RESEARCH DOSSIER
Motor Vehicle Years Mileage and Exposure Measurement
Abstract. Motor insurance exposure may be measured in contract years, vehicle-years, driver time, covered miles or kilometers, trips, and vehicle operating hours. Each denominator describes a different dimension of insured risk. This monograph develops the conceptual and mathematical relationships among these measures, specifies the population and coverage conditions necessary for defensible motor claim frequencies, examines the observation and selection mechanisms of mileage and telematics systems, and evaluates the comparability of insurer claims with road-traffic statistics. The research analyzes actuarial practice standards, administrative insurance statistics, road-traffic measurement systems, and peer-reviewed empirical studies in the United States, the United Kingdom, Belgium, Canada, and Spain. It presents explicit exposure integrals, frequency transformations, stratified comparisons, constructed numerical counterexamples, and validation checks. The synthesis demonstrates that annualized and per-mile claim rates can rank the same groups differently, that distance is predictive under conditions varying by population and model, and that coherent actuarial inference requires alignment of contracts, travel, event definitions, and recording periods. The numerical illustrations are hypothetical. No original insurer microdata, independently estimated causal effect, new authorial theory, or external peer review is claimed.
Keywords: motor vehicle years; car-years; earned exposure; risk units; driving distance; mileage; vehicle miles traveled; policy-years; driver exposure; trip frequency; telematics; pay-as-you-drive; claim frequency; loss cost; actuarial measurement; exposure bias; automobile insurance.
Introduction
The denominator of a claim frequency determines what the resulting rate measures. A statement that one class of vehicles produces twice as many claims as another remains ambiguous until the insured objects, coverages, observation periods, driving distances, drivers and countable claim events are identified. Two cars can be insured throughout the same year while one travels a few hundred miles and the other tens of thousands. Their time-based insured exposure is equal, while their movement-based exposure differs by orders of magnitude.
The ASOP 53 exposure-base definition describes the exposure base as a unit used to measure future costs of risk transfer and emphasizes its relationship to those costs, practicality and verifiability (Actuarial Standards Board, 2017, sections 2.2 and 3.6). The actuarial discussion in Basic Ratemaking (Werner and Modlin, 2010) establishes the framework in which earned exposure, claim frequency, severity and expected losses are evaluated together. This framework remains necessary when modern sensors produce detailed, but unevenly observed, travel data.
Time-based exposure originates in the contract ledger: the dates for which an insurer bears defined obligations. Distance-based exposure arises from actual travel: miles or kilometers driven during covered periods. Trip counts capture beginnings, endings and travel purposes, while active driving hours capture time spent operating. None of these measures is interchangeable without assumptions about vehicle use, exposure to hazards, and the event being counted. For parked-vehicle theft, driving miles may be only a partial explanation; for moving collision events, it is often an informative factor.
The administrative distinction is visible in authoritative datasets. The U.S. Traffic Volume Trends program estimates public-road vehicle-miles from traffic monitoring, while the NAIC auto insurance statistical description records earned insured exposures, incurred losses and claims. One dataset is built around movement on roads; the other is built around insurance obligations and claim accounting. Joining their numerators and denominators requires a justified bridge between populations, periods and definitions.
The research program addresses three central questions. How should vehicle-year and driver exposure be reconstructed from contracts and movement data? How can observed mileage be distinguished from missing, self-selected or behaviorally induced travel, and how can its relation to claims be estimated? What comparisons of claim frequency survive changes of denominator, coverage, legal system and statistical sampling?
This work is a comparative interpretive and mathematical monograph. Its numerical demonstrations are constructed and individually reproducible. Empirical claims remain attributed to the cited investigators. The study does not reestimate proprietary telematics datasets or claim an independently validated national insured-mileage panel. Its aim is to establish the requirements for correct measurement and defensible inference, while preserving the evidentiary differences between actuarial, transport, legal and academic sources.
Conceptual, Historical, and Primary-Source Foundations
The scientific object
Motor insurance exposure is a measurable quantity assigned to a defined insured object and coverage over an observational interval. An earned vehicle-year denotes one covered vehicle present for one stipulated year-equivalent. A covered vehicle-mile denotes travel by an insured vehicle during a period within the selected contract conditions. A trip denotes an operational movement separated by an explicitly declared trip-boundary rule. Driving time measures hours of vehicle operation; driver exposure attributes activity to a person rather than exclusively to a vehicle.
The unit must be identified before rate calculation. A single vehicle may have several named drivers. Their combined time listed on the policy could exceed one driver-year even though their vehicle contributes only one distinct covered vehicle-year. One driver may use several insured vehicles; attributing the entire driving history to each multiplies mileage artificially. If a car carries four occupants for five miles, it has traveled five vehicle-miles, whereas occupants together have accumulated twenty person-miles. The National Household Travel Survey carefully distinguishes household travel, vehicle trips, individual travelers and trip duration (FHWA, 2026). Such distinctions transfer directly to the integrity of insurance denominators.
An insurance claim is not identical to a crash, an injury, a collision or a repair invoice. One crash can trigger several insurers and multiple contractual coverages, or no insured claim when damage falls below deductibles. A stationary vehicle may generate a comprehensive claim from theft, weather or fire. A liability claim may depend on driver status, legal responsibility and applicable policy terms. Therefore, even perfectly measured mileage cannot yield an all-coverages frequency unless the relationship between the event and driving activity is appropriate.
Development of exposure measurement
Contract registers have historically made effective dates, premium accounting and covered units readily observable. A vehicle-year can be reconstructed even where the insurer lacks direct records of annual mileage. The CAS ratemaking treatment formalizes earned exposure as a central element in ratemaking, with premiums, claims and expenses attached to compatible time and rating populations (Werner and Modlin, 2010). Distance adds an explanatory dimension that is useful when vehicle use varies substantially across policyholders.
PAYD insurance charges in relation to reported or verified miles or kilometers. UBI may incorporate acceleration, speed, braking, nighttime use, trip location or road context in addition to distance. The per-mile insurance analysis by Ferreira and Minikel (2012) examined Massachusetts experience associated with approximately three million individual automobile car-years for the 2006 policy year and found mileage informative alongside conventional rating characteristics. The study's estimates reflect its historical data and models; a projected 5.0%-9.5% mileage reduction under hypothetical broad PAYD adoption is a scenario rather than an observed universal causal effect.
Subsequent sensor-based work extended the analytical questions. Verbelen, Antonio, and Claeskens (2018) examined the predictive power of telematics in a Belgian product focused on young drivers, and Ayuso, Guillén, and Nielsen (2019) investigated mileage with driving behavior in frequency models. Boucher and Turcotte's (2020) longitudinal driver study used repeated observations and fixed effects to examine a distance-frequency relation. These investigations are complementary: they use different insurance populations, event definitions and empirical specifications. Their conjunction supports asking more precise questions about exposure instead of declaring one constant distance-risk coefficient for all portfolios.
Event-frequency-severity decomposition
Let N denote a specified claim count, C the corresponding ultimate losses and expenses, E the matching earned vehicle-years, and M verified covered miles from the same vehicles and periods. Then the time frequency is N/E and the mileage frequency is N/M. When N is greater than zero, mean severity is C/N, and pure premium or loss cost per vehicle-year is C/E. The familiar identity (N/E)(C/N) = C/E depends on using a consistent definition of claims and losses. Similarly,
is an exact algebraic transformation only when N, E and M represent the same covered cohort. It makes no empirical assertion that the marginal risk of each additional mile is constant. If mileage data describe only volunteering device users while claims cover the entire book, the transformation loses population integrity.
The ASOP 53 historical consistency guidance requires attention to the exposure base, historical adjustment and selection of data suitable for prospective cost estimates. That standard is U.S. professional practice guidance within its stated scope; it is not a universal statute. The ASOP 12 classification guidance concerns risk classification and is relevant to the separate task of allocating expected losses among classes. Exposure measures how much observed activity or time enters the denominator; rating variables describe differences in conditional cost.
Roadway activity versus contractual protection
Road network figures use different boundaries. The U.S. long-run highway travel series is a public-road travel history. The Great Britain 2025 estimates report 342.6 billion motor vehicle-miles in 2025, of which 262.4 billion involved cars and taxis (Department for Transport, 2026). These totals include travel by many vehicles and policy conditions that a particular insurer does not cover. Roadway miles cannot be inserted into an insurer claim rate simply because both data systems use the word vehicle.
Likewise, road fatalities include victims with and without insurance claims, pedestrians and passengers, while private insurer loss records follow policy and reporting rules. The NHTSA final 2023 road statistics give 40,901 U.S. traffic deaths and 1.26 fatalities per 100 million vehicle-miles in 2023; this is a road safety measure, not an insurance claim frequency (NHTSA, 2025). The unit mismatch is both mathematical and institutional. A transparent denominator needs identification of the covered fleet and the event process before further actuarial analysis.
Why definition changes matter
A portfolio may shift from named-driver coverage to permissive-use coverage, from personal use to delivery operations, or from ordinary collision handling to an automated claims channel. Reporting thresholds, recoveries and property versus bodily-injury benefits may also change. The resulting apparent frequency trend mixes physical risk with contract and administrative changes unless historical series are adjusted. A common label such as vehicle-year does not guarantee a common risk definition across time, coverage or jurisdiction.
A full measure should consequently carry a versioned scientific specification: the insured object, contract basis, coverage, event, time scale, distance source, reporting population and conversion assumptions. These specifications allow later scholars to interpret the statistic without relying on tacit local knowledge.
Vehicle-Year and Driver Exposure
Earned vehicle-time as an integral
For vehicle i and coverage c, define I_ic(t)=1 during legally effective covered time and zero otherwise. Under a fixed 365-day denominator, earned exposure across an interval T is
In a daily implementation the integral counts covered days. Alternative actual-calendar-year conventions replace the denominator for each year by 365 or 366 as appropriate. Each method is mathematically coherent when openly declared and applied consistently. A vehicle covered for 365 days during a 366-day calendar contributes 365/366 = 0.997268 actual-calendar-year exposure, while it contributes 1.000000 under the fixed-365 convention. The appropriate convention depends on the rating, statistical and regulatory purpose.
A policy can cover several cars. A one-year three-car contract contributes three earned vehicle-years if all cars have that coverage for the year. A single car renewed at midyear produces two policy records but, absent a gap, one continuous vehicle-year. A vehicle changing carriers midyear may create two insurer exposure records that together describe one distinct car-year across the market. An excess insurance layer can coexist with primary coverage; the financial obligations should remain separately identifiable even where distinct physical vehicle-time is not duplicated.
Overlapping policy terms need careful treatment. If one vehicle has records covering days 1-100 and 51-150 for the same coverage, the distinct covered period has 150 days under the convention that these are inclusive day labels, although the raw sum is 200. A construction based on half-open timestamp intervals would use different precise boundaries but the same principle: form the union of coverage intervals to count distinct vehicle-time. For insurer-level liability accounting, duplicated contractual layers may be tracked separately instead of silently collapsed.
Policy, accident and calendar cohorts
A policy-year cohort groups records by inception period; a calendar-year cohort groups observations by the date on which exposure occurred; an accident-year cohort groups claims by loss occurrence, with exposure aligned to its relevant risk period. Written exposure tied to policies issued in a period can differ substantially from the time-based exposure earned during it. The NAIC report specifications distinguish written premiums and exposures from earned premiums, incurred losses, earned exposures and number of claims. Rates must use numerator and denominator conventions that correspond to the intended cohort.
Claims can develop after the coverage interval ends. The analyst may use ultimate developed claims by accident period and current earned exposure for the matching accident period, while transparently explaining development methods and selected evaluation dates. Dividing paid claims in one calendar year by newly written policy units in the same year is not a generally valid frequency. Changing deductibles or claims reporting behavior can change the count of claim files without any change in accident incidence.
Vehicle and driver relations
Driver exposure requires its own observational unit. A driver listed as an authorized operator for an entire policy year may drive rarely. Two named drivers do not generate twice as much vehicle travel simply by being named. Driver-years inferred from listing periods measure the duration of eligibility, whereas driver-miles measure attributed driving activity and driver-hours measure operation duration. These quantities should be stored under different variable names and never pooled without an explicit purpose.
For households with alternating operators, trip-level driver attribution can sum to total vehicle mileage, subject to unknown or unassigned journeys. When a driver uses two vehicles, the driver's aggregate mileage is the sum of unique attributed trips across vehicles; the insurer still maintains separate vehicle coverage records. If driver identity is unavailable for twenty percent of travel, assigning it all to the most common named driver creates artificial precision. The system should preserve unassigned mileage and test alternative allocations.
Risk classification is adjacent to, but distinct from, observation. ASOP 12 provides a framework for classification systems, while the earned vehicle-year is calculated from actual coverage intervals. Comparing claim rates for age classes requires a defensible way to attribute driver exposure; the mere existence of age data in the policy does not prove who was driving at each event. Commercial fleets may have rota, dispatch and telematics identifiers that improve attribution, but they also encounter shift work, subcontracting and device misassignment.
Multiple coverages and parked risk
Each coverage can have a different exposure-response relation. An insured vehicle can incur fire, theft or storm losses while parked. Physical collision and third-party bodily-injury liabilities tend to be connected more directly to operation, yet may contain incident-specific legal and contractual complications. A motor product combining these coverages may warrant separate movement- and duration-sensitive components. This is an ordinary component decomposition within existing actuarial analysis, not a newly instituted theory.
Let s identify mutually exclusive use states such as active driving, stationary parking and other covered time. If state-specific covered durations are measurable, total covered time is the sum of these state durations. A distinct measure such as miles is accumulated only during applicable movement states. A model can include time- and mileage-related predictors for different claim types while avoiding the assertion that covered stationary time represents zero exposure.
Reproducible miniature portfolio
Consider three hypothetical vehicles during a 365-day observation year. Their insured covered days are 90, 180 and 365; their recorded covered miles are 1,200, 9,000 and 18,000. Earned exposure equals (90+180+365)/365 = 1.739726 vehicle-years, and matching mileage totals 28,200 miles. Mean mileage per earned vehicle-year equals 28,200/1.739726 = 16,209.45. Suppose exactly one matched claim is recognized within the selected policy and claim definition. Then frequency per vehicle-year is 1/1.739726 = 0.574803, while frequency per million covered miles is 1,000,000/28,200 = 35.460993.
The two rates are not competing empirical estimates. They are two normalizations of the same constructed claim record. Their transformation is recovered by multiplying 35.460993 claims per million miles by 0.01620945 million miles per vehicle-year. This example also demonstrates the administrative danger of using the three policy records as three vehicle-years, which would understate the corresponding annual claim rate.
The provenance check is as important as the arithmetic. If a car drove outside the effective cover interval, those miles belong to its full activity history but not automatically to insured mileage for that coverage. If the telematics system recorded only the first two vehicles, the rate for the entire three-vehicle book should not use their partial distance total with the full claim count. A valid denominator requires the same insured population as the numerator.
Mileage Measurement and Selection
Methods and observation errors
Insurers may observe mileage through start-and-end odometer readings, inspection records, vehicle sensors, telematics devices, GPS, mobile applications or reported annual mileage bands. Odometer readings are cumulative and may bracket uninsured time; interpolation assumes a driving pattern over the gap. Sensors can fail, reset, be reinstalled or transmit late. GPS traces can miss tunnels or suffer map matching errors. Mobile systems can confuse passenger trips with vehicle operation or stop recording when permissions change. Self-reports may round, understate or otherwise misrepresent distance.
For trip-level data, the definitions of start, stop, short interruptions and journey purpose determine counts. A sixty-mile uninterrupted highway journey and twelve urban five-mile journeys have equal distance, but different turning, parking, congestion and trip-start environments. Driving time adds a further variable: a motorist traveling thirty miles through congestion may spend much longer operating than one covering thirty miles on an open highway. Therefore miles, trips and hours support different hypotheses and require separate empirical validation.
The telematics frequency study by Ayuso, Guillén, and Nielsen (2019) incorporates distance with characteristics of use and driving behavior. Verbelen, Antonio, and Claeskens (2018) analyze observed driving context in Belgian motor insurance. Both substantiate the value of detail beyond a blunt annual mileage category within their studied populations, but neither establishes a universally transportable coefficient applicable to every vehicle class.
Offsets and exposure functions
For an observation i, a common count model uses conditional expected claims
Its log link contains log(E_i) as a fixed-coefficient offset. This states conditional proportionality between expected claims and the chosen exposure measure. A mileage-based offset substitutes measured miles M_i for E_i, requiring M_i>0 for the logarithm and assuming proportionality to miles under the other modeled variables. The assumptions are statistical, not dictated by the legal existence of an insurance policy.
One can instead compare a duration-offset model with mileage as a predictor, a mileage-offset model with contextual variables, and a flexible model of mileage response. A nonlinear function can accommodate rate differences between occasional and intensive drivers. If the vehicle's claim exposure includes parked losses, a log-mile offset cannot model a zero-mile insured vehicle with a positive non-driving event expectation unless a separate component or appropriate zero-handling rule is used. Composite coverages may need a mixture of time- and movement-sensitive risk mechanisms.
The Boucher–Turcotte longitudinal analysis (2020) used repeated observations and Poisson fixed-effects methodology and found an approximately linear within-driver distance-frequency relation under its specifications. Cross-sectional comparisons among drivers can reveal a different relation because driver type, route and car use vary together. The Ferreira–Minikel study found mileage relevant in a large Massachusetts sample, alongside class and territory. The empirical relationship therefore has to be interpreted conditional on whose distance changes, which claims are counted and whether the analysis compares different drivers or follows the same driver.
Missing data and biased mileage
Let M_i be true covered miles and M_i* recorded miles. In the additive measurement expression M_i* = M_i + u_i, errors can affect denominators and mileage regression coefficients differently. If unrecorded trips are omitted while the corresponding insurance claims remain visible, claims per recorded mile will be inflated. If device failures correlate with older vehicles, urban operation or high-risk behavior, the missingness becomes informative rather than random. Standard regression adjustments do not automatically repair that selection.
Two odometer readings across a year with a two-month insurance lapse cannot tell the insurer exactly how many of the elapsed miles were insured. Equal daily allocation is an assumption; seasonal reconstruction requires additional evidence. A program enrolling only device-compatible vehicles may exclude some vehicle classes or vintages. An automobile policy renewed with a different mobile device can create discontinuous records for an otherwise unchanged vehicle. These cases justify completeness indicators, source flags and sensitivity bands instead of artificially exact point measurements.
The ASOP 23 data quality requirements govern the selection, review and appropriate use of actuarial data within their scope. The standard specifically differentiates review from audit. An exposure dataset should record observation origin, device or odometer identity, time coverage, validations, imputations and uncertainty. It should not treat an externally supplied mileage estimate as a fully measured fact when provenance or accuracy remains unknown.
Selection into PAYD and UBI
Enrollment into mileage-based products is often voluntary. Drivers anticipating low mileage, favorable premiums or perceived safe behavior may select PAYD disproportionately. The same group could have lower observed claim frequency without changing their driving after enrollment. Self-selection, premium response and monitoring effects must therefore be distinguished. A simple comparison of UBI subscribers with other policyholders is a selection comparison until the design provides credible evidence about intervention effects.
The Massachusetts mileage analysis modeled potential travel reductions under assumed pricing responses, but its scenario is not an experiment in which every driver was randomly assigned to mileage-based insurance. A prospective causal study could consider staggered rollout, eligible comparison cohorts and pre-enrollment trends, with adjustment for changes in fuel prices, fleet composition, commuting needs and economic conditions. Even longitudinal observation can be affected by time-varying confounders and attrition.
Suppose high-mileage customers exit after a rate change, leaving a lower-mileage set of participants. Average reported mileage falls even if no enrolled driver reduced driving. Device adoption can also shift reported miles upward when previously unrecorded trips become visible. A reported change after adopting telematics is therefore a combination of physical activity, measurement and selection. Distinguishing these components is a central identification question.
Outcomes and driving context
Distance to first at-fault accident, all-claims counts, bodily injury claims and total claim cost use different random variables. Ayuso, Guillén, and Pérez-Marín's distance-to-first-accident analysis studied young drivers using GPS and survival methods, reporting associations involving speeding and nighttime driving in analyzed groups (2016). This endpoint should not be substituted for the all-coverages claim count used in a conventional frequency model. A first-event study also changes its risk set as drivers experience events or leave observation.
Road type, weather, time of day, speed environment, work use and driver concentration can affect event opportunity per mile. Mileage alone cannot attribute causality to particular behaviors. A claim may emerge from a vehicle collision, another driver's liability, parked-vehicle damage or theft, each requiring its own event and coverage definition. When several different mechanisms are combined, the output must be described as a portfolio loss rate rather than the physical probability of collision.
Data rights and minimization
Telematics can reveal home and work locations, habits and other identifiable activity. In the European Union, GDPR Articles 5 and 25 address accuracy, purpose limitation, data minimization and data protection by design for processing within the regulation's application. The EDPB connected-vehicle guidelines provide regulatory guidance for vehicle and mobility data (European Data Protection Board, 2021). These are jurisdiction-specific legal and supervisory considerations; applicability depends on the processing and governing law.
Annual mileage totals may suffice for a distance-based tariff where detailed location trajectories are unnecessary. More granular use variables require independently justified collection and retention. Technical capability to observe a trip does not itself create authorization to retain it or use it for underwriting. Measurement governance therefore includes contractual transparency, lawful processing, appropriate access restrictions, retention schedules and a replicable derivation from raw data to privacy-conscious statistical measures.
Comparability of Claim Rates
A constructed denominator reversal
Consider two hypothetical classes observed for 100 earned vehicle-years each. Class A travels 800,000 covered miles and produces 20 matched claims; Class B travels 2,000,000 miles and produces 40 claims. Class A has 0.20 claims per car-year and 25 claims per million miles. Class B has 0.40 claims per car-year and 20 claims per million miles. Class B therefore has twice the annual claim frequency yet a lower frequency per mile.
| Hypothetical group | Car-years | Covered miles | Claims | Claims/car-year | Claims/million miles |
|---|---|---|---|---|---|
| A | 100 | 800,000 | 20 | 0.20 | 25 |
| B | 100 | 2,000,000 | 40 | 0.40 | 20 |
This reversal is an arithmetic property of changing the exposure basis. It is not evidence that A's drivers are fundamentally riskier or B's drivers safer, because routes, loss definitions, deductibles, severity, reporting and vehicle characteristics are unspecified. Both results can be true and useful for their respective questions. The analyst should report the associated annual mileage distribution alongside the two rates.
Standardized comparisons
Portfolio rates may differ because the books contain different mixtures of driving conditions. Let k index comparable cells defined by vehicle, territory, road and use classes. If n_gk denotes claims and e_gk earned car-years for group g in cell k, an unadjusted time rate is the sum of claims divided by sum of earned years. A direct standardized rate can be calculated with common nonnegative weights w_k summing to one:
The selected reference weights must be stated, and sparse cells require careful estimation. A mileage-standardized counterpart substitutes matched miles for car-years and may need its own reference distribution. Conditioning and standardization can address observed mix effects; neither establishes the absence of unmeasured confounding. Denominator choice and class mix can jointly create Simpson-type reversals.
The risk classification practice standard pertains to appropriate construction and review of classes. The future-cost estimation guidance calls for practical exposure measures and appropriate organization of historical data. A claim rate can be standardized to compare frequency, while a pricing decision additionally requires claim severity, future trends, expenses, capital considerations and legally permitted risk classification.
Claims and road incidents
A police-reported crash may produce zero claims, one claim file or many claims under different insured coverages. A single claim can also reflect non-crash hazards. Injury victims, accident fatalities and insured claimants form different numerator populations. The NHTSA fatalities per VMT include a broad social road-safety outcome. The NAIC earned-claims statistics concern insurer financial and claim records. The difference in institutional recording cannot be fixed by multiplying one rate by a mileage constant.
An increase in recorded claims may reflect lower reporting thresholds or simplified digital claims processes even when accidents remain unchanged. Conversely, a higher deductible can lower observed small claims without making travel safer. Injury liability laws and policy benefits influence eligible compensation and claim-file counts. International comparison requires harmonized event definitions or explicit crosswalks from physical events to legal obligations and accounting records.
Road scope and international transport statistics
The FHWA monthly count program employs monitoring locations across the United States, and its estimates are subject to ongoing revision. The UK Department for Transport's traffic statistics methodology uses manual roadside counts, automatic traffic counters and road lengths to generate estimates. The 2025 Great Britain release disaggregates travel by vehicle and road type. These transport methods create national physical activity denominators, but do not identify whether each mile falls within a particular insurer's covered contract.
Comparisons across countries must examine public versus private roads, vehicle classes, domestic versus cross-border travel, registration practices, car-sharing, active vehicles and coverage requirements. Even a common kilometers-per-year unit is insufficient when one denominator counts insured passenger cars and another includes taxis, vans, trucks or motorcycles. An accurately converted numerical unit may still represent a different population.
Frequency, severity and total cost
Let C be ultimate losses and N matched claims. Mean severity C/N and pure premium C/E must be assessed alongside frequency N/E. A reduction in accident frequency per mile can coexist with rising repair expense, larger vehicles, medical inflation or expensive technology. Mileage-based exposure analysis should not be used to infer a monetary insurance rate without a severity and expense component.
Suppose a portfolio's annual claims per car-year rise ten percent while annual miles per car-year rise twenty percent. If populations and definitions match, claims per mile fall by approximately 8.33 percent because 1.10/1.20 = 0.916667. The calculation is an identity conditional on consistent numerator and denominator, not evidence of safer behavior. The same decomposition can be applied by class and period to identify which empirical mechanism requires further investigation.
Model uncertainty
Count observations can be modeled by Poisson, negative binomial, hierarchical or other formulations appropriate to the data. A Poisson rate N/E has approximate standard error sqrt(N)/E when its independence and mean-variance conditions hold. Repeated claims per driver, household clustering, catastrophic weather and portfolio heterogeneity often violate simple independence. Mileage may also be predicted rather than measured; uncertainty in that denominator needs to be incorporated rather than treated as fixed.
Cross-validation should preserve the structure of the exposure. Splitting trips from the same driver randomly between training and testing sets can overstate generalization. Temporal holdouts, driver-level clusters and truly new vehicle cohorts answer stronger prediction questions. Interpretation and legal permissibility must remain separate from predictive accuracy. A model can perform well for one observed insurance product and still fail when coverage, driver population, region or telemetry adoption changes.
Comparative Evidence, Data Uncertainty, and Validation
Evidence by research function
A direct insurance exposure study requires linked coverage, vehicle activity and claims data. Official traffic programs quantify movement on roads; insurer reports quantify covered exposure, claim counts and financial cost; actuarial standards establish practice and disclosure; peer-reviewed studies examine statistical relations for specific cohorts. Their sources must be used to support the claims they actually measure. The 2022/2023 NAIC auto database announcement reports a $1,438 national combined average premium per issued vehicle for 2023. That premium figure does not estimate insured miles, accident probability or individual loss severity.
The 2025 British road statistics report 342.6 billion vehicle-miles for Great Britain and use transport monitoring rather than insurer exposure records. A calculation that divides a U.S. insurer's claim count by British national motor miles would be devoid of a shared population. A subtler version of the same error occurs when an insurer uses its total claim count against distance from telematics participants, failing to remove claims and vehicle-time from customers whose mileage was not observed.
Required data architecture
A reproducible research dataset should contain a coverage ledger, a vehicle and driver identity register, a movement table and a claims table. The coverage ledger records insurer, policy, vehicle, coverage version, effective and termination timestamps, risk territory and endorsement history. The movement table records odometer or device provenance, trip times, validated miles or kilometers, attribution to vehicle and possibly driver, missingness flags and imputation. The claims register identifies accident, report and settlement periods, claim and occurrence IDs, coverage, liability status, paid and incurred costs, recoveries and development maturity.
A vehicle can have many trips, several policies and numerous claim files. A naive relational join can multiply miles or claims if each trip is joined to every claim or coverage endorsement. The analyst should aggregate each table to the intended observational unit before joining and reconcile totals after every operation. Device duplication, overlapping coverage intervals and changing driver IDs should produce explicit exceptions.
A useful dictionary states whether days are inclusive, whether actual leap-year or fixed-365 denominators are used, what is meant by a claim, how non-driving losses are treated, how a trip starts and ends, and which miles are excluded. It also records time zones, territorial coverage, loss development, calendar version, road class and the effective dates of policy and legal changes.
Deterministic numerical checks
The following tests use only constructed inputs. Their expected values follow directly from the definitions and can be reproduced with any programming language.
| Check | Inputs | Expected computation |
|---|---|---|
| Time aggregation | 90, 180 and 365 covered days | 635/365 = 1.739726 earned car-years |
| Distinct covered time | Inclusive day intervals 1–100 and 51–150 | 150 covered days, rather than 200 |
| Multi-vehicle policy | Three cars insured for 365 days | 3 car-years for each specified coverage |
| Mileage aggregation | 1,200, 9,000 and 18,000 miles | 28,200 miles |
| Claim-rate conversion | 1 claim; 635/365 car-years; 28,200 miles | 0.574803 per car-year; 35.460993 per million miles |
| Group A distance rate | 20 claims / 0.8 million miles | 25 per million miles |
| Group B distance rate | 40 claims / 2 million miles | 20 per million miles |
| Distance units | One million miles | 1.609344 million kilometers |
| Calendar convention | 365 days in a 366-day year | 365/366 = 0.997268 actual-year units |
Rounding should occur after aggregation, and timestamps should be normalized to an explicit rule. The interval union example assumes discrete inclusive day labels; implementations using timestamp intervals must document their own end-boundary logic.
Audit and sensitivity design
A comprehensive check should reconcile earned car-years to policy records; verify mileage against sensor and odometer increments; flag negative or implausible distances; measure missing covered-mile fractions; and reconcile accident, vehicle, coverage and claim identifiers. Separate claim counts by at-fault, non-fault, collision, comprehensive, injury and liability outcomes when the records permit. Compare recent and older claims under compatible development and reporting conventions.
Then fit time-only and mileage-only exposure specifications alongside models incorporating both, with nonlinearity and relevant use covariates. Evaluate overdispersion, zero counts, repeated-policy dependence, cluster-level calibration, out-of-sample prediction and systematic differences across mileage bins. Analyze telematics selection separately, including a comparison with eligible but unenrolled insureds where credible data exist. Conduct sensitivity analysis under plausible missing-trip imputations and alternative driver attribution.
The actuarial data-quality standard frames the data review and disclosure issues, and prospective risk-cost guidance frames exposure-base selection and treatment of changes in historical records. Meeting a technical test or finding a valid DOI does not prove that a causal claim or policy-rate conclusion is scientifically correct; substantive evidence-to-claim judgment remains part of scholarly QA.
Specific literature limits
Ferreira and Minikel's large Massachusetts car-year analysis reflects their studied policy year, classes and coverage information. Verbelen, Antonio, and Claeskens's Belgian telematics analysis reflects a selected product and driving population. The Canadian longitudinal study applies a panel model whose within-driver estimand differs from a general cross-sectional effect. The first-at-fault-distance study uses a survival outcome that differs from frequency of all claims. The GPS-based pricing study combines exposure and driving behavior under its own design. The findings cannot be pooled as interchangeable estimates of one universal marginal risk per mile.
This research reviews accessible official webpages, source records and scholarly publication texts or abstracts through November 14, 2026. Original insurer microdata, external proprietary telematics databases, source-specific coefficient replications, and original national insured-mile measurements were not obtained. The mathematics and illustrative calculations are independently checkable; source-specific observational limitations are maintained. The monograph supports interpretation and research design rather than personal insurance quotes or underwriting advice.
Integrated Results and Open Research Questions
Combined findings
Insurance time and road movement measure different properties of motor exposure. The first derives from a contractually covered risk interval; the second from observed operation. The actuarial exposure-base standard supports choosing a denominator with a strong relation to cost and practical verifiability, while road travel monitoring and insurance exposure reporting illustrate the separate statistical systems that supply the observations.
A common numerator divided by vehicle-years and by covered miles yields two rates with different meanings. In the illustrative two-group comparison, the annual rate and per-mile rate reverse their ranking. This is a general algebraic possibility when annual distance varies; it does not establish a new causal or scientific doctrine. Because frequency, severity and cost are distinct, the appropriate reporting suite includes matched claims per covered year, claims per covered mile where observed, miles per covered year, and loss cost under a consistent claim convention.
Mileage is demonstrably informative in the cited actuarial literature, though the strength and shape of the relation depend on sample and model. The Massachusetts evidence retains class and territory, the Belgian telematics evidence examines observed travel context, and the longitudinal Canadian evidence emphasizes the value of comparing repeated observations of the same driver. These contributions make different empirical claims and are most informative when read together with their assumptions.
The risk unit must remain traceable through every transformation. A coverage day, mile, trip or claim requires an identity, timing rule, unit and data source. An in-force car may be stationary and nonetheless produce certain covered losses. A journey may take place during a policy lapse. A claim can belong to a different coverage or be reported long after the accident. Each case changes which denominator is defensible.
Questions for further empirical work
The most important next research design would link effective coverage intervals, verified mileage, driver attribution and fully developed claims for a carefully described insured cohort. Models could then compare duration and mileage as offsets, flexible mileage response, trip counts, driving hours and contextual road-use factors. Primary outcomes should include collision events, at-fault incidents, non-driving comprehensive losses, injury liabilities and ultimate loss costs, rather than assuming a single claim process.
A second investigation should estimate the extent of recording bias from missing trips and compare aggregate odometer measurements with more detailed telemetry. A third should distinguish voluntary selection into PAYD from subsequent behavioral change through a credible identification strategy. A fourth should test the transportability of published findings across passenger cars, commercial fleets, rural and urban settings, and mature and emerging insurance markets. A fifth should develop documented mappings between national road activity and covered insured mileage for compatible vehicle classes and dates.
These questions concern data construction, established statistical techniques, actuarial interpretation and governance. Their resolution requires original observation and independent validation. The present synthesis defines the relationships and the evidentiary requirements that make such work meaningful.
Conclusion
Motor vehicle years, mileage, driver-time, trips and operating hours are complementary exposure quantities. Earned vehicle-years measure contractually protected time. Covered mileage measures observed travel in the corresponding risk period. Claims per car-year and claims per million miles answer related but different questions, and the two can rank groups differently whenever miles per covered year differ. Their valid numerical transformation requires exactly matched claim populations and risk intervals.
The peer-reviewed literature establishes the relevance of driving distance and travel context while demonstrating the importance of risk class, observation design, selection and model specification. National traffic measures and insurance accounting measures remain distinct evidence systems. Careful alignment of their population, coverage, period and event definitions is essential to international comparison.
The resulting operational specification is to reconstruct the exposure ledger, validate measured and missing travel, link appropriate claims and costs, calculate transparent rates, check algebra and cohort composition, test model sensitivity and disclose data-processing and legal constraints. These findings synthesize existing actuarial and empirical knowledge. No proprietary theory, claim of scientific priority or independent peer review is asserted.
References
Actuarial Standards Board. (2005; revised 2011). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). https://www.actuarialstandardsboard.org/asops/risk-classification-practice-areas/
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality. Effective April 30, 2017. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2017). Actuarial Standard of Practice No. 53: Estimating Future Costs for Prospective Property/Casualty Risk Transfer and Risk Retention. Effective August 1, 2018. https://www.actuarialstandardsboard.org/asops/estimating-future-costs-prospective-propertycasualty-risk-transfer-risk-retention/
Ayuso, M., Guillén, M., & Nielsen, J. P. (2019). Improving automobile insurance ratemaking using telematics: Incorporating mileage and driver behaviour data. Transportation, 46, 735–752. https://doi.org/10.1007/s11116-018-9890-7
Ayuso, M., Guillén, M., & Pérez-Marín, A. M. (2016). Using GPS data to analyse the distance travelled to the first accident at fault in pay-as-you-drive insurance. Transportation Research Part C: Emerging Technologies, 68, 160–167. https://doi.org/10.1016/j.trc.2016.04.004
Boucher, J.-P., & Turcotte, R. (2020). A longitudinal analysis of the impact of distance driven on the probability of car accidents. Risks, 8(3), Article 91. https://doi.org/10.3390/risks8030091
Department for Transport. (2026, updated September 30). Road traffic estimates in Great Britain, 2025: Headline figures. UK Government. https://www.gov.uk/government/statistics/road-traffic-estimates-in-great-britain-2025/road-traffic-estimates-in-great-britain-2025-headline-figures
Department for Transport. (2026). Road traffic statistics information: Methods and quality. UK Government. https://www.gov.uk/guidance/road-traffic-statistics-information
European Data Protection Board. (2021, March 9). Guidelines 01/2020 on processing personal data in the context of connected vehicles and mobility related applications, Version 2.0. https://www.edpb.europa.eu/documents/guideline/guidelines-012020-on-processing-personal-data-in-the-context-of-connected_en
European Parliament and Council. (2016, as amended). Regulation (EU) 2016/679: General Data Protection Regulation, Articles 5 and 25. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2016/679/
Federal Highway Administration. (2026). National Household Travel Survey: Survey descriptions and source information. U.S. Department of Transportation. https://www.fhwa.dot.gov/Policyinformation/nhts.cfm
Federal Highway Administration. (2026). Traffic Volume Trends: Travel Monitoring. U.S. Department of Transportation. https://www.fhwa.dot.gov/policyinformation/travel_monitoring/tvt.cfm
Federal Highway Administration. (2026). Our Nation's Highways 2026: Travel. U.S. Department of Transportation. https://www.fhwa.dot.gov/policyinformation/pubs/our_nations_highways_2026/travel.cfm
Ferreira, J., Jr., & Minikel, E. (2012). Measuring per mile risk for pay-as-you-drive automobile insurance. Transportation Research Record, 2297(1), 97–103. https://doi.org/10.3141/2297-12
National Association of Insurance Commissioners. (2026, February 13). NAIC releases 2022/2023 Auto Insurance Database Report. https://content.naic.org/article/naic-releases-20222023-auto-insurance-database-report
National Association of Insurance Commissioners. (2026). Publications: Auto Insurance Database Report (AUT-PB 2022–2023). https://content.naic.org/publications
National Highway Traffic Safety Administration. (2025). NHTSA releases 2023 traffic deaths, 2024 estimates. U.S. Department of Transportation. https://www.nhtsa.gov/press-releases/nhtsa-estimates-39345-traffic-fatalities-2024
Verbelen, R., Antonio, K., & Claeskens, G. (2018). Unravelling the predictive power of telematics data in car insurance pricing. Journal of the Royal Statistical Society: Series C (Applied Statistics), 67(5), 1275–1304. https://doi.org/10.1111/rssc.12283
Werner, G., & Modlin, C. (2010). Basic Ratemaking (4th ed.). Casualty Actuarial Society. https://www.casact.org/abstract/basic-ratemaking