RESEARCH DOSSIER
Microinsurance Claims Frequency Severity and Data Scarcity
Abstract. Microinsurance claims frequency and severity are investigated as distinct but linked quantities formed by effective coverage, covered risk, notification, adjudication, payment and loss development. The analysis combines selected primary supervisory and national regulatory documents, reported market aggregates, practitioner measurement guidance, scientific studies, established actuarial probability methods and transparent synthetic numerical examples. It demonstrates how exposure mismeasurement, selective reporting, contract transformations and immature cohorts affect frequency and severity estimation; explains the use and limits of credibility, aggregate loss models and external data; and relates expected cost to sustainability and realized household protection. A comparison of Philippine regulatory reporting, international guidance and Indian experimental evidence illustrates the boundaries of inference. The study contributes an original analytical synthesis without claiming a new theory or estimating a global claims distribution.
Keywords: microinsurance; claims frequency; loss severity; exposure; data scarcity; reporting bias; actuarial credibility; compound losses; incurred claims; reserving; inclusive insurance; financial sustainability; household protection
Introduction
Microinsurance claims are the monetary and operational consequences of contractual coverage in populations for whom the cost of an adverse event can be large relative to disposable resources. The monetary size of the insurer's payment is a poor measure of the social scale of the underlying loss. A modest funeral benefit may prevent an emergency loan; a hospital reimbursement may arrive too late to finance treatment; a crop index payment may be legally accurate and economically insufficient. An actuarial account of these outcomes requires systematic separation of the insured hazard, the contractual benefit, the notification process, the insurer's liability and the household's recovery.
The parent discipline, Inclusive Insurance and Microinsurance, establishes the relationship between accessible coverage, financial sustainability and effective protection. The present research is narrower: it studies the simultaneous determination and imperfect measurement of claim counts and amounts. Its central questions are which exposure and outcome definitions support frequency estimation, how policy structure and reporting influence observed severity, and what inferences remain justified when records are thin, selective or structurally incomplete.
The analysis combines a targeted documentary comparison, derivations from established probability models, and transparent synthetic examples. The documentary corpus covers international supervisory guidance, dated Philippine legislation and statistical returns, international practitioner measurement, U.S. actuarial guidance used comparatively, and a randomized Indian enrollment study. These are purposive cases chosen for distinct evidentiary functions rather than a probability sample of countries. The institutional survey overview and journal abstracts establish only the statements actually reported there. The study does not estimate a worldwide microinsurance frequency or severity distribution: neither the reviewed public aggregate tables nor the accessible study summaries provide harmonized individual exposure, loss and claim histories. Figures labeled synthetic are calculations under stated assumptions, not observed market data.
The analytical contribution consists in integrating the insurance observation process with actuarial estimation and institutional performance assessment. It applies established methods and develops implications from their interaction; it makes no claim to a new named theory, empirical discovery or priority. The review date is November 18, 2026. Dated texts are examined as evidence of their respective editions rather than a comprehensive statement of currently enforceable law.
Definitions, History and Conceptual Framework
The unit of insurance risk
A risk unit is a person, asset, transaction, household or other contractually defined exposure under a specified benefit and period. A claim is a request for performance under insurance or an insurer-recognized payment obligation, depending on the operational definition. These two records must be linked by coverage identifier, dates and benefit. A single death might activate a funeral benefit, credit-life payment and family assistance under distinct contracts. A single hospitalization can generate several provider invoices under one episode. Counting either invoice lines or beneficiaries as independent claims changes the observed frequency without changing the hazard. A claim register must therefore preserve event, claim, payment and coverage identifiers as separate entities.
A benefit may be fixed, indemnity-based, service-based or index-triggered. For fixed benefits the insurer's severity often depends principally on the insured sum and contract conditions. For indemnity products, severity depends on the realized eligible loss, limitations and adjustment. In cashless health coverage it is essential to distinguish provider reimbursement from cash received by the insured. An index contract can generate a valid payment without individualized damage verification; the household's uncompensated damage and the index payment are distinct variables.
Insurance for an underserved population is institutionally defined through product, distribution and target group. A small insured sum alone cannot identify that population. The IAIS (2012) inclusive markets application paper placed proportional supervision within its framework of insurance principles and standards. Its financial inclusion program later positioned inclusion within the 2025-2029 societal resilience theme and describes revision of the 2012 paper. This chronology connects access and prudential judgment without treating historical application guidance as national statute. Throughout this study, microinsurance refers operationally to regulated or identifiable insurance coverage designed for underserved groups, with local statutory definitions retained when interpreting jurisdictional records.
A sequence of observable states
Consider a cohort eligible to purchase a product. Enrollment is the transition into a contractual relationship, and effective coverage is the period during which the relevant hazard is protected. An adverse event may occur within or outside that period. Even during the effective period, an event may fall outside the benefit definition. A potentially compensable event must become known to the insurer or administrator, satisfy the applicable conditions, survive any review or appeal, and yield actual payment or service. In many datasets only the last two stages are visible.
For a specified insured event type, let E be earned exposure measured in risk-unit years, M the number of covered occurrences, R the number notified by a defined cutoff, A the number ultimately accepted, and P the number paid by that cutoff. A, R and P must count the same object, and an accepted claim can still await payment. Automatic payments and non-claim-triggered benefits require a separately defined notification or recognition state. Let Y be total ultimate insured benefit and H the household's economic loss, including the dimensions actually measured. E, M, R, A, P, Y and H answer different questions.
Occurrence frequency is M/E. Reported frequency is R/E; accepted frequency is A/E, and paid frequency is P/E. For positive ultimate settlements, the arithmetic mean of benefits is Y divided by the corresponding ultimate positive claim count. In a mature cohort the product of positive-claim frequency and average positive severity is exactly the aggregate benefit per exposure, provided numerator and denominator are aligned. For open cohorts it is an estimate only after appropriate development adjustment. This accounting identity is independent of a particular stochastic model.
Access statistics involve a separate denominator. Enrolled lives divided by eligible population measures a form of reach only after definitions, duplicates and effective periods are reconciled. Payment incidence among those who experienced covered events requires an event survey or comparable independent source. A household that could not file a claim will be absent from a paid-claim ledger. Thus an increase in recorded claims can indicate improved service, rising event incidence, changing selection, altered benefits or a mixture of these causes.
Time and causality
The accident or occurrence date locates the hazard; the report date locates the insurer's knowledge; the decision date locates adjudication; the transfer date locates operational payment; the receipt date locates benefit delivery. Calendar-year counts can mix accident years. A newly launched program accumulates exposure before all occurrences mature, while an older program may pay claims from earlier periods. Comparison without a defined valuation date creates an artificial trend.
The practical distinction is between insurance frequency as a contractual actuarial quantity and observed reporting frequency as the product of hazard and institutional behavior. An institution can alter the latter merely by changing a form, agent, language or payment channel. Research must hold these mechanisms apart before interpreting experience as evidence of changing underlying risk.
Primary Institutional Evidence and Comparative Framework
Legal boundaries and product classification
The Philippines provides a useful example of legal specificity. In the statutory text of Republic Act No. 10607 (2013), section 187 links the daily charge ceiling to 7.5 percent of the applicable Metro Manila nonagricultural daily minimum wage and the guaranteed benefit ceiling to 1,000 times that wage. Section 188 assigns the Commissioner a regulatory role. The thresholds establish a legally bounded product category at the time and under the text examined. No peso-denominated contemporary threshold is inferred without checking the applicable wage and subsequent legal changes.
The Insurance Commission's January 2025 Guidelines on Inclusive Micro Products identifies multiple categories and distinguishes micro health insurance from micro HMO services. This matters for the observation unit: the number of preventive service encounters under a service product need not equal the number of indemnity claims. A pooled statistical series spanning different products may show frequency changes when the product mixture changes. The circular's dated language is evidence of classification, not a finding about compliance by particular providers.
In the cross-jurisdictional comparison, the IAIS application paper supplies general supervisory context while the Philippine texts establish a specific historical and regulatory example. U.S. professional standards examined later provide methodological comparison. Their authority and intended audience differ; the interpretation does not treat them as one worldwide law.
Exposure evidence and numerical reconstruction
The Philippine Insurance Commission's first-quarter 2024 statistical report gives 53,136,762 reported insured lives and approximately PHP 3.316 billion in premium production or contributions across mutual benefit associations, life insurers and non-life insurers (table 1). This is an administrative market snapshot, not an exposure file. Division of production by reported lives yields approximately PHP 62.41 per reported life, but the quotient neither establishes an annual premium nor a risk-unit-year rate. The records do not show individual enrollment intervals or whether the same person is insured under more than one contract.
The Commission's later preliminary 2020-2024 key data reports for its non-life microinsurance table 2024 earned premiums of PHP 3,175.2 million and incurred claims of PHP 1,172.9 million. Their ratio, 1,172.9 / 3,175.2, is approximately 36.94 percent, matching the table. The source explicitly cautions that reported lives can include short-period policies and multiple policies and notes separate MBA data. This narrower table cannot be used to calculate a loss ratio for the combined sectors in the quarterly report. It also contains no matched claim-count denominator for the displayed incurred amount. Count and size remain unidentified even though their financial total is disclosed.
The algebraic consequence is immediate. Any two portfolios producing the same incurred total and earned premium can have distinct count and size distributions. A ratio of 36.94 percent could arise from numerous small medical reimbursements or fewer larger property claims. The statistic establishes the relation between losses and premiums within that reporting convention, not the frequency of household damage or the adequacy of compensation.
International reach and selection of evidence
The Microinsurance Network's 2024 landscape overview reports 344 million people covered by reported products across 37 countries and USD 6.2 billion of written premiums for 2023. These reported-product aggregates describe the survey frame and published reference year. They cannot be silently converted into a globally deduplicated count or an actuarial exposure base, and written premiums cannot be substituted for earned premiums in an occurrence-period loss ratio. The full survey microdata and claim cohorts are outside the evidence accessed here.
Wipf and Garand's (2010) performance indicators handbook illustrates the complementary role of claims settlement and rejection metrics. Its examples sometimes use interpolation to fill gaps. Its measurement definitions inform this monograph, but its illustrative examples are not imported as raw observations. The GIIN (2022) claims rejection metric uses a reporting-period formulation; analysts must make explicit whether rejected decisions are counted from the same notification cohort or simply during the accounting period. These denominators can diverge significantly when backlog grows or is cleared.
Microinsurance Network's low claims ratio research request lists potential explanations related to sales incentives and disclosure. It documents an institutional research question, not evidence of the prevalence or causal weight of those mechanisms. This distinction is particularly important in interpreting international comparisons where participation and accounting practices vary.
The evidence hierarchy employed here is functional: legal text establishes the product rule under its edition; regulator returns establish defined reported aggregates; a practitioner handbook establishes a measurement convention; an experimental article establishes a reported treatment contrast within its sample; a mathematical derivation establishes a conditional result. None substitutes for the absent joint event-exposure-payment data. This hierarchy makes the conclusions reproducible without giving documentary sources more inferential power than they possess.
Claims Incidence
Exposure construction and estimation
Frequency estimation starts with effective coverage intervals. For risk i with eligible insured days d_i within a valuation period, a natural annual exposure is e_i = d_i / 365.25 when daily exposure is appropriate. Leap-year conventions may be defined differently but must be consistent. An annual policy censored after three months contributes approximately one-quarter year. A waiting period requires benefit-specific treatment. A credit-life exposure may terminate when debt is discharged, while a renewal interruption may create an uninsured interval. Aggregation over all coverages is meaningful only within a commensurate risk unit.
Let N_i be the count of specified claims for exposure e_i. The independent homogeneous Poisson model assumes N_i has mean and variance e_i times lambda. Given total N and E, the maximum-likelihood estimate of lambda is N/E. For a zero count the maximum-likelihood estimate is zero, while the one-sided 95 percent Poisson upper confidence bound is -log(0.05)/E, approximately 2.996/E. For 200 risk-unit years with zero observed claims, the upper bound is about 1.50 per 100 risk-unit years. A zero observation in a small portfolio therefore coexists with economically relevant risk. This confidence statement applies to the modeled observed event process, not to hidden unreported losses.
The Poisson assumptions fail where repeated episodes, geographic clustering or heterogeneous propensities produce overdispersion. A negative binomial count model with mean m and variance m + alpha times m squared can represent certain mixtures of rates. It is not automatically adequate for epidemic or catastrophe dependence. A two-part model can represent genuine structural zeros if a class has no eligible hazard during an exposure interval. Administrative non-observation, however, is not proof of structural zero. Where each exposure can give rise to at most one event, a Bernoulli or survival formulation may fit the contract more closely than the recurrent-event model.
A generalized linear count model can express log expected count as log exposure plus an intercept and a vector of covariates. The log exposure term is an offset; its coefficient is fixed at one. Candidate covariates can include benefit type, age band where legally and ethically permissible, region, season, tenure and channel. A model estimated on claims alone without a complete exposure roster has no reliable nonclaim comparison population. The analyst must keep the size of the sample and its effective independent information separate.
Credibility and sparse experience
Biener (2013), in Pricing in Microinsurance Markets, examines the particular pricing constraints of microinsurance. His 2011 conference pricing analysis discusses the tradeoff between insufficient premiums and uncertainty-driven unaffordability, as well as drawing on additional experience and simulation. The practical implication is that a local short history should be integrated with explicitly comparable evidence, not replaced by an unexplained market average.
The Actuarial Standards Board's (2013) Credibility Procedures distinguishes experience under study from relevant additional experience and emphasizes comparison of characteristics and homogeneity. It is a U.S. professional standard; the mathematical reasoning below illustrates broader actuarial practice without representing a universal legal requirement.
A standard Poisson-gamma example makes the updating process explicit. Assume the annual observed-claim rate lambda follows a gamma distribution with shape a and rate b. If the portfolio produces N claims over E unit years under a Poisson likelihood, the posterior has shape a+N and rate b+E. Its mean is (a+N)/(b+E); equivalently, the sample rate N/E receives weight E/(b+E), and the prior mean a/b receives weight b/(b+E). For N=30, E=1,000, a=20 and b=400, the prior mean is 0.05 and sample mean 0.03. The posterior mean is 50/1,400 = 0.035714. The hypothetical prior contributes the equivalent of 400 exposure-years at its assumed average. A prior is a modeling assumption, not an independent observation; the calculation cannot make it relevant if the contract or reporting process is different.
For a new portfolio of E_new exposure-years, the posterior predictive count has mean E_new times the posterior mean and variance equal to that mean plus E_new squared times the posterior variance of lambda. For 1,000 future exposure-years, this example gives mean 35.714 and variance approximately 61.224 (posterior variance 50/1,400 squared). A pure Poisson forecast using the posterior mean as a fixed rate would give variance only 35.714, omitting uncertainty in the rate. Both approaches remain conditional on unchanged coverage and reporting.
A prior sensitivity exercise should repeat the calculation under plausible means and exposure-equivalents. Choosing an external portfolio requires reconstructing policy limits, enrollment selection, benefit definitions, time periods and reporting completeness. A heavier prior can yield apparently stable but wrong predictions. Experience from the same source must not enter twice as both prior and likelihood. Cross-validation must exclude the target portfolio from any external credibility training set if the evaluation is intended to test transfer.
Reporting and acquisition effects
Under independent homogeneous thinning, if true covered events occur at rate lambda and each is notified with probability q, R conditional on E is Poisson with mean E lambda q. The observed likelihood depends on the product lambda q. With only E and R, both parameters cannot be identified separately. This limitation is structural rather than a consequence of insufficient computational power. If more exposure is recorded under the same observation mechanism, the product becomes increasingly precise while the individual parameters remain unidentified.
For a synthetic example, 30 notifications over 1,000 exposure-years establish an observed rate of 0.03. An underlying rate of 0.05 with q=0.60 and a rate of 0.0375 with q=0.80 are observationally equivalent for this model. If independent evidence supports q between 0.50 and 0.90, the corresponding underlying rate lies between 0.03333 and 0.06, ignoring sampling error. A proper uncertainty statement also propagates count variation and uncertainty in the external range. The range is a sensitivity assumption, not an empirical estimate of reporting propensity.
Selection into coverage changes the hazard population. Banerjee, Duflo and Hornbeck (2014), in their Karnataka bundling experiment, report that a compulsory insurance bundle increased microfinance dropout by 16 percentage points. The publisher abstract attributes the result to poor demand and reports absence of adverse selection in enrollment in that experiment. It does not quantify claim propensity or establish that adverse selection is absent in other markets. The lesson for incidence estimates is that the observed in-force population is a potentially endogenous subset of the intended market.
An insured person's probability of claiming may also increase after learning about benefits or gaining access to an agent. Effective claims assistance can raise recorded frequency while leaving occurrence risk unchanged. Conversely, difficult documentation may suppress reporting. Pricing on the reduced count without recognizing possible restoration of access creates a feedback problem: a future improvement in claims service becomes an unexpected increase in liabilities. Estimation should anticipate the promised service standard rather than merely reproduce the historical level of administrative friction.
Loss Severity
Contract transformations
Severity must be defined at the chosen stage. Let X be nonnegative eligible ground-up economic loss and B the insured benefit. For a deductible d and a maximum payment L with no coinsurance, a simplified indemnity contract gives B = min(max(X-d, 0), L). For coinsurance fraction c applied before a payment cap the corresponding formula is B = min(c times max(X-d, 0), L). Actual wordings can apply an aggregate rather than per-claim cap or order coinsurance differently. A fixed payment b following a covered event has B=b unless contract conditions change that amount. An index trigger defines payment using the recorded index, not necessarily X.
The distribution of B is therefore determined jointly by ground-up losses and the benefit function. A mass of observations at the policy limit is a genuine feature of payments and a censoring problem if estimating losses above that limit. For an insurer pricing the same limit, that mass need not be expanded to estimate paid severity; for analysis of a proposed higher limit or household need, tail estimation requires independent loss observations or explicit scenarios. Payment zero can result from a loss below deductible, rejection, absent reporting, lack of coverage or unresolved status. These mechanisms have different statistical meanings.
For positive payments, gamma and lognormal distributions can describe right-skewed amounts, while an empirical distribution can preserve recurring benefit schedules. A mixture of fixed sums and hospital reimbursements is generally better represented by distinct components. Heavy-tail models require sufficient tail observations or defensible external assumptions. A maximum benefit creates bounded single-claim insurer severity but does not remove aggregate dependence or concentrated events. In a small sample the apparent maximum may instead reflect the largest observed loss, not a contractual maximum. Those cases require different stress tests.
Reporting selection and payment measurement
Suppose an eligible loss of size X is reported with probability q(X). Among reported events the expected ground-up severity is E[X q(X)] / E[q(X)], provided the denominator is positive. The difference from unconditional mean is Cov(X,q(X))/E[q(X)]. If larger losses are more likely to be reported, and other things are equal, reported severity may exceed the underlying mean. A fixed administrative burden relative to the benefit makes this mechanism plausible. But the sign can reverse where severe events make documentation, travel or communication unusually difficult. The formula provides a conditional explanation rather than evidence about which mechanism dominates a given country.
Consider two equally common covered loss classes, 50 and 200 monetary units, with reporting probabilities 0.40 and 0.80. The population mean is 125, the mean reporting probability is 0.60, and the mean reported severity is (0.5 times 50 times 0.40 + 0.5 times 200 times 0.80)/0.60 = 150. Expected recorded benefit per underlying event is 90 under full indemnification of reported losses. If reporting rises to one in both classes, expected recorded benefit becomes 125 while average reported severity falls from 150 to 125. This example demonstrates why changes in frequency and severity can offset one another. A falling observed severity can accompany improved effective protection and rising aggregate cost.
Payment severity measures money delivered under the contract, whereas the beneficiary's recovery depends on whether it arrives when required, whether an intermediary retains part, and whether an insured service is accessible. Medical direct payments should be recorded as insurer cost and separately as benefits in kind. Credit insurance payments to a creditor should be distinguished from liquidity delivered to the household. A high total of settled insurer liabilities does not prove that claimants received the necessary resources at the right moment.
Aggregate losses and concentration
In the standard independent compound model, S is the sum of N independent identically distributed positive severities X_j, independent of N. Conditioning on N gives E[S] = E[N] E[X] and Var(S) = E[N] Var(X) + Var(N) (E[X]) squared. With Poisson N of mean m the variance reduces to m E[X squared]. These are distributional identities under specified independence assumptions. If N and severity respond to common shocks, the model must condition on them or otherwise model dependence.
For 1,000 exposure-years at a hypothetical rate of 0.04, expected count is 40. If every claim pays 100, the mean aggregate is 4,000 and variance 400,000, with standard deviation approximately 632.46. If the conditional payment mean remains 100 but the payment variance is 10,000, the aggregate variance doubles to 800,000 and standard deviation rises to approximately 894.43. Pure expected cost alone cannot establish capital needs. The count process may also be overdispersed, adding another source of aggregate volatility.
A community exposed to a flood or a single employer's closure may have simultaneous losses. Splitting sales among thousands of policies inside the same community provides limited diversification against that event. A catastrophe scenario should specify affected share, average contract benefit, claims recognition, settlement time, reinsurance recovery and liquidity. Maximum payout per person limits an individual exposure, not the simultaneous aggregate. Correlated claims can also overwhelm the processing channel precisely when fast service is most valuable.
Development and valuation
An insurer's ultimate liability includes paid benefits and appropriate estimates of outstanding obligations, including incurred but not reported claims when the obligation has arisen. A paid total is a cash-flow measure. For recent accident cohorts it usually understates ultimate cost if obligations are open, although later recoveries or corrections can reduce net amounts. A canceled payment must not be counted twice when reissued.
Development triangles can organize paid, reported and incurred amounts by occurrence cohort and valuation lag. In small portfolios many cells are zero or dominated by single payments. Chain-ladder factors can be unstable, and changes in service speed can invalidate historic patterns. An expected-loss-based approach may supplement them, with assumptions about exposure and ultimate loss levels made explicit. The Bornhuetter-Ferguson logic combines losses already observed with a prior expectation for the undeveloped share; reliability depends on the relevance of that expected loss and the development pattern. For a simple synthetic example, if cumulative reported losses at a valuation date are 2,400 and a justified model estimates 60 percent reporting completeness, a development-factor estimate is 4,000. With an independent prior ultimate of 3,500, a reported-loss-plus-expected-unreported calculation gives 2,400 + 0.40 times 3,500 = 3,800. These are alternative conditional estimates, not additive reserves. Outstanding amounts then depend on the precise basis and paid-to-date values.
Reporting lag should not be conflated with settlement delay. When observed delays include only paid cases, long open claims are systematically omitted. A time-to-payment analysis with right censoring preserves open claims, but rejection is a competing outcome rather than an ordinary pending payment. Changes in claims practice should be modeled by calendar period, origin cohort and product. A useful operational series reports receipt distributions, backlog age and denials together with the actuarial liability estimate.
Data Reporting Bias
Architecture of missing information
Sparse data take several forms. A thin but complete cohort has few independent claims; a truncated ledger begins after historical events; a selective ledger excludes events that never become claims; missing covariates block risk adjustment; and aggregate returns destroy linkage even when the totals are accurate. These conditions demand different responses. More sophisticated regression can improve efficiency in a thin complete cohort but cannot recover unrecorded losses without external assumptions or measurements.
The U.S. Actuarial Standards Board's (2016) Data Quality standard addresses data selection, review, use and disclosure of materially uncertain or potentially biased results. For this investigation the analogous scientific requirement is to identify what each file was collected to accomplish. A premium collection system is optimized for transactions; a claim ledger for liabilities; a household survey for reported experiences; a regulator return for supervisory aggregates. No single source necessarily has the correct observation unit for every research question.
Missing completely at random is a strong condition under which missingness bears no relationship to observed or unobserved variables relevant to the analysis. Missing at random is conditional on variables available to the model. Missing not at random allows dependence on unobserved outcomes. Microinsurance reporting after a loss can plausibly depend on severity, awareness, transport access or documentation. Without evidence on these factors a complete-case estimate should be interpreted as a property of records, not a population estimate.
Minimum linked data and audit
A minimally useful longitudinal dataset contains pseudonymous policy and risk identifiers, product and version, benefit identifiers, enrollment and effective dates, termination or suspension, exposure units, premium basis and coverage amount. The claims table adds occurrence, notification, decision and payment dates, event and claim IDs, claimed amount, approved amount, outstanding estimate, rejection reason, reopening and payment recipient. A payments table should preserve reversals and transaction references. A separate sample or administrative linkage can measure potentially covered occurrences that were never notified, with legal authorization and clear consent arrangements where required.
The required validation is relational. The study should check duplicate active coverages; ordered dates within the relevant contractual conventions; links from claims to valid policy versions; negative or reversed payments; reconciliation of liability and payment totals; missing states; and consistency in open and rejected cohorts. An exception does not automatically prove fraud. A reported claim for an event before enrollment can reflect an error, a retroactive coverage rule or a genuine dispute over effective dates. Retain original records and trace corrections.
Data should be minimized and restricted to legitimate purposes. Household vulnerability and health status can be sensitive, and pseudonymous records can still allow reidentification when combined with detailed geography or rare events. An analytic dataset does not need unrestricted access to claim documents. Role-based access, separation of identifiers, retention policies and controlled aggregation improve both protection and reproducibility. Local legal requirements need independent confirmation for operational deployment.
Detection, validation and partial identification
An external sample of covered and unreported occurrences can estimate reporting probability under a stated sampling design. One approach is to follow an independently selected cohort with periodic questions about specified events and then compare self-reports with claims records. The sample should include discontinuers where the estimand concerns eligible or enrolled households, while incidence calculations restrict effective exposure appropriately. Validation of self-reported events may require independent records. Telephone-only or digital-only surveys risk omitting households with weak connectivity; nonresponse weights require credible inclusion probabilities and observed covariates.
If reporting propensity q_i can be estimated from external evidence and is strictly positive for all modeled strata, inverse probability weighting uses a reported event weight 1/q_i to estimate covered incidence or loss. This estimator becomes unstable for tiny q_i; weight truncation trades variance for bias and must be disclosed. If a class never reports, recovery from the claim register is impossible by weighting alone. Capture-recapture can sometimes combine two lists, but the simple independent two-list estimator relies on comparable populations, reliable matching and capture independence. A provider and insurer probably share referral pathways, so independence cannot be assumed.
When external measurement is unavailable, interval analysis offers a transparent alternative to point estimation. Suppose 20 out of 100 eligible surveyed event occurrences were paid and 15 remain unresolved; if all remaining outcomes are definitively negative, ultimate paid incidence lies between 20 and 35 percent of that event cohort. The lower bound treats all pending as unpaid and the upper bound treats them as paid. If 100 was instead the number of reported claims, this is a claim-cohort settlement interval and reveals nothing directly about unreported events. A distinction between these denominators is mandatory even for elementary arithmetic.
A reporting audit can compare channels with matched product and exposure periods and then explore differences in submission, documentation and rejection. Channel choice is often endogenous. A digital channel attracting younger users can show different incidence for demographic rather than administrative reasons. Randomized assistance, phased rollout or a defensible quasi-experiment can estimate the effect of a service change on notification. To estimate prevented damage or welfare effects, the outcome and causal estimand must instead concern independent household measures. A difference in reporting alone cannot establish a difference in loss occurrence.
Reproducible sensitivity and performance
The following synthetic cohort has 1,000 risk-unit years, 60 underlying covered occurrences, 36 reports, 30 accepted claims and 24 paid by valuation. Six claims are finally rejected; the six accepted and unpaid claims remain outstanding. All reports have been adjudicated. Paid benefit equals 2,400, and each accepted claim has an ultimate benefit of 100. Assume for illustration that the six rejected claims have zero final liability and the 24 unreported occurrences each have a legally payable benefit of 100 if validly reported. The example assumes the latter obligations survive notice requirements; actual contracts and applicable law may differ.
| Measure | Numerator and denominator | Result |
|---|---|---|
| Covered occurrence frequency | 60 / 1,000 | 6.0% per unit year |
| Reported frequency | 36 / 1,000 | 3.6% per unit year |
| Accepted frequency | 30 / 1,000 | 3.0% per unit year |
| Paid frequency | 24 / 1,000 | 2.4% per unit year |
| Rejection among notifications | 6 / 36 | 16.67% |
| Paid share of accepted by valuation | 24 / 30 | 80% |
| Mean accepted ultimate severity | 3,000 / 30 | 100 units |
| Paid amount per exposure | 2,400 / 1,000 | 2.4 units |
| Scenario total ultimate obligations | 3,000 + 2,400 | 5,400 units |
The underlying covered occurrences are distinguished from ultimately valid claims: the scenario includes six occurrences that are rejected for a final contractual reason, for example exclusions specific to the submitted benefit or a final valid procedural bar. If the term covered occurrence were defined to include only fully payable events, such rejections would instead have to be excluded from the underlying count. Here the occurrence label denotes events falling under the nominal risk category before full adjudication. Under the assumed legal obligation, estimated outstanding is 5,400 minus 2,400 = 3,000, of which 600 is accepted but unpaid and 2,400 is unreported. This deterministic example is not a reserve recommendation. It exhibits the size of the potential gap between cash transactions and the modeled contractual obligation.
Without independent information, the insurer would observe 36 reports and might estimate 3,000 ultimate accepted benefits for known claims. A separate event survey could reveal the remaining 24 occurrences, but only a legal and contract assessment establishes whether these are payable. The result separates statistical discovery from liability recognition. It also shows that a claims ratio can rise with improved identification of contractual obligations.
Wipf and Garand (2010) use a service-time schedule to characterize claim processing and warn against discarding unresolved cases. The resulting practical measurement program should retain the entire notification cohort through its last valuation. A reporting-period metric such as the GIIN definition must be explicitly labeled as a flow statistic where its numerator and denominator are not cohort-matched. An insurer can publish both views, but should not present them as interchangeable.
Integrated Scholarly Synthesis and Further Research
Pricing and financial sustainability
A prospective pure premium is expected ultimate insured benefit per unit exposure. Under independent frequency and severity assumptions it equals the product of expected positive-claim frequency and expected positive severity. More generally it is the expected aggregate cost per unit exposure, potentially conditional on shared risk factors. Expenses, cost of capital, taxes, commission and reinsurance are distinct components. A technically adequate premium should be evaluated on the same gross or net basis as the benefit and expense projections.
Consider a hypothetical product with 0.04 expected positive claims per exposure-year and conditional severity 100, giving pure cost 4. Suppose fixed and variable servicing costs collectively contribute 2 monetary units per exposure-year and a risk/capital provision contributes 1. With no additional expense components the illustrative required premium is 7. If price is 8 and earned exposure 1,000, the expected claims ratio is 4,000/8,000 = 50 percent, and the specified expense and risk provision leave 1,000 before other omitted items. These are expected-value calculations, not a recommended price. A 50 percent claims ratio could coexist with strong service and significant fixed costs or with modest benefits; the ratio alone does not resolve value.
Suppose an analyst fitted the reported frequency of 0.03 at unchanged mean insured severity 100, producing estimated pure cost 3. Improved reporting raises it to 0.05, resulting in pure cost 5 if acceptance and severity are unchanged. At a fixed earned premium of 8 per year the expected claims ratio rises from 37.5 to 62.5 percent. The increase of 25 percentage points is consistent with unchanged true hazard and better reporting. Once the same change raises processing expenses, the sustainability effect depends on total liabilities, expenses, reinsurance and reserves. This is why a low historical ratio caused in part by claim friction is not a prudent baseline for a product committed to better service.
Conversely a high claims ratio can reflect underpricing, external shocks, temporarily adverse development, richer benefits, or an intentional social subsidy. An institution may legitimately receive a public or donor subsidy that changes the amount charged to customers but not the cost of the insurance promise. An actuarial study must distinguish the technical cost, the customer contribution and the funding source. It must also make sustainability scenarios under subsidy expiration and growth explicit.
Evaluation of competing explanations
For any observed decline in claims per earned premium, at least four explanations deserve consideration: lower underlying exposure risk or hazard, higher price relative to unchanged benefit, contraction or narrowing of coverage, and reduced probability or cost of reporting and obtaining settlement. A composition change can produce any of these patterns through altered product or demographic mix. Development lag and inconsistent accounting basis may create additional apparent change.
The appropriate test depends on the explanation. Independent event data inform hazard. Comparable contract versions establish benefit changes. Exposure and enrollment histories establish population and mix. Claim and complaint histories illuminate the reporting and settlement process. Financial statements reconcile earned premium and ultimate incurred amounts. A multivariate statistical model cannot replace these observations when its dependent variable merges the mechanisms. The comparative institutional sources examined here illustrate precisely this point: a legal definition can bound benefits, a regulatory table can establish reported aggregates, and a claim metric can define service, while none independently identifies all elements of the causal process.
Scale also has ambiguous implications. Increasing portfolio size can reduce process variance per independent exposure and spread fixed systems costs. Geographic concentration, multiple family members exposed to the same event and shared delivery bottlenecks reduce those gains. Pooling across programs can improve estimation if benefit and reporting mechanisms are comparable. It can also erase important differences in access. A defensible research design uses a common minimum data schema while permitting product-specific risk units and benefit definitions.
Comparative research program
The immediate empirical priority is a cohort dataset linking coverage and claims with an independent event sample. For health products, define care episodes and clinical eligibility under the contract; for funeral insurance, define death event and beneficiaries; for property insurance, identify assets, damage and occurrence; for index products, retain both trigger data and independent loss outcomes. Product-specific catastrophe grouping makes cross-product aggregate capital assessment possible without assuming all policies are independent.
Prespecify primary outcomes, valuation dates, exclusion rules, missingness analyses, and how overlapping policies will be counted. Distinguish occurrence frequency from reported count and define whether severity is claimed, accepted, paid or ultimate. Document cross-currency comparison: convert only at an explicitly chosen rate and price date. Ratios to household consumption or income may inform adequacy, but require independent household measures and an ethical data protocol. Inflation adjustments for medical and property losses should reflect relevant cost indices where available rather than a single general price index by default.
Validation should test out-of-time and out-of-location calibration, predicted-versus-realized frequency for exposure cohorts, severity quantiles and aggregate loss behavior. A narrow model with a known observation mechanism is often more informative than a high-dimensional prediction whose labels could encode access barriers. Predictive performance on recorded payment alone should not be used to allocate claims assistance in a way that suppresses legitimate reporting. Changes in the claims system should be versioned and models recalibrated explicitly.
Risk classification should be tested for both predictive relevance and distributional consequences under applicable law. A group with low recorded claims because of poor access may appear inexpensive. If the resulting classification removes outreach to that group, the policy converts a data artifact into a persistent exclusion. Conversely, community rating or simplified classification may increase cross-subsidy while minimizing costly screening. These are contractual and institutional choices involving affordability, sustainability, fairness and public policy, not consequences that can be decided by a statistical fit alone.
Scope of the contribution
The original synthesis is a linked interpretation of exposure measurement, selective reporting, settlement development and financial performance. It yields a specific conclusion: contractual claim cost and household protection must be evaluated using distinct but connected observation units, because improving the observation and service process can simultaneously increase financial liability and effective protection. This conclusion follows from standard thinning, selection and compound-loss calculations and the institutional distinctions examined in the source record. No priority, original empirical estimate or new theoretical system is asserted.
The literature and legal scope is bounded. The accessible national comparison focuses on the Philippines; the enrollment experiment concerns India; the IAIS and practitioner documents have international scope, and U.S. standards are methodological comparators. The Landscape overview is not a full survey evaluation. The analysis does not establish current compliance in any market or reproduce underlying study microdata. A more expansive international empirical comparison requires current product contracts, harmonized administrative cohort files, field survey data and permission to process them. These constraints delimit the interpretation rather than support an inference about an unobserved global distribution.
Conclusion
Microinsurance claims frequency and severity are estimable only after defining the precise contract, effective exposure, event and count unit, benefit and valuation stage. A payment ledger establishes recorded transactions; claims liabilities and household loss require further evidence. Sparse experience creates statistical uncertainty, while systematic nonreporting creates an identification problem that cannot be resolved by additional observations of the same selective process alone.
The mathematical analysis establishes conditional relations among occurrence, reporting and payments. In the simple thinning model, the reported rate identifies only the product of the underlying rate and reporting probability. Severity is altered by contract transformations and the selective probability that an event becomes a claim. Aggregate liability depends on frequency, severity, dependence and development. Credibility and sensitivity analysis provide useful approaches when external experience is genuinely comparable and uncertainty is explicitly represented.
The documentary comparison supports a disciplined interpretation of market performance. The Philippine regulator's statistics report premiums and coverage measures on defined bases, and the non-life loss ratio can be reconstructed from its preliminary incurred and earned amounts. The ratio contains neither component of a separate frequency-severity distribution. International indicator definitions distinguish important aspects of processing, but require well-defined cohorts. The Indian experiment illustrates why enrollment may change when an insurance product is bundled with another service; its result should not be generalized into a universal claim behavior relationship.
A sustainable microinsurance program must estimate the obligations it promises under an effective claims system, including the possible cost of better reporting and settlement, and must measure whether that promise translates into timely usable protection. A low claims ratio alone neither establishes inadequate protection nor proves efficient pricing. It is a starting point for investigation. The defensible research outcome is a coherent set of exposure, occurrence, reporting, acceptance, payment, benefit and household measures, with clearly stated assumptions and uncertainty. Reliable protection and financial viability can then be assessed together rather than inferred from an isolated statistic.
References
Actuarial Standards Board. 2013. Actuarial Standard of Practice No. 25: Credibility Procedures. December; effective May 1, 2014. United States. Source.
Actuarial Standards Board. 2016. Actuarial Standard of Practice No. 23: Data Quality. December. United States. Source.
Banerjee, Abhijit, Esther Duflo, and Richard Hornbeck. 2014. Bundling Health Insurance and Microfinance in India: There Cannot Be Adverse Selection If There Is No Demand. American Economic Review 104(5): 291-297. DOI: 10.1257/aer.104.5.291. Source.
Biener, Christian. 2011. Pricing in Microinsurance Markets. Conference paper, Madrid 2011, International Actuarial Association archive. Source.
Biener, Christian. 2013. Pricing in Microinsurance Markets. World Development 41: 132-144. DOI: 10.1016/j.worlddev.2012.05.028. Source.
Global Impact Investing Network (GIIN). 2022. Claims Rejection Ratio (PI7902). IRIS+ System, version 5.3. Source.
International Association of Insurance Supervisors (IAIS). 2012. Application Paper on Regulation and Supervision Supporting Inclusive Insurance Markets. Approved October 9. Basel: IAIS. Source.
International Association of Insurance Supervisors (IAIS). n.d. Financial Inclusion. Accessed November 18, 2026. Source.
Microinsurance Network. 2024. The Landscape of Microinsurance 2024. Report overview, 2023 reference year. Source.
Microinsurance Network. n.d. Request for Information on Reasons for Low Claims Ratios in Inclusive Insurance. Accessed November 18, 2026. Source.
Philippines Insurance Commission. 2024. Microinsurance Industry Quarterly Statistical Report as of the Quarter Ending on 31 March 2024. Manila, table 1. Source.
Philippines Insurance Commission. 2025. Circular Letter No. 2025-01: Guidelines on Inclusive Micro Products. January 16. Manila. Source.
Philippines Insurance Commission. 2026. Annex A: Key Data 2020-2024 (Preliminary Figures). Manila. Microinsurance non-life table and footnotes. Source.
Republic of the Philippines. 2013. Republic Act No. 10607, amending Presidential Decree No. 612 (Insurance Code), Title 6, sections 187-188. Lawphil statutory text. Source.
Wipf, John, and Denis Garand. 2010. Performance Indicators for Microinsurance: A Handbook for Microinsurance Practitioners. 2nd ed. Microinsurance Network. Source.