Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Health Claims Cost and Utilization Models

Abstract. Health claims cost is the product of health needs, decisions to seek and supply care, contractual coverage, the organization of services, prices, and the administrative recognition of a payment obligation. This monograph develops an integrated actuarial examination of how health exposures, treatment events, service utilization, unit costs, provider practices, morbidity composition, and financial claims can be measured and projected without confusing their distinct denominators or institutional meanings. It derives the relationship between member-month exposure, service rates, episodes, allowed expenditures, payer liability, and claims development; evaluates count, two-part, compound, and generalized linear modeling approaches; examines provider- and patient-level heterogeneity; and demonstrates exact price–utilization decomposition and risk-standardized forecasting with fully identified hypothetical examples. Its international comparison distinguishes national health expenditure accounts, health service statistics, social protection systems, and private insurer claim ledgers. The analysis incorporates U.S. and international statistical documentation, actuarial professional standards, and methodological research on skewed costs, count processes, and risk adjustment. An original analytic synthesis finds that model validity rests on consistency across three linked representations: clinical service production, insurance contract valuation, and accounting development through time. Their correspondence must be explicitly estimated, tested, and reconciled. The results include operational specifications for cohort construction, utilization metrics, price indices, case-mix adjustments, experience trend, prospective trend, calibration diagnostics, and sensitivity testing. All numerical cases are illustrative; no patient-level dataset has been newly estimated, and no novel actuarial theory or proprietary prediction model is asserted. The monograph provides a documented scientific basis for further insurer-specific empirical research, actuarial review, and international comparison.

Keywords: Health insurance; medical claims; medical utilization; medical cost; actuarial modeling; member months; claim frequency; treatment episodes; allowed expenditure; insurer liability; frequency–severity modeling; two-part models; zero inflation; generalized linear models; Gamma regression; provider variation; case mix; risk adjustment; medical trend; price–quantity decomposition; incurred claims; paid claims; claims runout; prediction calibration; data quality; clinical coding; catastrophic claims; international health accounts.

Introduction: Research Object, Questions, and Evidence

The scientific object of this dossier is the cost of insured health care as generated by exposure to covered risk, the occurrence and treatment of health conditions, the utilization of medical resources, the payment terms of coverage, and the statistical process by which expenditures become observable in claims. Each stage changes what a numerical claim means. A submitted hospital bill is a request for payment; an allowed amount is a price after adjudication and contractual adjustments; a member’s cost share is a contractual allocation; a paid claim is a cash event; an incurred claim concerns an underlying service or obligation whose full financial settlement may occur later. A model that does not say which of these it predicts is incompletely specified.

The System of Health Accounts 2011, issued by OECD, Eurostat and WHO (2017 revised edition), provides a comparative accounting vocabulary for health care consumption, financing schemes, functions, and providers. It supplies a rigorous reason to separate all-system health expenditures from a given insurer's benefits. The OECD health indicators (OECD, 2025) document the variation in access, financing, resources, and spending among countries. Neither dataset identifies a universal equivalence between national medical spending and the expense obligation of any individual insurance carrier.

In the United States, National Health Expenditure Accounts (Centers for Medicare & Medicaid Services, CMS, 2025–2026 access) report $5.3 trillion in 2024 spending, a 7.2% annual increase, and $15,474 per resident. Those economy-wide values include financial and administrative categories beyond an insurer's claim ledger. A 7.2% national expenditure increase cannot be inserted unchanged into an insurer's prospective medical claims trend without examining its different exposure, morbidity, product, reimbursement, and accounting boundaries.

The core research questions are these. How can exposure and claims units be defined so service rates are comparable over time? Which probability models appropriately represent counts, zero spending, positive severity, clustering, and high-cost tails? Under what conditions are changes in price separable from changes in care quantity, intensity, and diagnostic composition? How should an analyst distinguish morbidity from provider effects and contractual payment differences? How is observed claim experience transformed into a defensible prospective forecast? Which tests establish whether a model's predicted costs are sufficiently calibrated for its intended actuarial use?

The method comprises mathematical derivation, close examination of methodological and institutional primary materials, a selective review of relevant scholarly literature, international accounting comparison, and explicitly hypothetical numerical demonstrations. The literature review is targeted rather than a registered systematic search. Evidence is traced to original publishers, government statistical programs, and professional standards. Empirical claims attributed to studies are limited to those inspected in accessible full text or publisher abstracts and documentation. Mathematical examples are reproducible arithmetic, not estimates from confidential claims. This analytical work contributes a structured interpretation of established models; it does not claim a new theory, an original fitted algorithm, or causal identification from newly collected data.

Professional practice imposes a further discipline. ASOP No. 23 on data quality (Actuarial Standards Board, 2016 version) addresses data selection, review, limitations, and disclosure. ASOP No. 56 on modeling (Actuarial Standards Board, 2019) addresses model use, appropriateness, assumptions, and controls. These are U.S. actuarial standards with specified professional scopes; their techniques and evaluative questions are useful comparatively, while their formal applicability does not extend automatically to every jurisdiction. Accordingly the monograph separates general statistical reasoning from professional obligations in particular systems.

Health Exposure Units

Coverage, time at risk, and person denominators

The exposure unit fixes the quantity against which every frequency or cost is divided. For a population of eligible persons indexed by i, define d_i as eligible coverage days during the observation year. Person-years of exposure are E_y = sum_i d_i/D, where D is the chosen day convention, and member months are E_m = sum_i m_i when true enrolled calendar-month records exist. The conventional annualization E_y = E_m/12 is appropriate only after the monthly rule has been documented. Dividing exact covered days by a fixed annual denominator and dividing discrete enrollment months by twelve can produce different totals for partial-month entrants, retroactive enrollment, and leap years. Both can be valid; mixing them changes the rate.

The basic insured cost measure is PMPM = sum_i C_i/E_m, where C_i is the claim cost attributed to member i in the specified service period, valued on the selected financial basis. The parallel annualized rate is PMPY = 12 × PMPM for consistently defined person-month exposure. A frequency of 60 emergency visits per 1,000 member years equals 5 visits per 1,000 member months when event and coverage time windows coincide. It is not equivalent to a 6% probability that a member had an emergency visit, because repeat users create more events than distinct users.

Exposure also requires a coverage boundary. Eligibility, product enrollment, membership in a risk-bearing pool, and assignment to an accountable provider population are not interchangeable. An insured person can contribute several benefit exposures for different products or a single individual exposure for a combined medical benefit; a household policy has a different counting unit from a per-person medical contract. An employer may fund health claims while an insurer administers them; the payer represented in the model must therefore be expressly named.

A proper denominator records entry and exit, births and deaths, midyear eligibility changes, benefit transitions, and gaps. If a dataset identifies only members who submitted claims, dividing by that group systematically overstates spending per covered member and suppresses the zero-claim component. MEPS household survey files (Agency for Healthcare Research and Quality, AHRQ, current data catalog) explicitly include full-year consolidated person-level files and separate medical event files. Their existence illustrates why a researcher must join utilization information to a population denominator rather than treating a sample of positive claims as the entire insured population.

Episodes, claim lines, services, and encounters

A claim is an administrative artifact whose granularity depends on the payer, provider, and billing system. One inpatient admission may produce a facility bill, physician claims, diagnostic service lines, separate drug claims, and subsequent rehabilitation. A denied claim may be resubmitted with altered identifiers. An encounter record may report capitated services without a positive fee-for-service payment. Consequently claim count, line count, visit count, episode count, and unique-user count answer different questions.

A defensible study documents its event construction as a reproducible algorithm: identify the subject; choose clinical or administrative episode start; set allowable gaps; determine whether repeated services join an earlier episode; treat transfers, observation stays, and readmissions explicitly; retain revised and voided claim lineage; and specify when an episode remains incomplete at the observation cutoff. These choices should be stable across comparison periods or accompanied by a bridge between definitions.

For example, 1,000 claim lines may represent 250 outpatient encounters with four billed services each. Treating those lines as 1,000 treatment episodes multiplies apparent utilization fourfold. Conversely aggregating four independently required procedures into one encounter removes service intensity information that is essential for unit-cost projection. The appropriate representation often preserves a nested hierarchy: person → episode → encounter → service line → unit, with amounts and event dates at each level.

Claim completion and benefit coding create further ambiguity. Inpatient payment on a diagnosis-related group basis, outpatient fee schedules, pharmacy dispensing claims, and capitation all embody different relationships between service quantity and payment amount. Comparative modeling should preserve comparable resource units or explicitly model an institutional difference in the payment rule. The HCUP utilization methodology (AHRQ, current online documentation) also illustrates the significance of survey and database redesigns: the U.S. National Inpatient Sample changed methodology in 2012, requiring trend weights for comparable long-run inpatient estimates. A discontinuity in a data system can masquerade as medical utilization change.

The statistical unit and the financial unit

An actuarial claims dataset should have at least four linked dimensions: service occurrence time, financial valuation time, contractual benefit terms, and exposure. Let s identify service line, i member, t service period, and v valuation date. The allowed expense A_{ist} arises from adjudication of a covered service. Member liability M_{ist} may depend on all earlier eligible spending within a benefit year, so it is not necessarily a simple fixed fraction of A_{ist}. Insurer incurred benefit L_{ist}(v) must distinguish amounts fully adjudicated from estimates for services already incurred but not yet recorded.

The relation A = M + L is an accounting identity only when both sides cover the same expense base and all relevant components are assigned; uncovered services, noncovered charges, coordination of benefits, recoveries, provider risk sharing, and separate payer arrangements require further explicit fields. Gross billed charges are not an additive substitute for net allowed spending. A retrospective reconciliation should be possible from original line, adjustment, payment, reversal, and recovery transactions to the final claim value.

This establishes three connected but distinct quantities: the clinical quantity of care, the contractually recognized value of care, and the insurer's eventual cash settlement. Pricing analysis uses the second, exposure-based risk analysis may use the first or third depending on purpose, and liquidity analysis additionally requires the dates of cash flows. A model of provider efficiency should not automatically treat high allowed prices as high clinical resource use; a model of insurer loss should not substitute hospital gross charges for claim benefits.

Observability and administrative reconciliation

The data dictionary should identify eligibility tables, member identifiers, service-date conventions, claims finalization, patient cost sharing, billing units, provider keys, diagnosis and procedure taxonomies, prescription quantities, and risk factors. The MEPS file-type documentation (AHRQ, current online description) distinguishes person-level consolidated files and event files, illustrating the importance of retaining source-level grain. A sampling survey with weights and public-use restrictions must be analyzed with its survey design; an administrative claims extract may approach a contractual population census but can omit services, outcomes, enrollment corrections, and nonclaimant experience.

Data accuracy has temporal dimensions. Diagnosis and procedure codes can be backdated, codes can be split or consolidated across annual versions, provider identifiers can change through merger, and payment processing can change independently of service consumption. Missing or zero payment may indicate a genuinely uncovered event, a full deductible, capitation, a failed match, or an unprocessed claim. Inference about health need from observed paid claims alone therefore imposes a contractual observation mechanism.

The practical quality test proceeds from raw transactions to audited aggregates. Analysts should reconcile aggregate allowed and paid claims by service period to finance totals, reconcile covered lives to enrollment systems, verify uniqueness after reversals, and quantify the effect of unresolved records. ASOP No. 23 is relevant because its data-review and disclosure expectations emphasize material limitations and reasonable use of available information. Where the dataset fails reconciliation, estimation may continue for exploratory purposes, but the uncertainty must be explicitly included in the scientific interpretation.

Frequency and Utilization Measures

Rates, events, and count distributions

Let N_{is} count services of type s for member i over effective exposure T_i. The crude event rate for service s is lambda_s = sum_i N_{is}/sum_i T_i, with T consistently measured in years or months. A Poisson intensity model assumes E[N_{is}|X_i,T_i] = T_i exp(X_i beta_s); the logarithm of exposure enters as an offset, not an unconstrained clinical covariate. Its conditional equidispersion Var(N|X,T)=E(N|X,T) is frequently unrealistic for medical use, in which recurring illness, persistent patient behavior, and dependence among events create overdispersion.

A negative binomial specification provides a conditional variance greater than the mean, often written Var(N|X,T)=mu+alpha mu² under an NB2 parameterization. The dispersion parameter alpha captures residual heterogeneity at the chosen unit but does not prove a particular clinical cause. Count methods are reviewed in Winkelmann's Econometric Analysis of Count Data (2008, fifth edition). Poisson quasi-likelihood may also estimate conditional means consistently under weaker distributional conditions when its systematic mean is correctly specified, but its standard errors must reflect variance misspecification and dependence.

Zero utilization deserves careful interpretation. A zero count may arise because a covered member had no relevant illness, did not seek care, lacked local access, used services billed to another payer, or was observed for only a short time. Hurdle models separate whether use occurs from the positive count process. Zero-inflated models posit a mixture in which one latent regime yields structural zeros and another generates counts including zeros. They can fit similar empirical distributions while assigning different meanings to the zero component. Model selection should be guided by the study's causal and measurement context, not only a lower information criterion.

The frequency distribution of unique users is related to but not equal to the event distribution. If a simple Poisson model holds for a full year with lambda=0.6, the probability of at least one event is 1-exp(-0.6), approximately 45.1%, while the mean event count remains 0.6. Patient clustering and omitted exposure weaken this simple equality. Portfolio statistics should report both use prevalence and event rates because a cost increase concentrated among frequent users has different operational implications from an increase in first-time users.

Two-part and compound cost specifications

Positive medical expenditure is usually skewed, with a substantial population having zero annual claims for a given service. For total cost Y, a two-part conditional mean is E[Y|X] = Pr(Y>0|X) × E[Y|Y>0,X]. The first part can be modeled by logistic regression; the second by a positive-cost model such as a Gamma generalized linear model with log link. This specification permits the probability of any covered spending and positive spending intensity to respond differently to age, diagnosis, access, product, and provider characteristics.

An alternative compound specification writes Y_i = sum_{j=1}^{N_i} Z_{ij}, where Z is event-level cost. Under conditional independence and appropriately specified count and severity expectations, E[Y_i|X_i] = E[N_i|X_i] × E[Z_i|X_i]. If event severity depends on event count, member severity, cumulative deductibles, or unobserved morbidity, the product of two separately estimated unconditional means is generally inappropriate. A richer model needs the conditional dependence or aggregation structure explicitly represented.

The framework of Nelder and Wedderburn's generalized linear models (1972) supplies link–variance modeling that avoids an automatic assumption of Gaussian costs. Manning and Mullahy's estimator comparison (2001) emphasizes that ordinary log-linear regression, Gamma GLM, and related methods exhibit different bias and efficiency under skewness, heavy tails, and heteroscedasticity. Their published abstract supports method sensitivity; it does not establish that a single family is best for every insurance population.

For a log-cost regression fitted only on positive values, exponentiating a predicted log mean typically estimates a geometric mean rather than the arithmetic mean needed for aggregate pricing. Duan's smearing estimator (1983) supplies a nonparametric retransformation approach under its assumptions. When the residual distribution varies with X, a single global smearing factor may itself be misspecified; group-specific or model-based conditional retransformation can be required. Every predictive method must ultimately be judged on the original financial scale when predicting total dollars.

Cost sharing, provider capacity, and endogenous utilization

Utilization is partly a behavioral outcome of coverage. A visit rate depends on morbidity and on perceived and actual prices, plan network conditions, waiting time, physician referral, authorization rules, and local supply. The randomized RAND Health Insurance Experiment, summarized by Brook and colleagues (2006), found that participants facing cost sharing used fewer services than those receiving free care, with differential health consequences for selected vulnerable groups. Its design establishes credible causal information for the experimental populations and historical arrangements, not a universal current price elasticity for every medical service.

The Oregon Medicaid lottery study (Baicker et al., 2013) found increased use of several services and reduced catastrophic out-of-pocket burdens after Medicaid coverage, together with mixed measured physical-health results within the study period. These findings demonstrate that observed utilization reflects access and insured demand as well as medical need. A forecasting model that extrapolates a newly insured cohort from previously uninsured claim records would confound zero observed claims with zero demand if it ignored enrollment and coverage transitions.

Utilization can be constrained by provider capacity. A shortage of specialists, beds, laboratory throughput, or drugs may produce lower billed volume even as underlying need rises. Conversely the expansion of telemedicine, outpatient capacity, elective surgery schedules, or diagnostic access can shift the site and timing of recorded events. The analyst should distinguish potential need, realized services, and paid services. A rise in realized utilization may accompany improved access; the same numerical rise might reflect induced low-value care in another institutional context. Actuarial cost prediction alone does not identify social welfare or clinical appropriateness.

Contract design also changes observed counts. A bundled payment can turn multiple service claims into one reimbursed episode, while fee-for-service coding exposes numerous service lines. Increased use of capitation may reduce paid fee-for-service event counts without reducing clinical appointments. The model should include changes in measurement and reimbursement regime, rather than attributing all observed count movement to treatment decisions.

Temporal dependence and portfolio aggregation

Medical service events cluster in time. A hospitalization may produce imaging, surgery, outpatient follow-up, pharmaceuticals, and readmission, generating correlated costs within a person and episode. The correlation can survive risk adjustment and spread across providers. Treating service lines as independent observations yields understated uncertainty and unstable apparent significance. Inference should account for patient, household, provider, and time clustering as appropriate.

For a portfolio Y = sum_i Y_i, variance is sum_i Var(Y_i)+2 sum_{i<j} Cov(Y_i,Y_j). Diversification reduces independent idiosyncratic variability but leaves systemic covariance from epidemics, provider price shifts, reimbursement changes, and common data processing disruptions. Estimation may use hierarchical random effects, longitudinal generalized estimating equations, clustered inference, bootstrap resampling at a genuinely independent unit, or explicit covariance models. The choice depends on whether the research object is prediction of a portfolio aggregate, attribution of a provider difference, or inference on an individual risk.

Seasonal design matters. Respiratory infections, influenza seasons, school cycles, and elective procedures affect service-date frequencies. A calendar-month model should distinguish service occurrence from reporting and payment timing, because claim submission seasonality may reflect billing offices rather than care. The estimate should compare like months after appropriate adjustment and test for administrative shocks rather than relying on a naive twelve-month average.

Portfolio aggregation has an accounting requirement: if a member has four episodes, the member's annual cost equals the sum of episode benefits only when reversals, capitation, coordination, and all other attributed amounts are reconciled. Cohort-level dollar predictions should be summed from individual predictions using weights consistent with enrollment. Averaging subgroup PMPM values without exposure weights generally produces the wrong portfolio rate.

Unit Cost and Severity

Billed, allowed, paid, and incurred amounts

Cost has several financial definitions. Billed charge is the provider's submitted amount. Allowed expense is the adjudicated amount eligible under the applicable pricing arrangement. Paid amount is the cash transferred by the particular payer at a specified time, net or gross according to the chosen convention. Incurred claims reflect recognized service-period obligations including late and unreported emergence. An insurance benefit may include direct payments, provider capitation, delegated-risk settlements, or related arrangements that do not correspond one-to-one with individual claim lines.

For a member with an annual in-network deductible D, coinsurance c after deductible, and no other accumulation or exclusions, aggregate member liability on eligible allowed amount A is M(A)=min(A,D)+c max(A-D,0), subject to any applicable in-network out-of-pocket limit. The illustrative insurer benefit is L(A)=A-M(A). Actual contract valuation also needs copayment priorities, separate deductibles, formularies, family accumulators, coordination, noncovered amounts, benefit limits, and statutory protections. Therefore a line-by-line simulation may be required even to obtain correct average liability.

Severity analysis must state whether Z is allowed cost per service, allowed cost per episode, insurer cost per episode, annual member spending, or a benefit paid for a particular contractual contingency. These variables have different tails and depend differently on plan design. A frequency–severity model fitted to annual member claims may be valuable for premium forecasting yet uninformative about the negotiated unit price of one laboratory test.

The ASOP No. 5 incurred-claims guidance (Actuarial Standards Board, 2017 online adopted version) addresses relevant health claims estimates, reporting and processing periods, seasonality, operational changes, and the need for documentation. It supports the scientific distinction between actual incurred experience and partial observed payment development. Its scope is defined for specific U.S. actuarial work and should be applied with attention to effective version.

Unit prices, service intensity, and provider contracting

Define service family s with utilization Q_{st} and observed mean allowed unit price P_{st}. The basic identity A_t = sum_s Q_{st} P_{st} is exact if services are exhaustive, mutually exclusive for this accounting purpose, all line amounts are correctly allocated, and P is calculated as category allowed spending divided by category quantity. It is an accounting decomposition. It becomes an economic measure of price change only when units have a stable definition and comparable case composition. A shift from straightforward imaging to high-complexity imaging raises the observed category average even with unchanged price for every individual procedure.

Service intensity can be measured in units per encounter, work relative value units where appropriate, cost-weighted episodes, or another explicitly defined resource metric. An intensity index must avoid counting the same payment twice across nested units. Coding intensity changes can reflect genuinely different severity, documentation, incentive responses, or a new codebook; no one interpretation follows from the metric alone.

Provider price variation is substantively important. Meiselbach and colleagues' hospital price comparison (2023) used disclosed U.S. negotiated prices and found commercial prices commonly substantially above Medicare Advantage prices within matched hospital–insurer–service cells. The paper's descriptive design, sample coverage limitations, and lack of utilization counts are essential to its interpretation. The finding demonstrates that the payer's price environment affects cost at fixed service, rather than identifying a causal explanation for all differences.

International accounts further complicate price. Tax-financed hospital global budgets, negotiated social insurance tariffs, public-sector salaries, capitation, bundled episodes, and private fee schedules distribute resource costs differently. An apparent zero marginal insurer claim for a capitated visit does not imply zero economic cost. The international health-accounting framework records financing and functional categories that can support principled system comparisons, but transposing them into claim-level unit prices requires carefully defined bridges.

Skewness, high-cost events, and prediction on the currency scale

Annual insured medical cost distributions typically contain many zeros, a long right tail, and a few exceptionally expensive individuals. Means determine aggregate funding, while medians characterize typical member experience. For example, if nine members cost $100 each and the tenth costs $9,100, mean expense is $1,000, while median expense is $100. Neither number is an error; each answers a different question. Any tail cap, winsorization, exclusion, or reinsurance adjustment modifies the estimand and must be documented.

Lognormal, Gamma, inverse Gaussian, Tweedie, and compound distribution models can represent different cost features. Their names alone do not resolve the fundamental data problem: the source variable may combine heterogeneous services, limits, claims processing, and provider arrangements. Tail-aware alternatives can include mixtures, quantile analysis, excess-of-threshold estimation, or separate catastrophic claim modules. Extreme-value methods have domain-of-attraction and threshold assumptions, and insurer-specific stop-loss terms change net severity. A mean estimate is sometimes highly sensitive to a small number of real cases, so removing them simply to stabilize a model may eliminate essential insured risk.

A useful predictive report provides both expected aggregate cost and tail-risk measures under explicit assumptions, with a direct distinction between gross and net payer liability. It should assess the calibration of large claims in holdout data and not let excellent median prediction disguise systematic underestimation of very expensive episodes. The Manning–Mullahy model comparison and Duan retransformation method remain relevant because actuarial costs must be aggregated on a currency scale. An apparently well-fitted log-cost relationship can be financially biased when retransformed without attention to heteroscedasticity.

Exact decomposition of spending changes

For a homogeneous service category, expense in period 0 is A_0=Q_0 P_0 and in period 1 is A_1=Q_1 P_1. The exact difference can be allocated symmetrically:

Quantity effect = (Q_1-Q_0)(P_0+P_1)/2.

Price effect = (P_1-P_0)(Q_0+Q_1)/2.

Their sum equals A_1-A_0. This is the two-factor symmetric decomposition, equivalent to averaging the two possible sequences of updating price and quantity. It allocates the interaction without arbitrary dependence on change order. It is an arithmetic identity; interpreting price and utilization causally demands additional research.

Suppose Q_0=1,000 units, P_0=$200, Q_1=1,100 and P_1=$220. Expense rises from $200,000 to $242,000. The symmetric quantity effect is $21,000 and the symmetric price effect is $21,000, summing exactly to $42,000. The example illustrates a 21% expense increase with 10% increases in both unit quantity and unit price. Quoting only a 10% price trend would miss half of the actual dollar change under the stated decomposition.

For multiple service types, apply the identity within properly harmonized cells and sum contributions. If services enter or exit the taxonomy, price is undefined for the absent cell, and a new-service convention or explicit entry/exit component is required. A change in average cost per visit is likewise decomposed into price per service, services per visit, and mix of visit types when the data permit. This distinguishes an actual contract price increase from higher clinical complexity or a reallocation toward more expensive modalities.

Indices based on fixed baskets, chain weights, or reimbursement schedules can each be appropriate for a stated goal. A fixed basket helps estimate like-for-like prices; a chain-weighted expenditure index better reflects evolving compositions but can have drift and comparability problems. At each stage analysts must specify the observational unit, normalization and weights. Statistical decomposition cannot by itself establish whether a higher price purchased improved quality.

Case Mix and Risk Adjustment

Population composition, health status, and classification

Case mix is the distribution of health states, expected morbidity, ages, benefit types, and other characteristics that affect expected cost. If one portfolio gains more chronically ill members, observed PMPM may increase without a change in prices or behavior for otherwise similar members. Conversely a healthier enrollment mix can conceal rising service prices. Risk standardization is thus central to separating within-group experience changes from portfolio composition changes.

Let g index mutually exclusive population strata; w_{gt} is the exposure share in period t and c_{gt} is expected spending per exposure in that stratum. Aggregate PMPM is C_t=sum_g w_{gt} c_{gt}. Its change contains both changes in stratum costs and changes in the exposure weights. To compare costs holding population mix fixed at reference weights w_g*, define C_t^std=sum_g w_g* c_{gt}. A standardized trend describes the experience of a fixed reference mix; a forecast of actual funding should use projected future mix. Confusing these targets causes errors even when each subgroup estimate is unbiased.

Suppose 700 member months are in a low-risk group at $100 PMPM and 300 in a high-risk group at $500 PMPM, producing $220 PMPM. A second period with 600 low-risk and 400 high-risk member months at unchanged group costs produces $260 PMPM. The $40, or 18.18%, increase is entirely a composition effect in this example. It would be incorrect to interpret it as an 18.18% increase in unit prices or care volume among equally situated members. If the high-risk group's cost were simultaneously changing, risk mix and within-group changes would need a stated allocation convention.

Risk classification itself is purpose-specific. ASOP No. 12 (Actuarial Standards Board, 2011 revision page) discusses classification principles, objectivity, and risk-related characteristics within its scope. A diagnosis code may be predictive but susceptible to selection, documentation or coding incentives; a demographic category may be empirically associated with cost yet restricted by law or inconsistent with the purpose of a financing system. Prediction and permissible use are separate questions.

Prospective and concurrent risk models

A prospective risk model predicts future-period cost from information available at an earlier cutoff; a concurrent model uses information from the expenditure period. Their forecast targets and allowable covariates differ. If diagnosis codes generated by the event being predicted are included as prospective predictors, information leakage makes apparent out-of-sample accuracy meaningless for actual prospective decisions. Features require recorded availability dates, not only service dates.

The U.S. CMS Medicare risk adjustment program publishes model documents and diagnosis mappings across successive years. Such systems demonstrate how demographic variables, diagnosis groups, calibration weights, coding rules, and payment years are jointly specified. They are institutional payment methodologies, not universally valid disease severity indices. A risk score built for one program's covered benefits and coding may be miscalibrated for another insurer, country, or population.

A common actuarial mean specification is E[Y_i|X_i]=exp(alpha+X_i beta) for nonnegative outcomes with a suitable link and variance family. Age functions, prior utilization, morbidity indicators, insurance type, geography, benefit generosity, and provider access can be included if available before the forecast and permitted for the intended purpose. Interactions may be important, for instance between comorbid conditions and benefit design. Strong correlation among diagnoses requires regularization, clinically defensible grouping, and scrutiny of feature instability.

It is useful to assess calibration separately from rank discrimination. A model may put expensive members near the top of its score distribution while underpredicting every category by 15%; this ranking can be useful for one research question but unsuitable for estimating total required claims funding. Conversely a constant PMPM forecast may predict an aggregate accurately while offering little member-level risk differentiation. Both the estimand and loss function should be stated before evaluating a model.

Provider effects and incomplete adjustment

Provider-level differences combine negotiated prices, treatment intensity, referral practices, patient selection, local capacity, and underlying disease severity. The unadjusted comparison of mean costs at two hospitals cannot isolate a provider practice effect. Provider attribution can be ambiguous when care is delivered by multiple teams or facilities and when severe cases are systematically referred to tertiary centers.

Multilevel models can represent member, provider, and geographic variation. In a simple conditional model, log E[Y_{ip}|X_i] = X_i beta+u_p, where u_p is a provider-specific effect. This association may support prediction, but treating u_p as an estimate of provider-caused spending requires stronger identification conditions. In particular referral patterns, coding intensity, and unobserved severity may correlate with provider assignment. Fixed effects, longitudinal changes within a provider, natural experiments, and instrumental-variable designs have different identifying assumptions that require independent scrutiny.

Provider contracting can make identical services carry different allowed prices, as illustrated by Meiselbach et al. (2023). Consequently a well-calibrated morbidity model need not predict spending under a new network where negotiated prices differ. Updating a network requires a separate contractual price bridge and a possible utilization response. The transfer is not purely a statistical intercept change if provider mix, access, and coding also evolve.

Provider effects may create correlated tail risks. A specialty network that treats severe chronic disease can concentrate expensive drugs and complex admissions. Extreme claims need to be identified and evaluated as parts of the same clinical pathway when a coherent episode definition is possible. Independent service regression without episode-level validation can understate the dependence of specialist consultations, diagnostics, and treatments.

Fairness, calibration, and transfer across systems

Model calibration can be measured by observed-to-expected ratios O/E within transparent strata, with uncertainty intervals where sample size is limited. For group g, O/E_g=sum_{i∈g}Y_i / sum_{i∈g}Yhat_i. A portfolio O/E near one can coexist with systematic underprediction of some subgroups and overprediction of others. Reporting should therefore include deciles of predicted risk, sex and age strata when appropriate, geography, benefit regime, enrollment duration, and the most consequential service categories, subject to lawful data governance and adequate cell sizes.

Predictive input variables can reproduce unequal historical access. Fewer observed claims may signal unmet need rather than lower morbidity; a model trained solely on payments may therefore predict lower cost for groups who previously encountered barriers to care. The OECD access indicators (OECD, 2025) document variation in unmet need and financial protection. Those differences warrant testing whether insured spending is an appropriate proxy for the intended concept of risk or need.

Fairness also requires distinguishing allocation questions. A model for aggregate claims budgeting, risk equalization among insurers, morbidity research, individual underwriting, and provider quality payment may be subject to different legal and ethical constraints. A predictive risk factor that improves actuarial forecast precision can nevertheless be impermissible for a particular decision. A monograph of cost modeling can establish what is measurable and identify risk of bias; lawful deployment requires the rules of the actual jurisdiction and contract.

Transportability has three parts: new population distribution, changed conditional cost relationships, and different institutional payment rules. Simple recalibration of the mean may address the first only partially and will not correct unknown care access or benefit exclusions. When transferring a model from one national health system to another, analysts should reconstruct the outcome definition, eligibility, codes, provider prices, calendar, and reimbursement mechanisms before relying on parameter similarities. International comparability begins with measurement discipline.

Forecasting Medical Cost Trend

Experience trend and prospective trend

Observed medical trend is a measured change in claims expense between comparable periods; prospective medical trend is the expected change under a future configuration of exposure, benefits, prices, disease patterns, technology, and care delivery. A retrospective growth rate is a starting datum, not automatically a forecast. Slater's medical trend decomposition discussion (Society of Actuaries, 2026) distinguishes persistent growth, adoption processes, discrete level changes, and temporary departures. That account is useful as an industry interpretation and should not be mistaken for externally validated universal parameter values.

For strictly comparable cohorts with positive PMPM expense, observed trend g=(PMPM_1/PMPM_0)-1. The analyst must identify the dates: service periods, runout through valuation, exposure coverage, claims gross or net of capitation, and any benefit or network normalization. Comparing a fully run-out year with a recent incomplete year usually biases the apparent recent trend downward. Comparing an incurred basis in one year to a paid basis in another is conceptually invalid.

A multiplicative representation for analytic attribution can write C_{t+1}/C_t=(1+g_P)(1+g_Q)(1+g_I)(1+g_M)(1+g_B) for price, frequency, within-encounter intensity, morbidity mix, and benefit or payer-share factors, provided the factors have been defined in a compatible sequential index or counterfactual calculation. With independently stated factors, their product is a projection convention rather than a unique empirical identification. Interaction terms are material, and different orderings require a documented allocation method when assigning dollar contributions.

If a portfolio assumes a 5% price change, 3% service quantity change, and 2% intensity change, then the combined factor is 1.05×1.03×1.02=1.10313, or 10.313%. Their simple sum, 10%, understates the combined effect by 0.313 percentage points. These figures are hypothetical. Forecasting should also consider whether increases in intensity are already embedded in per-episode price, because double counting the same mix change can yield a much larger bias than this arithmetic interaction.

Price, use, intensity, mix, and coverage interaction

A medical trend study should distinguish five mechanisms. First, a true change in the reimbursement price for the same service at the same provider and contractual tier. Second, a change in how frequently the covered population uses a service. Third, a change in the number or complexity of services per encounter. Fourth, a change in patient health composition or the distribution of providers and therapies. Fifth, a change in the share of allowed expense borne by the insurer due to benefits, deductibles, limits, and coordination with other payers.

These mechanisms are empirically interdependent. An improved drug can be more expensive per prescription and reduce hospital admissions. A new deductible may reduce low-cost utilization and increase average cost per observed episode because the least severe visits disappear. A facility merger can raise negotiated prices and shift referral patterns. When a service is newly covered, observed insurer claims can jump despite unchanged overall health resource consumption. Causal attribution requires more than disaggregating totals.

The Health Care Cost Institute reports (HCCI, current report portal) separate spending, prices, and utilization for members of employer-sponsored insurance in covered datasets. HCCI reports a rise in average annual spending in its 2019–2023 series from $5,676 to $6,743. These figures refer to a particular commercially insured population and methodology. Their interpretation differs from the U.S. national expenditure series, whose 2024 spending level and growth describe the full health system.

Expenditure growth can also arise from adding covered people. Total portfolio benefit is exposure multiplied by cost per exposure; it follows that aggregate cost growth combines enrollment growth and per-exposure medical trend. The distinction is essential when preparing a claim-cost budget. A carrier may see its total payments rise 20% because membership expands while its PMPM remains stable; a health ministry may see national spending increase due to population aging and greater access. These outcomes are substantively different from a 20% increase in price per service.

Cohort, time, and seasonal modeling

A practical forecast is constructed on service cohorts, with development data available through a stated cutoff. Each cohort's allowed and insurer-paid expense should be projected to ultimate on a comparable basis before studying trend. Separate models may be needed for hospital admissions, ambulatory care, drugs, behavioral health, emergency care, long-term therapies, and other major categories. The granularity must balance clinical relevance, financial materiality, and credibility.

A structured model might specify log C_{st}=alpha_s+beta_s t+season_s(month)+X_{st}gamma_s+epsilon_{st}, interpreted as a conditional mean only when the link and retransformation are appropriate. Here s indexes a stable category, X contains changing exposure, risk composition, and coverage indicators, and t represents service period. An untested linear time coefficient should not be extrapolated across epidemics, coding redesigns, major reimbursement reform, or sharp technology adoption. Structural breaks deserve explicit dates and estimation.

The HCUP trend documentation illustrates a concrete longitudinal hazard: methodological redesign can require reweighting historical years. Similar problems arise after insurer system migration, merger, revised network pricing, drug rebate treatment, or the introduction of a new episode grouping algorithm. An analyst should maintain both the original history and a restated comparison series, document the bridge, and test whether reported trend changes are robust to the adjustment.

Prospective forecasting should use information actually available before the forecast origin. If final claims development through a future period or diagnosis details learned after renewal are used as predictors, out-of-sample results acquire look-ahead bias. Realistic backtesting recreates historical decision dates, contemporaneous data completeness, and the model versions that could then have been fitted. Random row-level splits of claims from the same member across training and test samples may further exaggerate performance.

Scenarios, uncertainty, and validation

Uncertainty has several sources: random claim events, cost tails, parameter uncertainty, data revisions, model form, intervention effects, price negotiations, and scenario shocks. A single trend value cannot represent all of these. A forecast report should show a central estimate, sensitivity to major assumptions, and relevant prediction intervals or stress tests. For insurer claims funding, capital and liquidity questions can require the distribution of aggregate payments, not only the expected medical trend.

An illustrative scenario set may vary unit-price increases, service utilization, high-cost drug adoption, new membership mix, reimbursement changes, and claims completion speed separately. Parameter uncertainty can be explored through resampling at independent member or provider clusters where appropriate; process variance is generated through conditional event simulations; model uncertainty is tested with reasonable alternative functional specifications. The relationship among sources of uncertainty should be recorded to avoid double counting common shocks.

Backtesting assesses whether expected ultimate PMPM matches subsequently matured observed ultimate PMPM. It also measures calibration by service category, morbidity group, geography, provider and tail strata, and the stability of important coefficients across renewal periods. A good forecast should be able to explain the sources of error when it fails. The ASOP No. 56 modeling standard and ASOP No. 41 communications guidance support documenting model purposes, limitations, assumptions, and responsible communications under their professional scope.

There is a scientific boundary between prediction and evaluation. A policy change may coincide with declining hospital utilization, but without a suitable comparison design the data do not establish that the policy caused the decline. Causal evaluation can use randomized studies, controlled interruptions, difference-in-differences, or other designs when their conditions are defensible. The RAND experiment and Oregon lottery analysis show why carefully designed studies are relevant to utilization responses. Their historical findings cannot substitute for empirical evaluation of an unrelated insurer's current intervention.

Claims Cost Calibration

Integrated measurement specification

A complete claims cost model must connect five layers in reproducible order. First, define the insured exposure population and coverage intervals. Second, identify health events and measurable services within those intervals. Third, assign service volumes, classifications, and provider identities. Fourth, determine allowed expense and contract-specific insurer liability. Fifth, project incomplete reporting, adjudication, and payment to the target incurred or cash basis. The resulting model may be implemented as separate statistical modules, but its final predictions must reconcile to one designated portfolio output.

A concise mathematical representation for expected service-period insured benefit is E[L_t]=sum_i sum_s E[N_{ist} × U_{ist} × P_{ist} × R_{ist}], where N is event count, U is units per event, P is comparable unit allowed price, and R is the resulting insurer share only when an effective proportional interpretation is defensible. The expectation of the product is not ordinarily the product of marginal expectations. For nonlinearly accumulated deductibles and out-of-pocket caps, calculate L from the full member-year sequence of allowed services and applicable rules instead of multiplying by a single average share.

This can also be expressed as a hierarchical simulation. Generate eligible person time; predict episode initiation; conditionally generate service bundles and units; assign price based on network, contract, geography and date; adjudicate benefit at member level; simulate runout and cash settlement. Every component should retain a mapping to observed data, and aggregate totals should be checked against financial records. The simulator is an implementation pattern assembled from established statistical and contract accounting concepts, not a newly commissioned proprietary doctrine.

Model architecture must remain adequate to its intended use. A simple category-level exposure × utilization × allowed price model may be superior for an annual aggregate budget when available data are sparse and benefits are stable. A detailed individual-level model becomes necessary when deductible accumulation, disease pathways, provider contracts, and risk sharing materially determine liability. Complexity increases opportunities for error; methodological sophistication only has value when predictive performance, clarity, or decision relevance improve.

Worked exposure, benefit, and trend examples

Consider an illustrative health plan with 900 people insured for all twelve months and 200 people insured for six months. Recorded member months are 900×12+200×6=12,000. Exposure is exactly 1,000 member years under the member-month convention. If 450 emergency visits occur, the rate is 450 per 1,000 member years, 0.45 per member year, or 37.5 per 1,000 member months. The distinct-user proportion cannot be determined from 450 visits without member-level counts.

Suppose this portfolio has $3,000,000 in mature insurer benefits attributable to the same service year. PMPM is $3,000,000/12,000=$250 and PMPY is $3,000. If $240,000 of additional late-incurred benefit is expected from services in that period, the revised incurred estimate is $3,240,000, yielding $270 PMPM. Treating the $240,000 as new service utilization in the following year rather than development of the previous year's incurred cost creates an artificial time trend.

Consider a member with $4,000 in eligible in-network allowed charges, an unspent $500 deductible, 20% coinsurance after the deductible, and no relevant cap reached. Member liability is $500+0.20×$3,500=$1,200; insurer benefit is $2,800. If a one-time $100 charge is excluded from covered allowed benefits, it must not be added to the $4,000 eligible base without changing the calculation. If a separate copayment, family deductible, or prior accumulation exists, the arithmetic must be recomputed from the actual benefit wording.

Now suppose a service category changes from 1,000 units at $200 to 1,100 units at $220. The symmetric decomposition attributes exactly $21,000 each to price and quantity. Suppose a normalized portfolio projection additionally assumes price 5%, service frequency 3%, and intensity 2%. The combined hypothetical increase of 10.313% yields $2,206.26 from a $2,000 baseline monthly category budget, assuming no membership, risk mix, benefit, or other change. The result is derived by multiplication, not by treating the component percentages as empirical facts.

Finally consider the two-group example developed above. With unchanged conditional costs, changing the exposure mix from 70% low-risk/30% high-risk to 60%/40% increases PMPM from $220 to $260. Reweighting both periods to a fixed 70/30 reference leaves standardized PMPM at $220 in each. If instead high-risk unit cost subsequently rises from $500 to $550 while low-risk cost remains $100, the second-period unstandardized PMPM becomes $280 and reference-mix standardized PMPM becomes $235. The observed total increase from $220 to $280 combines mix and within-group changes; the standardized increase of $15 is a 6.818% change on the reference mix. The decomposition makes clear which trend estimand is being reported.

Incurred-to-paid reconciliation and reserving boundary

Claims arrive and settle with reporting and payment lags. Let I_t represent ultimate claims incurred by service period t, P_{t,v} cumulative payments for those claims through valuation date v, and U_{t,v}=I_t-P_{t,v} the remaining expected unpaid amount, under a consistent convention for recoveries, expenses, and liability recognition. This identity is a useful simplified bridge; real reporting categories may require separate case reserves, incurred-but-not-reported estimates, and processing liabilities.

A development triangle can organize cumulative paid or reported amounts by service month and development age. Simple age-to-age factors project immature cohorts using historical development patterns. Their validity depends on stable reporting, adjudication, policy benefits, contracts, seasonality, and the maturity mix. For health claims, changes in electronic submissions, pharmaceutical billing, claims processing backlogs, or emergency utilization can materially alter patterns. A stable-looking trend on a paid basis may therefore be an artifact of accelerated or delayed settlement.

ASOP No. 5 discusses incurred health claim estimation, use of utilization information, changes in administrative practice, credibility, and disclosures. Its technical scope is especially pertinent when service-period trends are calculated from claims at different development ages. Reserve valuation is a neighboring actuarial subject; this dossier treats runout as a necessary source of bias and calibration requirement for cost and utilization models, rather than giving a full reserve methodology.

Portfolio funding and liquidity also differ. Expected ultimate medical cost may be stable while the timing of settlement accelerates, increasing near-term cash needs. Conversely processing delays may temporarily reduce cash payments while unpaid incurred obligations grow. Prediction of service use, incurred benefit, and payment cash flow should be reconciled but not collapsed into one response variable.

Diagnostic tests, auditability, and comparative interpretation

A scientific quality review should begin with identity and completeness: eligibility-to-claim linkage, distinct member counts, duplicated claim lineage, negative adjustments, missing provider fields, service-date mapping, coding versions, capitation treatment, and reconciliation to actual financial totals. Numerical validation then tests frequency distribution, zero rates, severity tails, observed-to-expected dollar totals, risk strata, provider concentrations, seasonality, reporting age, and claim-category consistency.

The key predictive measures should be chosen for purpose. For aggregate pricing, dollar-weighted bias and PMPM calibration matter directly. For risk ranking, rank correlation and concentration can be informative but insufficient. For event frequency, Poisson or negative-binomial deviance, dispersion, and conditional residuals matter. For claim severity, quantile calibration, tail errors, and original-scale prediction are essential. For probabilistic forecasts, coverage of predictive intervals across independent future cohorts is more informative than an in-sample goodness-of-fit figure.

A credibility procedure may blend sparse experience with a relevant broader population. ASOP No. 25 (Actuarial Standards Board, 2013 revision) provides principles for assessing predictive value and communicating procedures. A purely mechanical volume threshold should not be equated with reliable experience when the provider, coverage, or population is rapidly changing. The external reference data must resemble the target in relevant exposure, benefit, price, clinical mix, and time dimensions.

International application requires context. The OECD cross-country indicators and System of Health Accounts help locate system-wide service, financing and expenditure measures, while MEPS survey data and HCCI reports illuminate narrower U.S. samples and insurance coverage arrangements. No cross-national interpretation should treat a public hospital budget, employer-sponsored claims, and national total spending as interchangeable. Harmonization may require a common service basket, purchasing-power or currency conversion, stable diagnostic definitions, and explicit financing-boundary adjustments. These choices change the estimand, and their sensitivity should be reported.

A substantive editorial audit of the current synthesis identifies the following principal interpretive safeguards. Service-line utilization must not be called incidence of disease; observed insurance payment must not be equated with complete resource cost; raw price differences must not be interpreted as provider-caused effects; a regression risk factor must not automatically be treated as causally responsible; a cohort mix shift must not be labeled medical inflation; and a developing paid series must not be treated as mature incurred experience. These are research conditions supported by the analytical distinctions established across the dossier, not a new doctrine.

Research limitations and reproducibility conditions

The manuscript synthesizes established actuarial and health-econometric methods from selected foundational, institutional, and peer-reviewed sources available at the research cutoff. It does not report an original insurer-level claims sample, confidential contracts, linked clinical outcomes, an exhaustive national reimbursement survey, or a fully estimated predictive model. Selected journal papers were reviewed at the level of accessible publisher text or abstracts, while principal methodological statements were independently limited to what those materials support. Operational ASOP descriptions are jurisdiction-specific and do not constitute a universal global regulatory conclusion.

All arithmetic demonstrations are explicitly hypothetical. A new empirical study should preregister or at least version the target estimand, data extraction, eligibility and benefit rules, code groupers, price and utilization definitions, modeling families, split dates, validation strata, treatment of zeros, censoring and claim development, and material model changes. It should retain reproducible code and data dictionaries under appropriate privacy protections and provide external replication where lawful. Absent those inputs, parameter values and premium effects cannot be generalized to an identified insurer.

The major unresolved scientific questions for future data-driven research concern joint identification of price and quantity under bundled payment, modeling rare technology-driven spending shifts, consistent international case-mix bridges, provider selection bias, omitted unmet medical need, and forecast resilience under rapidly changing enrollment. These are research frontiers and limitations, not defects concealed in the current analytic examples. A public release of this manuscript would require a separate editorial authorization and production verification process, independent of the present substantive investigation.

Conclusion

Health claims cost and utilization are linked by a sequence of observable decisions and accounting transformations. Eligible people experience conditions; some seek and receive care; providers deliver treatments and encode services; payment systems assign prices; contracts allocate expense; and claims administrators recognize and settle obligations. The actuarial estimate is reliable when its denominator, event definition, valuation basis, and development stage correspond to the question being asked.

The mathematical analysis establishes that service frequency, intensity, unit price, case mix, payer share, and reporting runout each affect observed spending, often through interactions. Count and two-part models clarify the extensive and intensive margins of utilization. Generalized linear and retransformation methods address skewed monetary outcomes. Symmetric price–quantity decomposition provides exact arithmetic attribution under stable units, while risk standardization distinguishes a population-composition change from a within-group cost change. Prospective forecasting adds the independent task of judging which observed effects will persist, reverse, or change under future contracts and care delivery.

The empirical and institutional evidence shows why health claims cannot be interpreted as purely clinical demand or as a universal measure of health system cost. National accounts, survey data, insurer administrative records, provider pricing studies, and clinical utilization datasets capture different populations and expense boundaries. They can inform one another when the mapping is explicit. The scientific task is therefore both statistical and institutional: to specify what each variable means before estimating how it behaves.

The integrated methodological conclusion is practical and testable. A defensible health claims cost model starts with correctly measured exposure, retains nested service and episode structures, separates quantities from prices and benefits, controls morbidity composition without overlooking unequal access, forecasts on a fully developed financial basis, and validates results against real future cohorts. Its claims are proportionate to the data it has seen. This synthesis consolidates established actuarial science and medical expenditure research into a coherent explanatory account without asserting a new theory or an unverified discovery.

References