizing trial evidence to a defined target, subject to measured determinants of selection. In the causal setting, Pearl and Bareinboim (2011) formalize differences between environments using selection diagrams. Their work gives rigorous foundations for distinguishing changes in covariate composition from changes in causal or outcome-generating mechanisms.

For insurance predictions, three forms of change demand separate attention. Covariate shift changes P(X), potentially making reweighting appropriate if P(Y|X) remains stable. Conditional-outcome shift changes P(Y|X), perhaps through inflation, repair practices, weather, policy design, legal definition, or claim-reporting changes. Selection-induced missingness changes which joint outcomes are observed and may remain unidentified from the observed sample alone. Some problems combine all three.

A predictive transport assessment therefore states the source and destination environment, target population, decision date, exposure denominator, observed and unobserved features, selection pathways, structural stability assumptions, and external data required. A model trained in one country or product cannot acquire worldwide validity through a single global weighting scheme unless the specified conditions are supported.

Partial identification and transparent sensitivity

When some exposure is absent without trustworthy response propensities, an honest result may be an interval rather than a point estimate. Suppose a target population consists of 80% fully observed equal risk units with a binary claim indicator mean of 0.08, and 20% units whose outcome distribution is entirely unknown. Without restrictions beyond the binary range from zero to one, the target indicator mean lies in [0.8×0.08, 0.8×0.08+0.2], or [0.064, 0.264]. This is an exact worst-case bound for the constructed setup. If outcomes are nonnegative and unbounded claim counts or monetary losses, an upper bound of this form requires a separately justified maximum outcome; the binary illustration must not be transferred unchanged.

The strength of external information determines whether this interval can be narrowed. Additional exposure covariates, audited claims-reconciliation studies, representative follow-up samples, or credible monotonicity assumptions may narrow uncertainty. Each extra assumption needs an evidence entry. The width of a justified interval is an informative scientific finding about what the records can support.

Governance and regulatory boundaries

Actuarial standards provide obligations relevant to selecting and evaluating data. U.S. ASOP No. 23, Data Quality addresses data selection, review, use, reliance, and disclosure for actuaries subject to its scope; its adopted version is listed as effective April 30, 2017. It is a professional standard in its stated jurisdictional context, not a universally binding global regulation.

The European Insurance and Occupational Pensions Authority's 2025 opinion on AI governance clarifies supervisory expectations for AI use under relevant insurance-sector legislation, with a risk-based approach and data-governance emphasis. The opinion's scope and its explicit distinction from high-risk or prohibited AI Act categories must be respected; it is not a general law establishing any single weighting method. In the United States, the NAIC's 2023 model bulletin describes governance and documentation expectations for insurer AI applications. A model bulletin's adoption by NAIC members does not make identical language independently binding in every state; applicable state enactments, bulletins, and insurance laws require jurisdiction-specific verification.

Representativeness is also relevant to fairness. A model that performs well for majority exposure and poorly for an underrepresented group may have uneven error or consumer impacts. Yet representativeness and fairness are distinct questions: a statistical adjustment can improve target calibration while legally permissible rating factors, privacy requirements, and discriminatory effects still require separate examination. Scientific assessment documents both inferential and governance consequences without asserting that one statistical criterion resolves all regulatory questions.

An integrated insurance-data audit

A reproducible audit has a logical sequence. First, register the estimand: the proposed population, exposure unit, claims/outcomes, decision date, horizon, and contract basis. Second, reconstruct the data path with counts and exposure totals at each participation and observation gate. Third, distinguish units absent from the frame, present with missing predictors, present with immature outcomes, and present with observed outcomes. Fourth, identify measured determinants of inclusion and compare their distributions in source and target. Fifth, examine positivity and explicitly identify unsupported subpopulations. Sixth, choose estimators linked to assumptions—direct standardization, design weighting, modeled inverse weighting, calibration, missing-data modeling, or partial bounds—and state why each applies. Seventh, perform sensitivity analyses that test omitted selection determinants, time dependence, alternative target weights, delayed claims, and extreme weights. Eighth, validate on appropriately matured and temporally separated data, reporting subgroup metrics, calibration, and uncertainty. Finally, preserve the population definition and data-quality limitations with each model version and use case.

The order matters. Selection correction chosen before the target population is defined can optimize the wrong quantity. A sophisticated algorithm cannot compensate for an unstated denominator. Conversely, even a simple frequency estimator can be scientifically adequate when the insurer possesses complete exposure and sufficiently mature outcomes for the exact in-force population being studied.

Original analytical synthesis

The foregoing literature supports a unified interpretation of insurance representativeness through four documented objects: an exposure-defined target; a chain of inclusion mechanisms; a time-indexed outcome observation process; and a transport assumption linking observed and intended environments. This is an analytical organization of established results, not a proposed new named model or doctrine.

The first implication is that representativeness is estimand-relative. Recording all contracts currently in force establishes census-like coverage of that book, but not of previously rejected risks or a market-wide customer population. The second implication is that selection corrections must conserve relevant exposure units: matching policy counts can leave vehicle-years, value at risk, or covered member-months distorted. The third implication is that the observation clock and the risk clock must be reconciled: claim maturity and telematics availability can select records even when policy exposure itself is complete. The fourth implication is that prediction and causal explanation require different transport claims; voluntary monitoring may change behavior, while underwriting may suppress support for rejected risk classes. The fifth implication is that validation can only establish accuracy over populations for which observable, comparable outcomes and a credible target sampling design exist.

These implications sharpen actuarial interpretation of model success. A reported score is a conditional statement about an observed target, measurement regime, and evaluation period. Its broader relevance must be established by population comparison and mechanism analysis. The appropriate response to absent support is documented uncertainty, targeted new data collection, and narrower claims of validity.

Conclusion

Insurance statistics inherit the selection structure of the institutions that produce their records. Portfolio entry, underwriting acceptance, contract purchase, renewal, voluntary telematics participation, claim reporting, settlement, and extraction are each potential boundaries between an observed sample and an intended population. Selection bias becomes scientifically assessable when a target estimand and exposure denominator are specified before estimation, and when inclusion and outcome observation mechanisms are made explicit.

The mathematical analysis shows that poststratification and weighting recover target quantities only under appropriate overlap and conditional stability. Calibration to known totals improves measured balance while leaving unobserved selection effects as an evidentiary limitation. Extreme weights sacrifice precision, complete-case analysis can induce distortion, and random validation within a selected dataset cannot demonstrate transport to a different population. Constructed examples illustrate these consequences transparently without importing unsupported empirical claims.

The insurance-specific literature confirms that telematics availability, monitoring participation, lapse, claims development, and differing data clocks have concrete implications for risk measurement. The resulting scientific requirement is clear: every insurer-facing estimate or predictive performance statement should identify the population, exposure unit, time horizon, collection regime, observation mechanism, and assumptions supporting its extension beyond the sampled records. This research dossier supplies an integrated interpretive and methodological synthesis of those requirements. It does not claim a new theoretical system, an original insurer-data experiment, or an independently established universal magnitude of selection bias.

References

  • Berkson, J. (1946). Limitations of the application of fourfold table analysis to hospital data. Biometrics, 2(3), 47–53. PubMed record.

  • Bücher, A., & Rosenstock, A. (2023). Micro-level prediction of outstanding claim counts based on novel mixture models and neural networks. European Actuarial Journal, 13, 55–90. Full-text research article.

  • Goldburd, M., Khare, A., Tevet, D., & Guller, D. (2020). Generalized Linear Models for Insurance Rating. Casualty Actuarial Society. CAS source document.

  • Cole, S. R., & Stuart, E. A. (2010). Generalizing evidence from randomized clinical trials to target populations: The ACTG 320 trial. American Journal of Epidemiology, 172(1), 107–115. https://doi.org/10.1093/aje/kwq084.

  • Deville, J.-C., & Särndal, C.-E. (1992). Calibration estimators in survey sampling. Journal of the American Statistical Association, 87(418), 376–382. https://doi.org/10.1080/01621459.1992.10475217.

  • Duval, F., Boucher, J.-P., & Pigeon, M. (2022). How much telematics information do insurers need for claim classification? North American Actuarial Journal, 26(4), 570–590. https://doi.org/10.1080/10920277.2021.2022499.

  • European Insurance and Occupational Pensions Authority. (2025). Opinion on Artificial Intelligence Governance and Risk Management (August 6). Official opinion.

  • Geyer, A., Kremslehner, D., & Muermann, A. (2020). Asymmetric information in automobile insurance: Evidence from driving behavior. Journal of Risk and Insurance, 87(4), 969–995. https://doi.org/10.1111/jori.12279.

  • Heckman, J. J. (1977). Sample Selection Bias as a Specification Error (with an Application to the Estimation of Labor Supply Functions) (NBER Working Paper No. 172). NBER working paper.

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

  • Horvitz, D. G., & Thompson, D. J. (1952). A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association, 47(260), 663–685. https://doi.org/10.1080/01621459.1952.10483446.

  • Jin, Y., & Vasserman, S. (2021; revised January 2026). Buying Data from Consumers: The Impact of Monitoring in U.S. Auto Insurance (NBER Working Paper No. 29096). NBER working paper.

  • Lindskog, F., & Wüthrich, M. V. (2025). Eliciting claims development patterns and costs hidden in backlogs. European Actuarial Journal, 15, 667–705. https://doi.org/10.1007/s13385-025-00431-w.

  • Little, R. J. A., & Rubin, D. B. (2019). Statistical Analysis with Missing Data (3rd ed.). Wiley. https://doi.org/10.1002/9781119482260.

  • Lundberg, I., Johnson, R., & Stewart, B. M. (2021). What is your estimand? Defining the target quantity connects statistical evidence to theory. American Sociological Review, 86(3), 532–565. https://doi.org/10.1177/00031224211004187.

  • National Association of Insurance Commissioners. (2023). Model Bulletin: Use of Artificial Intelligence Systems by Insurers (adopted December 4). NAIC adopted bulletin.

  • Peiris, H., Jeong, H., Kim, J.-K., & Lee, H. (2024). Integration of traditional and telematics data for efficient insurance claims prediction. ASTIN Bulletin, 54(2), 263–279. https://doi.org/10.1017/asb.2024.6.

  • Pearl, J., & Bareinboim, E. (2011). Transportability of causal and statistical relations: A formal approach. Proceedings of the AAAI Conference on Artificial Intelligence, 25(1), 247–254. https://doi.org/10.1609/aaai.v25i1.7861.

  • Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41–55. https://doi.org/10.1093/biomet/70.1.41.

  • Seaman, S. R., & White, I. R. (2013). Review of inverse probability weighting for dealing with missing data. Statistical Methods in Medical Research, 22(3), 278–295. https://doi.org/10.1177/0962280210395740.

  • Shimodaira, H. (2000). Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2), 227–244. https://doi.org/10.1016/S0378-3758(00)00115-4.

  • Society of Actuaries Research Institute. (2021). Telematics Car Driving Data Analytics. Research report.

  • Sugiyama, M., Krauledat, M., & Müller, K.-R. (2007). Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8, 985–1005. Full published article.

  • Actuarial Standards Board. (2016; effective April 30, 2017). Actuarial Standard of Practice No. 23: Data Quality. Official adopted standard.

  • Reck, L., Schupp, J., & Reuß, A. (2023). Identifying the determinants of lapse rates in life insurance: An automated Lasso approach. European Actuarial Journal, 13, 541–569. Publisher article.