Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Life Insurance Mortality Pricing and Anti-Selection

Abstract. Life insurance mortality pricing is the prospective valuation of death-contingent benefits for an insured portfolio whose composition is shaped by underwriting, product design, applicant decisions, distribution, and post-issue behavior. Anti-selection arises when one party uses information or risk characteristics to make choices that financially disadvantage the insurance arrangement. Its actuarial significance cannot be determined from aggregate death rates alone: the relevant process includes entry, acceptance, classification, benefit selection, persistency, and mortality conditional on each of those events. This monograph develops an integrated examination of mortality classes, experience-based assumptions, selection mechanisms, level-premium calculation, risk loading, model validation, and long-horizon sensitivity. It distinguishes theoretical possibility from empirical evidence and pricing assumptions from statutory reserves. An explicitly synthetic ten-year term model demonstrates how selective purchase and differential lapse behavior alter premium adequacy. A two-risk illustration yields a 12.27% increase in first-year accepted-life mortality under specified purchase probabilities; the extended projection yields a materially different required premium when selection is modeled directly. These figures are analytical demonstrations, not observed industry rates. The original synthesis is a practical chain of reasoning connecting selection-conditioned exposure, identifiable risk drivers, contract cash flows, credibility, and model governance. Existing actuarial and economic methods are applied without claiming a new theory. The documentary foundation includes original actuarial standards, Society of Actuaries experience and valuation materials, the 2026 NAIC Valuation Manual, and peer-reviewed research on asymmetric information. Jurisdictional differences and information-access limits are explicit.

Keywords: Life insurance; mortality pricing; anti-selection; adverse selection; risk classification; underwriting; selection effects; preferred risks; term insurance; premium adequacy; mortality assumptions; credibility; expected claims; exposure measurement; persistency; mortality improvement; principle-based reserving; actuarial governance.

Introduction

The actuarial price of a death benefit depends on whose lives enter the pool, how long their coverage remains in force, and when deaths occur. Each question is interdependent. Application and acceptance change the composition of the pool. Contract terms influence the amount of protection purchased and retained. Underwriting can reveal some information while leaving other mortality-relevant information private. Economic conditions and medical developments alter both expected mortality and demand. A forecast that begins with a population mortality table and ends with a premium without explicitly addressing these transitions assumes away a central component of the business being priced.

The governing research question is therefore: how should the future death-claim costs of a life insurance contract be calculated when the distribution of risks among actual policyholders differs from the distribution among prospective applicants? The corresponding methodological question concerns identifiability. Differences in observed mortality may reflect risk selection, age, smoking status, underwriting methods, coverage, duration, calendar period, lapse, claims reporting, or sampling fluctuation. A defensible analysis establishes which mechanisms the evidence can distinguish before interpreting the residual difference as anti-selection.

In the United States, the Actuarial Standards Board's life and annuity pricing standard (ASOP No. 54, 2018) addresses pricing models, assumptions, sensitivities, and controls. Its risk classification standard (ASOP No. 12, 2005/2011 posting) expressly defines adverse selection and addresses its potential effect when testing risk classifications. These professional sources make the problem part of ordinary actuarial practice. They do not prescribe a universal level of premium or establish that anti-selection must be present in every product.

The distinction matters empirically. The Cawley and Philipson life-insurance study (1999) reported results difficult to reconcile with simple canonical predictions of hidden-risk models. The Harris and Yelowitz purchaser study (2014), using linked life insurance holdings and mortality records, found no compelling evidence of adverse selection in its broad age cohort. Conversely, Hendren's research on insurance rejections (2013) documented private information relevant to exclusion from several nongroup insurance markets, including life insurance. The results address different populations and decisions; the diversity of findings is substantive evidence against applying any single market-wide selection adjustment indiscriminately.

This dossier adopts three research procedures. First, it derives mortality and premium relationships from conditional probability and prospective cash-flow identities, declaring all modeling choices. Second, it compares what cited studies and standards actually establish, distinguishing transferable methodological insights from product-specific findings. Third, it constructs transparent synthetic examples, recalculates the resulting quantities, and identifies the limitations of such examples. All numerical illustrations are original calculations from disclosed assumptions. No confidential portfolio data, fitted proprietary mortality model, or external peer review is asserted.

The subject is prospective mortality pricing and the financial implications of selection. General premium and reserve mathematics, the historical construction of mortality tables, annuitant longevity selection, medical underwriting, and insurance-market welfare each have a wider independent literature. Those subjects are brought into the analysis when they determine the death-benefit premium, while their comprehensive treatment belongs to their separate research objects.

Mortality Classes and Risk Exposure

The insured risk as a conditional distribution

Let X denote age and other observable classification characteristics at issue, U denote additional unobserved or imperfectly observed mortality characteristics, A the event of acceptance, B the event of binding a policy, and I_t the event that coverage remains in force immediately before policy year t. Let D_t identify a death during the year. For a death benefit, the pertinent one-year mortality rate is

qtinsured(x)=Pr⁡(Dt=1∣X=x,A=1,B=1,It=1).q_t^{\mathrm{insured}}(x)=\Pr(D_t=1\mid X=x,A=1,B=1,I_t=1).

The conditioning is essential. A general population rate estimates deaths among a different collection of individuals and may use different age definitions, observation conventions, period years, and data quality. An underwriting class is an operational subset of accepted applicants, not a natural mortality law. A policy-duration table reflects selection at issue, survival, and the subsequent composition of those who remain in force.

A useful representation is

qtinsured(x)=∫qt(x,u) f(u∣x,A,B,It) du.q_t^{\mathrm{insured}}(x)=\int q_t(x,u)\,f(u\mid x,A,B,I_t)\,du.

Under this representation, the insured mortality assumption changes if either the underlying risk for a given type changes or the relative representation of risk types changes. Even if every individual retains the same underlying hazard, reweighting alone can move aggregate mortality. Conversely, a change in population mortality need not translate one-for-one into the mortality of a selected insured portfolio.

The Society of Actuaries' 2015 Valuation Basic Tables furnish published distinctions by sex, smoking status, age convention, and relative-risk categories. Their existence is evidence that a single universal mortality multiplier is too coarse for many valuation tasks. The publication includes later corrections and an updated report; the relevant table edition and age convention must be preserved in any implemented calculation. The 2019 individual-life experience report examines experience for 2012–2019 by key policy characteristics, providing a basis for studying how actual insured experience compares with reference mortality.

Exposure, deaths, and denominators

Mortality estimation begins by defining exposure. An annual policy count, an annualized policy-year exposure, a life-year exposure, and a face-amount-weighted exposure answer different questions. If d_{it} is the indicator of a covered death and e_{it} the exposure fraction at risk, a crude count-weighted rate is

q^=∑i,tdit∑i,teit,\widehat q=\frac{\sum_{i,t}d_{it}}{\sum_{i,t}e_{it}},

under the specified exposure convention and an appropriate interpretation of the within-year event probability. If face amounts F_i differ markedly, the expected claim amount is driven by ∑Fieitqit\sum F_i e_{it}q_{it}, not merely by the frequency of deaths. The corresponding experience ratio should use actual claim amounts and their expected amounts under aligned coverage definitions.

Observed deaths must be tied to a precise event interval and to policies actually exposed to that risk. A claim reported months after death may belong to the earlier exposure period; an applicant who never bound a policy cannot be treated as a covered survivor. Policy loans, riders, multiple policies on one life, guaranteed-issue segments, contestability, exclusions, and benefit changes may require adjustments to the risk and claim definitions. Treating all claims as identical unit events can misstate the financial consequence of anti-selection concentrated in high-face-amount cases.

Four time scales deserve independent control: attained age, duration since issue, calendar year, and underwriting cohort. Attained age approximates biological aging; duration captures entry selection and early post-issue behavior; calendar year measures common shocks and trend; cohort differentiates underwriting practices and market environments at inception. A single pooled observed-to-expected ratio can conceal offsetting patterns across all four. The SOA individual-life study archive shows why the valuation and experience questions are studied using distinct datasets and periodic updates.

Selection at issue and its gradual reorganization

An issue-selection effect is often observed as lower early-duration mortality among medically underwritten lives relative to a comparator. That empirical relationship must be attributed to the underwriting process, its target market, and the comparator used. As time passes, the originally selected cohort ages, information acquired at issue becomes less predictive, and the relative survival and lapse of different risks change the cohort's composition. The shape of the duration effect therefore cannot be transported blindly from one carrier's underwriting cohort to another.

Suppose two latent types have death probabilities q_L and q_H, with q_H greater than q_L. Their proportions before selection are 1-w and w. If the respective probabilities of buying and being issued a specified contract are a_L and a_H, the higher-risk proportion among policyholders is

wissued=waHwaH+(1−w)aL.w^{\mathrm{issued}}=\frac{w a_H}{w a_H+(1-w)a_L}.

The mortality probability at issue is

qissued=(1−wissued)qL+wissuedqH.q^{\mathrm{issued}}=(1-w^{\mathrm{issued}})q_L+w^{\mathrm{issued}}q_H.

This is a conditional-probability identity. The inequalities a_H>a_L and q_H>q_L imply a higher expected insured mortality than under unselected participation, conditional on the model's assumptions. Neither inequality is guaranteed in real data. For example, more extensive financial planning or risk aversion among low-risk people could raise a_L; underwriting might lower a_H; premiums might induce different demand responses; and contract design may redirect higher-risk applicants into different products.

This distinction is developed in the broader Einav–Finkelstein survey (2011), which describes evidence of both adverse and advantageous selection across insurance markets. The finding does not assign a direction of selection to a specific insurer. The general implication is methodological: participation is an endogenous filter whose direction and strength require evidence.

Policy amount and concentration

The composition of policy counts can remain stable while the composition of insured sums changes. Consider a class with 1,000 policies and an average face amount of $100,000. If 50 relatively high-risk policies hold $500,000 each, the monetary risk weighted by mortality can exceed what a count-based test suggests. Financial underwriting, maximum face amounts, medical evidence thresholds, reinsurance, and retention limits therefore affect both expected claims and the capital distribution.

An adequate mortality-pricing model evaluates at least two distributions: the mortality of lives and the mortality of dollars at risk. The precise financial metric depends on whether the benefit is constant, graded, decreasing, adjustable, or contingent on other events. For term death coverage, a face-weighted death distribution is often central; for universal life, the relevant net amount at risk may differ materially from the stated face amount.

The analytical boundary of comparability

A mortality comparison has evidentiary force only after comparable exposures, periods, classifications, coverages, and reporting rules are specified. Comparing a recently issued preferred nonsmoker term cohort with a long-duration population table confounds many factors. Comparing companies without controlling for their distribution channels and underwriting standards introduces additional differences. The need to examine classification effects is also recognized in the ASOP No. 12 testing guidance.

The first operational result is thus an exposure statement: every claim about selection-adjusted mortality must identify the population entering the denominator and the economic quantity being predicted. The exposure statement is part of the scientific finding, not a technical afterthought.

Mortality Assumptions and Underwriting

Building the baseline mortality basis

A prospective assumption is constructed from sources with known purpose and limitations. A credible portfolio experience study may be the best starting point; a recent industry insured-life table may supply external structure; a general population table may be informative for trend or reasonableness checks when adjusted for insured selection. The appropriate hierarchy is evidence-dependent.

The 2015 VBT report and tables provide historical insured-life reference rates, not forecasts of every present-day applicant. The 2019 SOA mortality experience report offers a later observational window but remains subject to its years, participant composition, and reporting conventions. The 2026 mortality improvement guidance addresses assumptions for AG38/VM-20 valuation and should not be silently transcribed as an automatically appropriate commercial pricing trend.

The distinction between expected future pricing assumptions and regulatory valuation assumptions is particularly important. In the United States, the NAIC's principle-based reserving overview describes reserves that use prescribed minimum requirements alongside scenario-based estimates reflecting relevant company experience. The 2026 Valuation Manual, VM-20, section 9.C, specifies mortality segments, anticipated experience, credibility, margins, and company- or industry-based assumptions. Those are obligations for the specified valuation framework; a pricing actuary still needs to develop the premium on its own economic and contractual basis.

Underwriting as an information transformation

Underwriting changes the information set available to the insurer and the distribution of lives accepted at each price. In a fully underwritten process, medical records, examination findings, prescription information where lawful, smoking status, occupation, financial need, and application disclosures may be relevant. In accelerated or simplified underwriting, the mix of information, referral thresholds, accepted volume, and potential misclassification may differ. A pricing assumption requires documentation of the exact underwriting rule, its effective date, the distribution channel, applicant characteristics, referral outcomes, and any exception policy.

Underwriting cannot be represented adequately by a binary accepted/rejected flag if acceptance simultaneously determines a standard, preferred, rated, postponed, or declined outcome. Different classes have different expected mortality and may attract different benefit amounts. Classification rules can also change demand: an applicant receiving a high premium may buy less coverage, defer purchase, seek a competitor, or buy a different contract. The insurer observes a selected sample after these decisions.

The 2026 NAIC VM-20 underwriting-process provisions explicitly consider whether processes are similar when pooling mortality experience, including distribution systems and target markets. Similarity may require third-party research or retrospective statistical demonstration. The technical implication for pricing is that mortality experience from one underwriting process is transportable only after differences relevant to selection and classification have been addressed.

Credibility and heterogeneity

Let q^company\hat q_{\mathrm{company}} be a company experience estimate and qreferenceq_{\mathrm{reference}} a relevant external estimate. A generic credibility-weighted estimate is

q∗=Zq^company+(1−Z)qreference,q^*=Z\hat q_{\mathrm{company}}+(1-Z)q_{\mathrm{reference}},

where ZZ lies between zero and one. The weight is meaningful only with an explained credibility method, justified reference data, compatible risk definitions, and consideration of dependence between the two datasets. It is not a license to average incomparable quantities.

The Actuarial Standards Board's credibility procedures (ASOP No. 25, effective 2014) emphasize relevant data, homogeneity, predictive value, and practical selection of procedures. For life mortality, heterogeneity is often strong enough that a high overall number of deaths does not yield equally credible estimates for each small risk class, issue age, duration, or benefit tier. Combining groups may improve statistical precision while introducing systematic bias. A justified segmentation scheme must balance these effects.

A Poisson approximation for death counts D_k in a reasonably narrow exposure cell k uses

Dk∼Poisson(Ekλk),Var(Dk)≈Ekλk,D_k\sim\mathrm{Poisson}(E_k\lambda_k), \qquad \mathrm{Var}(D_k)\approx E_k\lambda_k,

with EkE_k measured as appropriate life-years and λk\lambda_k a hazard rate. This approximation is distinct from a Bernoulli formulation using annual death probabilities. For annual probabilities, the exact binomial structure may be preferable; for heterogeneous face amounts, claim amounts require additional variance terms. Apparent confidence derived from a large pooled E_k can be misleading when exposures are correlated or risk mixtures change.

Mortality improvement and deterioration

Mortality assumptions can evolve with attained age and calendar period. A simplified period adjustment is

qx+t,tproj=qx+tbase∏s=1t(1−mx+s,s),q_{x+t,t}^{\mathrm{proj}} =q_{x+t}^{\mathrm{base}}\prod_{s=1}^{t}(1-m_{x+s,s}),

where mm represents an assumed annual proportional improvement, which may be negative in periods of deterioration. The quantity is model-defined, and the same life must not be given improvement twice through both the base table and the adjustment. Emerging evidence may require explicit treatment of mortality shocks, changed causes of death, emerging therapies, or shifts in underwriting mix.

The SOA 2026 improvement-scale release underscores that mortality-improvement assumptions change as evidence becomes available. A pricing model should document data vintage, projection period, the relationship between insured and general population experience, and the uncertainty that arises when long-term extrapolation extends far beyond observed data.

Underwriting rules, legal admissibility, and fairness

A predictor can be statistically associated with mortality and still be legally restricted, unsuitable for the product, or inconsistent with approved classification practice. ASOP No. 12 identifies applicable law among the considerations for selecting risk characteristics. In the United States, the Department of Health and Human Services' GINA research guidance states that the federal statute's protections for health coverage do not extend to life, disability, or long-term care insurance. This federal boundary does not eliminate state-specific restrictions or other applicable privacy and insurance laws.

Within the European Union, the General Data Protection Regulation provides specific rules for special-category data and certain solely automated decisions. A mortality predictor's lawful use depends on the applicable legal basis and safeguards. The Solvency II life underwriting provisions address the solvency capital treatment of mortality, longevity, lapse, and catastrophe risks; they are not a blanket authorization for every pricing variable.

Pricing governance should therefore evaluate statistical usefulness, causal or predictive relevance, error rates, distributional effects, and legal admissibility together. A model's ability to rank observed mortality within historical insured lives does not establish that using the same variables for future applicants will produce a stable or permissible risk classification.

Selection and Anti-Selection

Meaning and possible channels

The ASOP No. 12 definition of adverse selection links the concept to a party's use of known or suspected information that disadvantages the insurance system. In life insurance, selection may occur before application, during coverage choice, after a rating decision, at conversion, when increasing benefits, or when deciding whether to lapse.

A higher-risk person may have greater willingness to insure, yet may also be declined or offered a less attractive price. A healthier individual may be more likely to buy protection because of financial planning, family circumstances, and risk aversion. Conditional on insurer information, the sign of the relationship between private risk and coverage purchase is consequently an empirical matter.

The classic Rothschild–Stiglitz insurance model (1976) established how information asymmetry can alter competitive contract design and equilibrium. Its stylized two-type framework is useful for deriving selection mechanisms but does not prove their magnitude in present-day individual life markets. The Einav–Finkelstein review (2011) situates such models in a wider empirical literature, including advantageous selection.

Adverse and advantageous selection as conditional results

Selection is adverse for a particular insurer when the distribution of expected covered claims among participants, given contractual price and risk controls, shifts unfavorably relative to the pricing expectation. It may be advantageous when participation selects lower expected claims. The definition is conditional on the cost basis, contract, and time horizon. A product may exhibit adverse selection in high sums insured but advantageous selection in participation counts. A conversion option may experience an unfavorable mortality mix even if the original term block showed a favorable issue selection effect.

Cross-product evidence requires care. Finkelstein and Poterba's UK annuity contract analysis (2004) found mortality sorting on some annuity features but not a substantive difference by annuity size. The direction of mortality risk for annuity payouts is generally opposite the direction for a death benefit: longer survival increases the expected payout under a life annuity. Their work illustrates contract-feature identification, not a numerical death-benefit loading. Similarly, Fang, Keane, and Silverman's Medigap selection study (2008) demonstrates how advantageous sorting can be associated with financial and cognitive characteristics in another insurance market; its quantitative findings are not transferrable directly to life premiums.

Evidence from individual life and rejection samples

Cawley and Philipson (1999) investigated life insurance ownership, perceived risk, pricing, and quantity. The publisher's empirical article record reports findings that do not conform to several simple theoretical predictions. Harris and Yelowitz (2014) examined a representative sample with linked death records and found no compelling aggregate adverse selection in their investigated population. Neither analysis establishes the absence of every form of anti-selection across products, policy durations, or face-amount segments.

Hendren (2013) studied populations at risk of rejection and identified private information among rejected applicants that may help explain missing markets. The dependent decision here is different from simple ownership among accepted lives. A coherent interpretation of these studies is that application, eligibility, underwriting, and issued-contract outcomes operate as sequential selection stages. Observational findings conditioned at one stage cannot be generalized automatically to another.

Chiappori and Salanié's conditional testing framework (2000) used French automobile insurance data and reported no evidence of asymmetric information in that sample. Its main transferable contribution is the methodological importance of conditioning on observed risk classification and contract details when testing risk-coverage correlations. The sample itself is not evidence about individual life mortality.

An explicitly synthetic entry-selection calculation

Consider a notional one-year term pool containing 85% lower-risk applicants and 15% higher-risk applicants. Assume annual death probabilities of 0.001 and 0.004, respectively. If all applicants purchase, the mixture death probability is

qall=0.85(0.001)+0.15(0.004)=0.001450.q_{\mathrm{all}}=0.85(0.001)+0.15(0.004)=0.001450.

Now assume, solely for illustration, that 60% of lower-risk and 90% of higher-risk applicants purchase and are accepted at the modeled contract terms. The high-risk share among policies is

wissued=0.15(0.90)0.85(0.60)+0.15(0.90)=0.2093023256.w_{\mathrm{issued}} =\frac{0.15(0.90)}{0.85(0.60)+0.15(0.90)} =0.2093023256.

Thus,

qissued=(1−wissued)(0.001)+wissued(0.004)=0.001627907.q_{\mathrm{issued}} =(1-w_{\mathrm{issued}})(0.001) +w_{\mathrm{issued}}(0.004) =0.001627907.

The proportional increase over unselected mortality is 0.001627907/0.001450−1=12.26945%0.001627907/0.001450-1=12.26945\%. For a $500,000 end-of-year death benefit discounted at an assumed annual effective rate of 5%, the single-year mortality present value is approximately $690.48 before selection and $775.19 after selection. The difference is $84.72 per issued policy on this deliberately simplified basis. Premium taxes, expenses, capital charges, and behavioral feedback have not been included in these one-year values.

The calculation proves a conditional result of the specified assumptions. It does not estimate the probability of adverse selection in any national market, give an underwriting class a scientific mortality rate, or justify a general 12.27% surcharge. A change in applicant acceptance or purchase behavior could reverse the result.

How selection changes after issue

Let nL,tn_{L,t} and nH,tn_{H,t} be the expected counts of in-force lower- and higher-risk policies at the beginning of year t per initially issued policy. Let qj,tq_{j,t} be the annual death probability and ℓj,t\ell_{j,t} the probability of lapse following survival in the year. An explicit state recursion is

nj,t+1=nj,t(1−qj,t)(1−ℓj,t),j∈{L,H}.n_{j,t+1}=n_{j,t}(1-q_{j,t})(1-\ell_{j,t}),\quad j\in\{L,H\}.

The high-risk proportion among remaining lives becomes

wH,t=nH,tnL,t+nH,t.w_{H,t}=\frac{n_{H,t}}{n_{L,t}+n_{H,t}}.

If higher-risk policies lapse less frequently, wH,tw_{H,t} tends to increase relative to a same-mortality scenario with equal lapse rates, although higher death rates work in the opposite direction by removing high-risk lives through claims. Which effect dominates depends on their magnitudes. Selective lapses can therefore make future mortality differ from the originally issued mix even without any change in the mortality rates of the underlying types.

Convertible term business raises an additional option effect. The SOA/LIMRA term conversion study covers 2009–2023 conversion incidence and post-conversion mortality and lapse experience, showing that conversion is an observable business event warranting separate study. Conversion selection should be analyzed using the precise options, timing, premiums, and data available; it cannot be reduced to a generic term-life multiplier.

Identification rather than a residual label

A convincing empirical claim of anti-selection requires a comparison that can separate private information from observable pricing and underwriting variables. For life insurance, a possible design uses cohorts of applicants and issued policies, underwriting scores or classes, offered prices, decisions to accept, face amount, subsequent death experience, and lapses. Ideally, it also records applicants who declined or were rejected. Such data are frequently commercially sensitive, incomplete, or available only after a long follow-up.

A simple positive association between face amount and death risk among policyholders is insufficient. Larger policies may undergo more intense underwriting; they may be purchased at different ages, through different channels, or by households with different socioeconomic profiles. Conditioning on issued business can itself introduce selection bias because issuance depends on risk and other causes of purchase. Estimation should specify the selection stage, the identifying assumptions, the comparator population, and which selection paths remain unobserved.

Practical approaches include within-class mortality comparisons by sum insured, purchase-timing analysis, comparisons around documented underwriting-rule changes, duration analyses, propensity or selection weighting where the data support it, and survival models with time-varying covariates. Each method has specific vulnerability: hidden confounders, selection into underwriting, nonparallel trends, competing events, or unstable extrapolation. A statistically significant coefficient is not automatically a causal anti-selection estimate.

Life Premium Risk Loadings

Expected present value of death benefits

For an annual premium term contract with benefit B payable at the end of the policy year of death, let v=(1+i)−1v=(1+i)^{-1}, where i is the assumed annual effective discount rate. Under a specified survival and persistency model, let ntn_t be the expected number in force at the start of year t per policy initially issued, qtq_t the conditional mortality for those lives, and CtC_t the death benefit payable per covered death. The expected present value of death claims over T years is

EPV(C)=∑t=1TvtntqtCt.EPV(C)=\sum_{t=1}^{T}v^t n_t q_t C_t.

If level annual premiums P are paid at the beginning of each covered year while the policy remains in force, the present value of premium receipts is

EPV(P)=P a,a=∑t=1Tvt−1nt.EPV(P)=P\,a,\qquad a=\sum_{t=1}^{T}v^{t-1}n_t.

Here nt+1n_{t+1} must reflect survival and lapse under the same timing convention. A contract paying midyear claims, offering premium waivers, containing surrender options, or allowing premium holidays requires corresponding modifications. The identity is the basis of a pricing calculation and is not itself a complete profitability or solvency test.

Expenses, margins, and premium adequacy

Assume an initial expense E0E_0, renewal expense ete_t paid at the start of each year for an in-force policy, and a specified explicit margin MM for modeled uncertainty and target return. The simplified gross level annual premium is

P=EPV(C)+E0+∑t=1Tvt−1ntet+Ma.P=\frac{EPV(C)+E_0+\sum_{t=1}^{T}v^{t-1}n_t e_t+M}{a}.

An investment return assumption, product guarantee, distribution commission, premium tax, capital cost, reinsurance premium, surrender expense, and shareholder hurdle could all modify the expression. The term M must be identified: a 6% markup on expected claims is not the same as a 6% increase in mortality, a 6% profit margin on premiums, or a regulatory prudence margin. These alternatives can generate materially different financial results.

ASOP No. 54 addresses pricing assumptions and risk evaluation, including sensitivity and stochastic analysis where appropriate. The underlying purpose is to assess prospective financial results under the characteristics of the specific product. The model should distinguish expected pricing assumptions, explicit risk allowances, financial capital requirements, and valuation rules; stacking unnamed conservative adjustments risks counting the same uncertainty multiple times.

A ten-year synthetic portfolio pricing experiment

To extend the earlier two-risk example, assume ten years of term coverage with a level gross annual premium. Take a $500,000 death benefit paid at each year-end; a 4% discount rate; $200 initial expenses per issued policy; $30 renewal expenses at each beginning-of-year in-force exposure; a 6% loading on the present value of expected death claims; starting risk-type death probabilities qL,1=0.001q_{L,1}=0.001, qH,1=0.004q_{H,1}=0.004; and mortality increasing by 5% each subsequent policy year within each type. Set lower-risk lapse probability to 4% and higher-risk lapse probability to 2%, conditional on survival. These percentages are synthetic assumptions chosen to expose mechanisms, not estimates from the cited SOA studies.

The baseline comparison begins with 15% higher-risk lives among issued policies. The selective-purchase scenario begins with the derived 20.93023256% high-risk share. Both use identical benefit, mortality curves, discount rate, expense basis, and type-specific lapses. The high-risk share then evolves endogenously from the state recursion.

Quantity Baseline issued mix Selective issued mix
High-risk share at issue 15.000% 20.930%
Present value of premium-paying exposure a 7.204395 7.233778
Present value of death benefits per policy $6,243.78 $7,062.24
Present value of renewal expenses $216.13 $217.01
Level gross annual premium with stated claim loading $976.42 $1,092.51
High-risk share among in-force policies at start of year 10 17.048% 23.564%

Under the modeled circumstances, the selective case requires $116.09 more annual premium, approximately 11.89% above the baseline premium. This premium difference is smaller than the relative change in first-year mortality because ten-year survival, lapses, discounted premium exposure, fixed expenses, and the margin rule all enter the ratio. The scenario demonstrates why a first-year death-rate uplift cannot be used as a direct substitute for a multi-year premium calculation.

The result also illustrates an accounting subtlety. The premium-paying annuity factor is marginally larger in the selective case despite higher deaths because the assumed lower lapse probability among the greater high-risk share increases later in-force exposures. Whether that effect occurs in actual business is an empirical question. The two-type model simply makes the competing mechanisms visible.

Price response and circular selection

The illustration treats purchase probabilities as exogenous. In an actual market, raising the premium can change who buys. Let aj(P)a_j(P) be the purchase probability for type j at premium P. Then the proportion of high-risk business and expected claims are functions of price:

wH(P)=waH(P)waH(P)+(1−w)aL(P).w_H(P)=\frac{w a_H(P)}{w a_H(P)+(1-w)a_L(P)}.

A candidate premium P determines participation and therefore the cost basis that a new premium must fund. A consistent pricing analysis would consider the resulting fixed-point condition

P=P(wH(P),q,ℓ,E,i,M),P=\mathcal{P}(w_H(P),q,\ell,E,i,M),

where P\mathcal{P} denotes the present-value pricing calculation. A solution may be unique, multiple, or absent depending on the demand responses and contract constraints. This is a statement of the ordinary interaction between demand and actuarial costs, not the introduction of a new economic theory. Estimating aj(P)a_j(P) is particularly demanding when private type j is unobserved.

Insurers also face a competitive response. A price change can alter distribution-channel mix, quoted-to-bound conversion, underwriting referral patterns, and face-amount selection, even if aggregate demand is stable. A monitoring system that treats these responses as fixed can systematically understate adverse future composition.

Margins and capital should be assigned to their risks

Mortality assumption uncertainty, selection uncertainty, stochastic mortality volatility, catastrophe accumulation, expense overruns, lapse risk, model risk, and discount-rate uncertainty are distinct sources of financial risk. They may interact, but they should be identified before combining them. A mortality parameter margin is a provision against its own estimation or future deterioration; a risk-capital charge is a cost of carrying financial resources against adverse outcomes; a profit target specifies a commercial objective.

Under the Solvency II life underwriting module, mortality, longevity, lapse, expenses, and catastrophe are separately recognized risk components for capital purposes. That regulatory structure underscores the need to keep pricing profit metrics and solvency capital charges conceptually distinct. Its exact capital formulas apply within the specified European regulatory framework, not to every global carrier.

Reinsurance can alter the retained exposure distribution through cession limits, facultative underwriting, and treaty definitions. A reinsurance premium may transfer claims but leave the ceding insurer with expense, persistency, counterparty, or model risk. Any claim that a risk loading has been eliminated by reinsurance must identify the risk genuinely transferred.

The contrast between death-benefit and annuity pricing

For death insurance, higher mortality generally increases expected claims for a fixed death benefit. For a life-contingent annuity, higher mortality generally reduces expected survival-contingent payout periods. Product options can complicate both patterns. The empirical annuity literature, including Finkelstein and Poterba (2004), is useful for understanding contract-level sorting, but a sign reversal in liability sensitivity must be respected when translating the method. The presence of the two products in a company's portfolio does not justify netting their mortality risks without an explicit dependence and capital analysis.

Experience Monitoring

A consistent observed-to-expected system

The basic monitoring ratio is observed claims divided by expected claims on the same exposure basis. For death counts,

A/Ecount=∑dit∑eitqitexpected,A/E_{\mathrm{count}}=\frac{\sum d_{it}}{\sum e_{it}q_{it}^{\mathrm{expected}}},

with an exposure convention compatible with the probabilities. For death-claim amounts,

A/Eamount=∑ditCit∑eitqitexpectedCit.A/E_{\mathrm{amount}}=\frac{\sum d_{it}C_{it}}{\sum e_{it}q_{it}^{\mathrm{expected}}C_{it}}.

These ratios can diverge materially if high-face-amount lives experience different mortality. Both numerators must use actual covered claim eligibility and comparable valuation periods. A ratio above one establishes that realized experience exceeded the modeled expectation on the tested basis; it does not by itself identify private information as the cause.

Expected mortality should be frozen to the basis actually being backtested, including the version of underwriting classes, improvement scales, duration adjustments, and reporting conventions. Replacing the expected table midstream changes the denominator and may obscure deteriorating experience. Analysts can present both an original-basis backtest and a contemporaneous re-estimation, with a reconciliation.

Diagnostic segmentation

A credible monitoring system examines issue age, attained age, policy duration, sex where lawful and relevant, smoking class, underwriting class, medical or accelerated underwriting pathway, policy form, benefit amount, distribution channel, issue cohort, calendar period, and lapse or conversion status. The precise segmentation must reflect available data, legal constraints, sample size, and actuarial relevance.

The critical methodological distinction is whether a comparison conditions on an outcome partly determined by the risk being measured. For example, analyzing only people who successfully pass accelerated underwriting may understate the higher-risk applicants referred to full underwriting. An analysis of issued mortality therefore answers a narrower question than an analysis of the underwriting rule's effect on the total applicant pool. The NAIC 2026 VM-20 documentation treats evidence on underwriting-process similarity as relevant when setting valuation mortality, reinforcing the need to retain and document underwriting vintages.

Statistical uncertainty and repeated testing

For small observed death counts, A/E can vary widely by chance. Exact or approximate count intervals should accompany mortality ratios, with explicit acknowledgement of exposure uncertainty and model parameter uncertainty. Repeatedly searching dozens of subdivisions increases the chance of an apparently extraordinary result arising by chance. Large exposure cells reduce random error but can hide change within important subgroups.

A practical diagnostic sequence is to define the hypothesis before inspecting results, assess data completeness, compare like-for-like exposures, estimate uncertainty, inspect alternative explanations, and test whether the pattern persists in a later or held-out cohort. Model calibration should include the relationship between predicted and observed death rates across risk bands, not only ranking statistics. An underwriting score can rank risks successfully yet produce the wrong absolute mortality probabilities and therefore an inadequate premium.

ASOP No. 25's credibility and relevance guidance is particularly pertinent when a segment contains few deaths. The SOA experience archive also demonstrates the importance of updating studies as later experience becomes available and distinguishing studies by the insured product and calendar window.

Selection-specific monitoring indicators

To investigate selection, a portfolio can track quoted-to-bound ratios by price band, changes in face amount after an underwriting decision, refusal and referral frequencies, applicant-to-issued class transitions, lapse rates by risk segment, term conversion take-up, and mortality by contract option. These are behavioral indicators, not automatic measures of hidden mortality information. Their relationship to actual claims must be examined.

A particularly useful decomposition treats portfolio change as arising from (1) new-business composition, (2) underlying mortality development within stable cells, (3) survivor and lapse mix, and (4) benefit amount mix. The decomposition can be implemented with standardized cell weights or an explicit multistate model. Standardization alone cannot recover unobserved private information, but it can prevent observable composition changes from being mislabeled as evidence of anti-selection.

For underwriting changes, the monitoring design should preserve a cohort map identifying the old and new eligibility rules, risk-class boundaries, evidence thresholds, acceleration algorithms, referral routes, and any concurrent marketing change. Before-and-after comparisons without this map confound model effects with changes in the business process.

Governance, validation, and escalation

Mortality governance benefits from a written owner for each assumption, a source ledger, reproducible calculations, versioned table files, stable definitions of exposures and claims, and explicit thresholds for review. An emerging deterioration should trigger a structured investigation of data, classification, duration, population mix, benefit amount, and external trend before translating the residual into a higher price.

The ASOP No. 54 governance provisions support documenting modeling choices, sensitivity analysis, and the use of information provided by others. In cross-border business, governance must also record which actuarial standards and legal requirements apply to each product and legal entity. A single international mortality assumption can be analytically compared across countries but cannot erase differences in data collection, product definitions, permissible risk factors, and policyholder protections.

Long-Horizon Sensitivity

Why ten-year mortality pricing is path dependent

A level premium is collected over time, while most expected death claims occur under future conditions that differ from those at issue. The distribution of surviving and persisting policyholders evolves year by year. An assumption that is adequate for first-year mortality may be inadequate over a long guarantee period, especially when premium rates cannot respond to later portfolio deterioration.

The relevant long-horizon sensitivities include mortality level, improvement or deterioration, class relativity, lapse by type, product conversion, discount rates, expenses, initial selection, and the relationship between premiums and applicant purchase behavior. These factors affect both expected death claims and the denominator of the premium calculation. The sign and size of a partial effect are conditional on the rest of the model.

Reproducible stress comparison

Retain the ten-year synthetic example specified in the previous chapter. Apply an illustrative 20% multiplicative increase to the annual death probabilities of both risk types, keeping discount rate, lapse assumptions, expenses, and loading rule unchanged. The selective-issue model then has a death-benefit present value of approximately $8,452.77, a premium-paying exposure factor of 7.223092, and a required level annual gross premium of $1,298.15. Relative to the base selective-issue premium of $1,092.51, this is an increase of about 18.82%. The premium does not increase by precisely 20% because changing deaths also changes future premium-paying exposure and survival-weighted costs.

If the high-risk annual lapse probability instead rises from 2% to 4%, with all other selective-issue parameters restored to their base values, the model produces an expected death-benefit present value of approximately $6,743.94 and a gross annual premium of $1,063.45. In this particular model, equalizing lapse rates reduces later concentration of higher-risk lives. It would be inappropriate to assume that a real policyholder can be induced to lapse for actuarial convenience, or that a higher lapse rate always improves profit: a lapse can also remove future margins, generate surrender costs, or change reserve release.

Synthetic assumption set Level gross annual premium
Baseline issue mix, 15% high risk $976.42
Selective issue mix, 20.930% high risk $1,092.51
Selective issue mix plus 20% mortality stress $1,298.15
Selective issue mix, with 4% lapse for both types $1,063.45

All values were produced using the same recursion and present-value equation. They should be interpreted as transparent model output, not pricing guidance for any insurer or population.

Mortality improvement and cohort replacement

A mortality-improvement assumption is often thought of as a uniform offset to aging, yet the insured population can change independently of general mortality. New sales may enter with different risk distributions, underwriting technology may change the selected class, and persisting business may become progressively less comparable with current new business. Improving general population mortality can coexist with deteriorating experienced mortality in a particular insurer if its insured composition changes sufficiently.

The 2026 SOA insured-life improvement scale is specifically informative in showing that valuation improvement assumptions require dedicated research rather than automatic reuse of an older scale. A pricing sensitivity should include alternate improvement trajectories and at least one scenario in which a favorable population trend is partly offset by selection or claims-pattern change.

A useful long-horizon attribution separates the change in expected claim cost into within-cell mortality and changing portfolio weights. Algebraically, for classes j with class weights w_{j,t} and death probabilities q_{j,t},

Qt−Q0=∑j(wj,t−wj,0)qj,0+∑jwj,t(qj,t−qj,0).Q_t-Q_0 =\sum_j (w_{j,t}-w_{j,0})q_{j,0} +\sum_j w_{j,t}(q_{j,t}-q_{j,0}).

The identity uses a particular decomposition order; another order allocates the interaction differently. Its benefit is transparent accounting: the first sum is a composition contribution calculated at baseline risk rates, while the second measures within-class change weighted at current composition. It does not identify the causal reasons for either term. The choice of class granularity and stable classification definitions remains decisive.

Financial sensitivity versus regulatory prudence

A commercial premium, a statutory valuation reserve, and a solvency capital requirement all use information about future mortality, but they answer different financial questions. The NAIC principle-based reserve framework and the life-product mortality rules of VM-20 in the 2026 Manual concern specific US valuation obligations. The European Solvency II capital rule concerns the solvency capital treatment of life underwriting risks. Neither replaces a carrier's documented prospective pricing analysis.

Sensitivity analysis should therefore label the objective: expected profitability, probability of an adverse pricing result, reserve adequacy under an applicable valuation framework, or capital response. Mixing them can lead to a misleading statement that a contract is "adequate" when only one measure has been tested.

Integrated analytical synthesis

The research supports an ordered set of substantive findings.

First, the mortality rate used for premium formation is a property of an issued, selected, time-varying contract portfolio. Conditional probability provides the appropriate starting point, and the actuarial exposure definition determines what can be inferred from deaths.

Second, selection has multiple stages. Private information can influence demand, but the insurer's acceptance and rating rules can remove or redirect high-risk applicants. After issue, lapse and conversion decisions alter the mix of remaining liabilities. An aggregate comparison of ownership and mortality cannot by itself resolve the complete sequence.

Third, expected premium adequacy depends on the joint evolution of mortality, persistency, benefits, expenses, and the discount basis. Loading an annual death rate by an inferred anti-selection percentage is not mathematically equivalent to pricing a multi-year policy under changed participation and survival dynamics.

Fourth, actuarial credibility and legal admissibility place substantive limits on using predictors or external studies. A statistically plausible relationship should be tested for comparability, class stability, lawful use, and financial materiality. A model can be accurate on a historical issued sample while poorly describing future applicants whose purchase and underwriting processes have changed.

Fifth, empirical disagreement across cited life-insurance studies is a methodological result. Evidence of limited aggregate adverse selection in a broad purchaser sample can coexist with material private information among rejected applicants or adverse sorting within a specific conversion option. The population and conditional decision under study must be named before a finding is generalized.

Sixth, long-horizon pricing requires a documented account of how new business enters, how the existing portfolio runs off, how mortality assumptions change, and how uncertainties are allocated among pricing margin, reserve, and capital analyses. The synthetic experiment provides an exact example of this distinction; its figures have no empirical authority outside the declared assumptions.

These findings constitute an independent analytical integration of established mortality mathematics, selection economics, empirical studies, and actuarial standards. They establish an auditable methodological interpretation of the subject. They are not presented as a newly founded doctrine, named model, or unverified claim of priority.

Reproduction note

The numerical experiment can be reproduced using the following explicit deterministic procedure. Initialize lower-risk and higher-risk policy masses at the stated issued proportions. For policy year t from 1 to 10, set each annual mortality probability to its initial value times 1.05t−11.05^{t-1}. Add the discounted start-of-year in-force mass to the premium annuity factor and $30 times that discounted mass to renewal expense. Add the end-of-year expected death amount, $500,000 times discounted type-specific deaths, to the claim present value. Multiply each type's current mass by (1−qj,t)(1−ℓj)(1-q_{j,t})(1-\ell_j) to advance to the next policy year. Finally divide the present value of claims multiplied by 1.06, plus $200 and the renewal expenses, by the premium annuity factor. Use full precision during calculation; round only reported final values to cents or the stated number of decimals.

This recursion treats lapses after survival within the policy year, pays death claims at year-end, pays annual premiums at the beginning of the year, applies a flat 4% effective discount rate, assumes fixed death benefits, and ignores premium taxes, surrender values, reinsurance, catastrophe correlations, and investment or solvency constraints. Any change in timing conventions requires recomputation; it is not merely a cosmetic adjustment.

Conclusion

Prospective mortality pricing is an exercise in forecasting the insured portfolio generated by a contract, an underwriting process, and a market. Anti-selection enters the calculation through conditional participation and through later policyholder decisions. Its magnitude and even its sign depend on the information available to each party, the population considered, the price and product offered, and the stage at which outcomes are measured.

The mathematical result is straightforward: when purchasing or retention probabilities differ across mortality types, insured risk weights change; when insured risk weights change, expected claims and the appropriate level premium generally change. The empirical result is more specific. The established literature contains evidence of varied selection patterns across insurance markets and across the stages of life-insurance access and ownership. A scientifically defensible mortality assumption should therefore be derived from comparable, credible data and tested for the particular underwriting and contract conditions to which it is applied.

The dossier's original contribution is the integrated reasoning that connects risk-conditioned entry, underwriting, duration-specific mortality, benefit amount, persistency, premium present values, credibility, stress analysis, and regulatory context in a single auditable argument. The numerical experiments substantiate conditional calculations and quantify their sensitivity under synthetic assumptions. They offer a reproducible demonstration rather than an estimate of market-wide anti-selection. Full actuarial implementation requires product-specific data, appropriate professional responsibility, and the relevant legal and valuation standards of each jurisdiction.

References

  • Actuarial Standards Board. 2005; last revised posting May 2011. “Risk Classification (for All Practice Areas).” Actuarial Standard of Practice No. 12. Effective May 1, 2006. Official ASOP No. 12.

  • Actuarial Standards Board. 2013. “Credibility Procedures.” Actuarial Standard of Practice No. 25. Effective May 1, 2014. Official ASOP No. 25.

  • Actuarial Standards Board. 2018. “Pricing of Life Insurance and Annuity Products.” Actuarial Standard of Practice No. 54. Effective December 1, 2018. Official ASOP No. 54.

  • Cawley, John, and Tomas Philipson. 1999. “An Empirical Examination of Information Barriers to Trade in Insurance.” American Economic Review 89 (4): 827–846. Publisher DOI.

  • Chiappori, Pierre-André, and Bernard Salanié. 2000. “Testing for Asymmetric Information in Insurance Markets.” Journal of Political Economy 108 (1): 56–78. Publisher DOI.

  • Einav, Liran, and Amy Finkelstein. 2011. “Selection in Insurance Markets: Theory and Empirics in Pictures.” Journal of Economic Perspectives 25 (1): 115–138. Publisher DOI.

  • European Insurance and Occupational Pensions Authority. Current online consolidated Solvency II Rulebook, “Calculation of the Basic Solvency Capital Requirement,” Directive 2009/138/EC, Article 105. Official rulebook article.

  • European Parliament and Council. 2016. Regulation (EU) 2016/679 of 27 April 2016 on protection of natural persons with regard to personal data processing (General Data Protection Regulation). Official Journal L119, 1–88. EUR-Lex official text.

  • Fang, Hanming, Michael P. Keane, and Dan Silverman. 2008. “Sources of Advantageous Selection: Evidence from the Medigap Insurance Market.” Journal of Political Economy 116 (2): 303–350. Publisher DOI.

  • Finkelstein, Amy, and James Poterba. 2004. “Adverse Selection in Insurance Markets: Policyholder Evidence from the U.K. Annuity Market.” Journal of Political Economy 112 (1): 183–208. Publisher DOI.

  • Harris, Timothy, and Aaron Yelowitz. 2014. “Is There Adverse Selection in the Life Insurance Market? Evidence from a Representative Sample of Purchasers.” Economics Letters 124 (3): 520–522. Publisher DOI.

  • Hendren, Nathaniel. 2013. “Private Information and Insurance Rejections.” Econometrica 81 (5): 1713–1762. Publisher DOI.

  • National Association of Insurance Commissioners. 2026. Valuation Manual, January 1, 2026 Edition, especially VM-20 §9.C and related PBR reporting requirements. Kansas City, MO: NAIC. Official Valuation Manual.

  • National Association of Insurance Commissioners. 2025, updated August 1. “Principle-Based Reserving.” Official regulatory overview. NAIC PBR overview.

  • Rothschild, Michael, and Joseph E. Stiglitz. 1976. “Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information.” Quarterly Journal of Economics 90 (4): 629–649. Publisher DOI.

  • Society of Actuaries. 2015, with 2018 report update. “2015 Valuation Basic Report and Tables.” Society of Actuaries Research Institute. SOA VBT materials.

  • Society of Actuaries, Individual Life Insurance Experience Committee. 2024. “2019 Individual Life Insurance Mortality Experience Report” (experience years 2012–2019). Society of Actuaries Research Institute. SOA experience report.

  • Society of Actuaries Research Institute. Current topic directory, accessed November 2026. “Individual Life — Mortality Experience Studies.” SOA study catalog.

  • Society of Actuaries Research Institute and LIMRA. 2025. “2009–2023 Term Conversion Experience Study: Covering Term Conversion Incidence, Mortality, and Lapse and Post-Conversion Mortality and Lapse.” Official study listing.

  • Society of Actuaries Research Institute, Mortality and Longevity Oversight Advisory Council, Mortality Improvement Life Working Group. 2026. “Individual Life Insurance Mortality Improvement Scale — for Use with AG38/VM20 — 2026.” Released September 22, 2026. SOA research release.

  • U.S. Department of Health and Human Services, Office for Human Research Protections. 2009. “Genetic Information Nondiscrimination Act (GINA): OHRP Guidance.” HHS official guidance.