RESEARCH DOSSIER
Road Traffic Crash Exposure and Insurance Risk Denominators
Abstract. Road-traffic safety statistics and automobile insurance experience measure different events in partly overlapping populations. This monograph develops an international analytic synthesis of crash exposure, travel-based denominators, police-report ascertainment, insurance eligibility, claim frequency, causal comparisons, and reproducibility. It establishes precise observational units for crashes, involved vehicles, injuries, people, policies, and claims; explains how vehicle-kilometers, driver-years, vehicle-years, trips, encounters, and monetary exposures support different estimands; and derives statistical identification conditions for the transition from road harm to covered insurance loss. The evidence base combines WHO, ITF, FHWA, and NHTSA primary statistical definitions with published methodological research on accident-count overdispersion and observational safety evaluation. All numerical examples below are explicitly synthetic, except identified figures directly attributed to official sources. The resulting analysis specifies a defensible pathway from descriptive public road safety data to insurer risk models while retaining the selection, temporal, legal, and cross-jurisdictional limits of each source.
Keywords: road traffic crashes; insurance exposure; vehicle-kilometers traveled; policy-years; fatality rates; claims frequency; IRTAD; FARS; sampling bias; overdispersion; cross-country comparability; causal road safety evaluation; actuarial statistics.
Scientific Foundations and Observational Populations
Road-traffic exposure is the amount or frequency of opportunities for a specified road-user population to experience a specified event. A valid exposure denominator identifies who or what is at risk, for how long or how far, and under which observation rules. The event may be a police-reportable crash, a fatal crash, an injury, a death, a collision involving a particular insured vehicle, a compensable liability event, a submitted claim, or a payment. These events have different counting rules. A fatal crash may involve several deaths, a multi-vehicle collision several insured automobiles, and a single collision multiple claims for property damage, medical treatment, injury, defense, and other insured benefits. Their counts must remain distinct before statistical inference begins.
The WHO Global Status Report on Road Safety documents a worldwide estimate of approximately 1.19 million annual road deaths in its 2023 report. The estimate is important as a public-health indicator, but it is a global mortality measure, not a census of automobile insurance claims, insured vehicles, or covered accidents. It combines national reporting, modeling, and definitions; the study's numerator therefore cannot simply be divided by premiums collected or the population of a particular insurer to infer a private portfolio rate. A cause-of-death registration system, a police traffic report, and an insurer claim file each represent a different institutional observation mechanism.
The first methodological operation is defining the target population. For a national fatality-risk study, the population may encompass all people injured on public trafficways; the unit can be a person-death. For automobile insurance pricing, the population may encompass covered policy contracts or vehicles, and the relevant event may be one claim of a specific coverage type during an earned exposure period. For road-segment engineering, the entity may be a section of roadway over a measured interval, with motor-vehicle traffic flow and crash count. For evaluating exposure of cyclists or pedestrians, motor-vehicle distance alone may capture only a contributing background force while omitting the road-user's own exposure in person-trips or person-kilometers.
Official administrative systems make these boundaries observable. The US Fatality Analysis Reporting System covers fatal motor-vehicle traffic crashes across the 50 states, DC and Puerto Rico, under its defined fatality-within-30-days inclusion rule. Its documented data inputs include police crash reports, death certificates, roadway records, and other state records. The WHO discussion of the fatality definition identifies the relevance of a 30-day convention and the interpretive complications of connecting the injury event to later death. A change in registration rules can create an apparent fatality trend even when physical danger has not changed. A study therefore needs event dates, outcome ascertainment windows, inclusion rules, and the version of each source's definitions.
The IRTAD road safety database makes a comparable distinction between road safety outcomes and exposure. It includes national crash and injury indicators alongside vehicle-kilometers, vehicle fleets, driver licensing and population measures, under a documented common-format framework. Each is useful for a different question. A rate of fatal injuries per population describes a societal burden. A rate per vehicle-kilometer approximates mortality per traffic volume. A rate per registered vehicle combines driving intensity and conditional crash risk. A rate per licensed driver includes participation, driver activity, and legal eligibility. One denominator is suitable only after the estimand and target population are fixed.
A scientific dossier linking road risk to insurance must also distinguish liability, loss, and claims. A collision creates physical damage. Whether a person is liable depends on substantive law and facts; whether the event triggers an insurer's duty depends on contract terms, coverages, limits, exclusions, deductibles and jurisdictional compulsory insurance regimes. An insurer may make several payments for one collision or none. An uninsured road user can appear in police fatalities without belonging to any private auto exposure pool. The population of insurance-reported accidents is typically selected by coverage, claim incentives, reporting practices, and policy administration. The distinction governs every subsequent comparison.
The Architecture of Crash Exposure Denominators
Distance denominators are widely used because they approximate the opportunity for vehicle movement. The US FHWA national travel series reports approximately 3.3 trillion vehicle-miles traveled on public roads in 2024; its detailed annual entries include 3,309 billion miles. This statistic describes road use across vehicle classes and public-road systems. It does not imply that all journeys were insured under one set of private contracts. The FHWA driving data summary reports vehicles registered, licensed drivers and vehicle-miles as separate quantities. National fatality rates often use deaths per 100 million vehicle-miles, as explained in the FHWA safety performance tables.
Let K denote vehicle-kilometers traveled, D the number of defined crash events, and r = D/K the distance-specific event rate. Reporting 10^6 r gives crash events per million vehicle-kilometers. The equivalent vehicle-mile denominator requires converting both unit and rate consistently: one vehicle-mile equals approximately 1.609344 kilometers, so a rate per million miles is approximately 1.609344 times the rate per million kilometers when the numerator is fixed. A conversion of units is distinct from a change of exposure population. Readers must not compare numerically similar miles-based and kilometers-based rates without checking the scale.
A vehicle-year is an earned temporal measure of risk-bearing. A car covered for six months contributes approximately 0.5 vehicle-years under an appropriate daily or monthly earning convention. A contract covering two vehicles for six months can contribute one vehicle-year but only half a policy-year if the policy is the counting object. This arithmetic explains why policy-years, vehicle-years, and driver-years cannot be treated as interchangeable. An insurer's monthly earned exposure may be affected by policy cancellation, midterm substitution, fleet turnover, seasonal use and changing coverage. If distance traveled is unavailable, vehicle-years approximate time eligible for claims, not realized driving opportunities.
Trips count departures or completed journeys and can be useful for crash-on-trip questions. Trips vary dramatically in length, road type, time, occupancy, and conflict opportunities. Person-kilometers weight travelers individually and are appropriate for passenger or vulnerable road-user burden, whereas vehicle-kilometers weight vehicles. Bicycle or pedestrian risk assessment requires its own exposure measures. Traffic counts such as annual average daily traffic combined with road-segment length estimate road-segment vehicle-kilometers. This is a model-based exposure, and errors in the count, functional road class, observation period or segment length propagate into its rate.
The FHWA 2024 VM-1 table separates travel by road type and vehicle group, showing why an aggregate national distance denominator masks composition. A commercial freight vehicle has different annual mileage, operating hours, vehicle mass, and driving context from a passenger car. If a high-mileage low-rate group becomes a larger share of national traffic, the aggregate rate can fall even as each group's individual rate remains constant. That composition effect must be examined before inferring that roads or vehicles became safer.
Theoretical exposure can also be defined through conflict opportunities. Elvik's event-based analysis of road accident exposure proposes examining events that create an opportunity for conflict between road users or require corrective action. The study's concept is helpful when equal kilometers produce different numbers of intersections, close encounters, overtaking maneuvers, merging events or pedestrian crossings. A busy urban intersection and a low-interaction highway may each generate a kilometer of travel but different conflict intensities. Conflict opportunities, however, must themselves be defined and measured reproducibly. Telematics observations, video-based conflict indicators and trajectory data can refine distance proxies but add coverage and privacy limitations.
A useful exposure table must consequently record numerator, denominator, unit, period, geography, vehicle class, severity definition, insured-coverage status, and ascertainment mechanism for each estimated rate. A rate should be interpreted as the expected number of clearly specified events per standardized observation unit under the sampled conditions. Cross-group comparison further requires checking that the denominator represents the same level of opportunity and risk eligibility in each group.
Crash Ascertainment, Underreporting, and Insurance Claims
Observed crashes are a measurement process applied to underlying physical events. Let T represent the true number of defined events, p the probability of their inclusion in a given database, and O the observed count. Under a simplified independently thinned process with common p, E[O] = p E[T]. In practice, p depends on injury severity, property damage amount, reporting law, police attendance, road location, vehicle type, driver incentives and other variables. An insured claim may be filed only if the expected indemnity or another benefit exceeds reporting and deductible costs. Thus observed police data and claim data are differently selected subsets of physical crashes.
The US Crash Report Sampling System offers a probability sample of police-reported motor-vehicle crashes, including property-damage-only and injury crashes. The documented sampling frame is explicitly police reports. Nationally representative weighting can estimate police-reported crash totals but does not by itself recover all events never reported to police. The Crash Investigation Sampling System uses an additional, more restrictive selected tow-away passenger-vehicle frame and detailed investigation. Its cases should not be pooled with general police-reported cases as if the same inclusion rule and measurement depth applied.
Insurance claim records introduce a further measurement layer. An insurer may have exposure and a claims system for collision, comprehensive, bodily injury liability, personal injury protection, uninsured motorist, or other coverages. Each claim type has a separate trigger, insured interest, reporting horizon, defense and settlement process. One crash may produce multiple claim records of the same or different types. Catastrophic vehicle damage and bodily injuries may generate claims that mature at different speeds. A collision between vehicles insured by different companies can be recorded by both insurers; a collision involving a pedestrian may be visible to a liability insurer without appearing in first-party collision coverage.
If a researcher wishes to estimate crash incidence from insurance claims, a claim-to-accident linkage table is essential. The link should preserve accident identifier, date/time, vehicle identifiers, participating parties, coverage, opened and reopened claim IDs, loss location, subrogation, recovery, policy effective dates, and event cause. Claims with the same date are not necessarily the same accident, and a single accident can contain several insured vehicles. Privacy law and authorized purpose govern linkability. Aggregate claim totals are insufficient to reconstruct unique accidents unless the record linkage logic has been established and validated.
Consider a wholly synthetic two-source capture example. A police database contains 100 defined crashes; an insurer-matched file contains 120; 60 appear in both. The familiar independent-source estimator is (100 × 120)/60 = 200 total crashes, implying 40 events unseen by either source after accounting for the union of 160. The arithmetic is straightforward, but the estimate requires a stable closed event population, source matching, and a conditional independence assumption. Police and insurance reporting are often positively associated through accident severity, which can bias this simple calculation downward. Different insurers, duplicate events, reporting windows and late claims can break the assumptions. The example establishes why a measured overlap supplies evidence about ascertainment but does not automatically identify true accident incidence.
Severity creates another denominator problem. Crash counts and fatalities refer to different outcomes. For a defined reporting period, an estimated fatality rate per vehicle-kilometer is deaths/K; a fatal-crash rate is crashes containing at least one qualifying fatality/K. Since a crash can involve several deaths, these two rates differ. Public health studies may further distinguish 24-hour, seven-day, and thirty-day deaths. A claim paid years after the accident belongs to an accident-year numerator when analyzing accident occurrence; it belongs to a payment calendar year when analyzing cash flow. Mixing these axes can create spurious changes.
An empirical audit should compare police, hospital, mortality, registration, traffic and insurance administrative series, document linkage quality, and quantify sensitivity to underreporting. National sources often differ in available coverage. Where an independent validation source is absent, the responsible result is a clearly labeled police-reported or insurer-reported rate with an ascertainment limitation rather than an unqualified estimate of all crashes.
Frequency Models, Heterogeneity, and Worked Calculations
Crash frequency modeling begins with an explicit stochastic observation equation. For group i, let N_i denote the number of specified crashes, E_i a positive exposure measured in agreed units, and X_i covariates measured before the outcome. A Poisson log-link model can be written log E[N_i | X_i, E_i] = log E_i + α + βX_i, where log E_i is the exposure offset. This structure imposes proportional scaling of expected count with exposure at a fixed covariate state. Its scientific meaning depends on the chosen unit: vehicle-kilometers, earned vehicle-years, intersection entering vehicles or another documented exposure.
The proportional exposure assumption should be tested. More driving may take place on relatively safe long-distance roads, while marginal trips may differ in time, congestion or route. Cumulative experience can change behavior. Event-based exposure may vary superlinearly with traffic density at intersections. For these reasons log exposure is a standard offset only when the model's underlying time-at-risk or distance-at-risk mechanism justifies proportionality. When traffic flow itself changes conflict intensity, including flow as a modeled predictor in addition to an exposure term may be more appropriate.
Poisson variance equals its conditional mean. Crash data often display overdispersion because unmeasured road, driver, vehicle, weather and exposure heterogeneity produces excess variability. Hauer's analysis of overdispersion on road sections shows that assumptions about dispersion and section length matter for regression estimation and empirical Bayes evaluation. A negative binomial, Poisson–lognormal, hierarchical count or mixture model may be useful, but model choice requires distributional checks rather than mechanical replacement. An inflated number of zero observations can also arise from heterogeneity in true exposure, road-network sampling or event definition.
A reproducible synthetic example demonstrates the danger of raw-count comparison. Suppose a road network records 150 eligible crashes during 100 million vehicle-kilometers in an earlier year and 168 during 140 million vehicle-kilometers in a later year. Counts increase by 12%, since 168/150 = 1.12. The exposure-normalized rate declines from 1.50 to 1.20 crashes per million vehicle-kilometers, a 20% decrease. This is compatible with growing traffic and falling risk per traveled kilometer. It does not independently identify a safety intervention effect: route mix, vehicle composition, severity reporting and road conditions might have changed at the same time.
A separate synthetic insurance calculation distinguishes counts from insurer risk. Consider 10,000 earned vehicle-years, with 800 distinct covered collision claims. The annual covered-claim frequency is 0.08 per vehicle-year. If the same portfolio generated 120 million vehicle-kilometers, its distance-based covered-claim frequency would be 6.6667 claims per million vehicle-kilometers. These rates answer different questions. If the next year saw identical 800 claims but 160 million kilometers and the same 10,000 earned vehicle-years, the vehicle-year rate would be unchanged while the distance rate would fall to 5.0 per million kilometers. Different conclusions follow because driving intensity per covered vehicle has changed, not because the arithmetic is inconsistent.
The expected insured-loss cost per earned vehicle-year can be expressed as λ_C × E[S_C | covered claim], where λ_C is the annual covered-claim frequency and S_C the appropriately defined ultimate claim severity, net or gross of recoveries according to the research question. A multi-coverage model may require a sum across coverage types, with careful treatment of correlated claims arising from a shared accident. Where accident occurrence is modeled first, the decomposition extends to crash incidence × probability of relevant coverage and reporting × conditional indemnity, adjusted for policy limits, deductibles, legal liability, claims management and exposure mix. This expression is analytical accounting of conditional probabilities and costs, not an assertion that every component is independently identifiable.
Source measurement error propagates through rates. If observed events are N and estimated exposure is Ê, a small multiplicative error Ê/E = 1 + δ approximately shifts the rate N/Ê relative to N/E by −δ for small δ. Underreporting creates a separate numerator factor. When both biases vary by group, their difference can produce an apparently significant risk contrast that has no equivalent in true crash hazard. Sensitivity analysis should therefore report plausible exposure and reporting intervals rather than only a model-fitting confidence interval.
A defensible actuarial forecast also accounts for delay. Accident date, first notification, closing date, final settlement, salvage and subrogation recognition may be separated by substantial time. The latest accident-year claims count is often incomplete. A crash-frequency comparison should use mature claims, adjust for reporting lag, or explicitly model delayed reporting. The forecast target can be notified claims by twelve months, ultimately reported claims, ultimately settled events, or paid-claims count, with each given a different definition.
International Comparison and Causal Identification
Cross-country comparisons require more than aligning units. The ITF Road Safety Annual Report 2025 presents comparative indicators for IRTAD participating jurisdictions, while its publication date is February 2026 despite the year in its title. Historical series may differ in completeness, legal reporting thresholds, vehicle categories and traffic observation systems. Inference must therefore identify each country's numerator system and denominator coverage and examine whether changes in one country's reporting practices coincide with the observed trend.
A population-based road fatality rate describes burden in a demographic group. A vehicle-distance rate describes fatalities relative to motor-vehicle movement. An insurer's covered-claim rate describes contracted loss frequency in a selected portfolio. Population rates may rise when motorization expands even if distance-based risk declines. An insurer's liability claims may rise because coverage penetration increases while the underlying fatal crash rate falls. These apparent disagreements are compatible once population, exposure and claim triggers are distinguished.
Direct standardization can separate exposure composition from within-class differences. Suppose countries A and B have rates r_{j,A} and r_{j,B} for road classes j, and choose a common reference set of traffic shares w_j summing to one. The standardized rates are R_A = Σ_j w_j r_{j,A} and R_B = Σ_j w_j r_{j,B}. The standard is explicitly chosen; the rates do not represent each country's observed aggregate unless the reference weights coincide with its observed shares. This is particularly important when comparing urban and rural networks, motorcycles and passenger cars, novice and experienced drivers, or daytime and night travel.
Insurance portfolios require analogous standardization for earned vehicle-years, annual mileage, vehicle age, use class, location, coverage, deductible, historical claims and driver composition. Changes in policy administration or statutory limits may change claim eligibility independently of crash incidence. A standardized road safety rate can support contextual risk analysis, but only a validated crosswalk between road occurrence and insurer insured exposure enables conversion to pricing or reserving forecasts.
Causal evaluation is more demanding. A road safety measure may be introduced on high-crash roads. Simple comparison of post-intervention counts with unusually high pre-intervention counts can show a decline caused partly by regression to the mean. Hauer's observational before–after methodology emphasizes traffic changes, comparison groups and empirical Bayes estimation. The FHWA evaluation protocols discuss empirical Bayes approaches and the importance of accounting for regression to the mean. An exposure rate adjustment alone does not solve treatment selection or unmeasured time trends.
A credible intervention study specifies treatment onset, affected road sections, comparison-network selection, anticipation, traffic diversion, and possible spillovers. It asks whether vehicles shifted to nearby roads, whether road user composition changed, whether detection and reporting increased, and whether policies changed at the same time. Effects can be conditional on vehicle class, severity, weather, light and road geometry. If enforcement changes travel demand, vehicle-kilometers may be a mediator rather than an exogenous offset; conditioning on post-treatment exposure can estimate a different causal quantity from the total effect on crashes.
The distinction between national public safety and private insurance loss is particularly important in cross-border comparison. Compulsory third-party insurance regimes, benefit schedules, liability rules, compensation systems, and claims reporting incentives differ. A lower police-recorded injury-crash rate may coexist with higher average insurance loss because medical prices, litigation costs, vehicle repair costs, coverage limits or injury compensation rules differ. Cross-country transport data should therefore not be treated as actuarial rate tables without an insurer-specific legal and financial observation model.
Validation, Source Provenance, and Integrated Findings
Validation begins with a reproducible dataset dictionary. For each analytical table record the collecting organization, publication version, access date, event inclusion threshold, observation start and end, severity threshold, geographic coverage, vehicle categories, original unit, transformations, and reconciliation to public totals. The FHWA historical FI-200 series shows several denominators for the same national fatality numerator; its rates should be recomputed after converting the source's published units and preserving its historical definitions. For detailed travel reconciliation, the analyst can use the FHWA class-specific tables.
Cross-source reconciliation should state whether the numerator is person fatalities, fatal crashes, police-reported property-damage crashes, all police-reported crashes, unique insured accidents, opened claims, reported claims or paid claims. Every source-to-claim mapping must distinguish observed original records from imputed or model-estimated outcomes. The 2025 traffic-death estimate announcement is an example of explicitly designated preliminary estimates; a manuscript must retain that distinction if citing the reported figures. Revisions to annual datasets require versioned records and an explanation of whether older rates have been recalculated.
Sampling validation should examine inclusion probabilities, frame coverage, unit duplication, event linkage quality and nonresponse. A probability sample of police reports may allow design-based inference for the police-reported universe; it does not guarantee generalization to unreported crashes. Administrative insurer datasets may include nearly all reported claims for one organization but have severe external validity limits across policyholders, uninsured motorists, jurisdictions and products. An encounter-based telematics data source may show detailed driving events while selectively representing motorists who have opted into monitoring. A validation sample covering only large, paid claims can be a poor test of small-claim frequencies.
Temporal validation requires claims-development alignment and forward-chaining. When claims records are available only after settlement, using ultimate outcome information as a predictor for accident-time risk creates information leakage. Historical backtests should construct the same information set that would have existed when a price, reserve, risk intervention or underwriting decision was actually made. Calendar-year changes in road traffic, fleet technology, fuel prices, reporting procedures and economic activity should be studied as distinct channels. Regional holdouts help evaluate transportability, whereas random record splitting can leak road-segment or driver identity between training and test partitions.
Model checking should include event-level aggregation, incident deduplication, exposure reconciliation, rate calibration, extreme-cell variance, uncertainty intervals, spatial correlation, underreporting sensitivity, and out-of-sample prediction. Stratification by road class, vehicle type, user vulnerability and urbanization helps reveal where aggregate metrics hide localized hazards. Tail insurance severity should be evaluated separately from frequency because a low crash-frequency environment can still generate costly injuries or catastrophic multi-vehicle losses.
The original analytical synthesis of this monograph is the consistent translation of one observed road event across four domains: physical occurrence, administrative ascertainment, insured eligibility, and financial claims. Each transition has a measurable population and a conditional selection mechanism. The transitions can be documented with public data, insurance administrative records, and insurer contracts when available. They cannot be inferred as identities from a single public road safety rate. This synthesis follows existing actuarial and transport-statistical theory and is intended as a practical scholarly reconciliation of established concepts.
An evidence-based international research program should test whether distance-normalized crash rates improve insurer claim prediction after vehicle-year exposure and coverage attributes are included; compare encounter-based models with conventional travel offsets; evaluate ascertainment variation using matched police and insurer records; and quantify uncertainty due to reporting delay and national coding standards. Identification assumptions should be declared before calculating a comparative figure. The scientific objective is a rate with a clear numerator, a compatible denominator, a known observation mechanism and an explicit decision purpose.
From crash occurrences to insured event frequencies
The inferential bridge from highway safety to an insurance portfolio can be represented with a sequence of observable and latent indicators. Let C be a physical road-traffic collision under a specified event definition; let I be involvement of an insured vehicle or person at the collision date; let L indicate that the occurrence produces liability or first-party loss recognized by the applicable contract; let R denote a submitted claim; and let P denote a recorded payment. These indicators describe separate decisions and institutions. A third-party claimant can be injured without holding the policy that ultimately responds; a claim can be reported and closed without payment; and a collision can cause multiple claims under different coverage types. The relationship among these events is conditioned on policy terms, peril, fault rules, insurer reporting practice, and the jurisdiction's procedural law.
A useful descriptive identity for a particular fixed population is
E[paid claims per unit exposure] = E[covered claim events per unit exposure] × Pr(payment | covered claim event)
when the numerator counts one payment decision per covered event under a common definition. If each claim may involve several indemnity payments, the last term must instead be the expected number of payments conditional on the covered claim event. The identity illustrates how a rate depends on its measurement object; it is not an assumption that independent stages or homogeneous populations exist. For an aggregate paid-loss amount, an alternative frequency-severity decomposition conditions on both the number of paid claims and their payment amount, with due adjustment for recoveries, reopening, litigation, deductibles, limits, and inflation.
The NHTSA Model Minimum Uniform Crash Criteria standardizes crash-level, vehicle-level, person-level, and driver-level descriptors for police-record systems. Its sixth edition, updated in 2024, aims to improve interoperability among U.S. traffic records. A standardized police record still does not identify the insurer liable for each participant, contractual benefit eligibility, loss-adjustment expense, or claim-reserve history. The central challenge of an insurance linkage study is therefore an explicit relational crosswalk, never a simple substitution of one database for another.
For an insurer carrying comprehensive and third-party liability policies, a proposed linkage table should distinguish the event's stable crash identifier, road-segment or geospatial reference, local observation date and time, vehicles involved, role of each insured unit, policy-effective and policy-expiration instants, applicable coverages, claim numbers, claim-reporting lag, and closure and payment transactions. Vehicle identity can be time-dependent when a fleet, rental, lease, transfer, or policy substitution changes the covered unit. A crash identifier can link many distinct claim records; one vehicle can sustain separate collision events; and an administrative correction can replace an earlier record without adding a new event. The proper unique keys and cardinalities must be documented before any multiplication of counts or calculated denominators.
Data protection is an additional scientific boundary. A source's use of a deidentified public case series does not justify claiming that insurer policy identifiers can be reconstructed from it. FARS documentation specifies that personally identifying information is excluded from its released public data, whereas insurer files typically retain identity and contract records for claims administration. Probabilistic linkage can use harmonized time, vehicle characteristics, location bins and event severity with error models, but every such linkage creates the possibility of false positives, false negatives, and selective omission. Its data-governance basis and lawful access arrangement must be reviewed independently of its predictive performance.
Source frames, selection mechanisms, and external validity
A road-safety database is a product of a reporting frame. FARS fatal-crash census rules require a motor vehicle traveling on a qualifying trafficway and a death within 30 days; the Crash Report Sampling System samples a much wider set of police-reported crashes using a probability design. The Crash Investigation Sampling System selects eligible crashes involving at least one towed passenger vehicle and reconstructs vehicle and occupant damage in greater detail. A cost predictor trained on such a tow-away frame will preferentially learn a particular severity and vehicle mix, even if an insurer's collision portfolio contains many low-severity incidents absent from that sampling frame.
The selection problem has several forms. A collision may be missing because no official report was filed; a valid police report may fall outside the survey frame; an insurer might lack the exposure or policy that would lead to a claim; a submitted claim might be denied or closed without payment; and a data analyst may retain only complete records. These filters select on mechanisms that are themselves related to loss severity, location, road-user demographics, or insured status. In causal language, conditioning on the reporting or claim indicator can introduce collider bias if both exposure characteristics and severity affect inclusion. The distribution of covariates in a selected file therefore need not match the distribution of risk in the intended underwriting population.
Suppose, only as a synthetic illustration, that a study captures 80 of 100 severe crashes and 40 of 400 minor crashes. Its severe-crash proportion is 80/120 = 66.7%, although severe crashes make up only 100/500 = 20% of the defined event population. An unweighted average cost based on the captured cases will generally overstate the average cost of all crashes if serious crashes are more expensive. If the two inclusion probabilities, 0.8 and 0.1, were correctly known and selection were otherwise ignorable within severity categories, inverse-probability weights of 1.25 and 10 would reconstruct 100 and 400 events, respectively. In practice, inclusion probabilities are rarely known for completely unrecorded crashes, and severity may be measured only after selection. That is an identification limit that a predictive algorithm cannot remove.
A legally valid insurance denominator is also selected. A premium measure is a financial amount reflecting rate, coverage, risk classification, taxes, and transaction timing; it is not directly an amount of road travel. Earned policy-years measure contract duration, but under a driver-specific policy, multiple vehicle drivers may have very different traveled distances and exposures. Fleet vehicle-days, registered vehicle counts, licensed drivers, passenger-kilometers and road-network travel count distinct populations. When national road measures are used in a portfolio model, their role should be declared as contextual covariates or calibration constraints rather than silently treated as insured exposure.
Frequency uncertainty and auditable quantitative contrasts
A statistical audit should distinguish the precision of the estimated rate from the quality of its numerator and denominator. Under a Poisson count model with fixed exposure, the variance of the event count equals its expectation and an approximate standard error for a log rate ratio comparing two independent counts is sqrt(1/N_1 + 1/N_0). Applying this approximation to the earlier wholly synthetic example of 150 crashes in 100 million kilometers and 168 crashes in 140 million kilometers gives a later-to-earlier rate ratio of 0.8 and a log-scale standard error of approximately 0.1122. The corresponding approximate 95% interval for the rate ratio is about 0.64–1.00. Rounding explains the endpoint; the method presumes independent Poisson events and treats both exposures as known. Overdispersion, correlated road segments, reporting changes, measurement error or shared vehicles will typically require a different variance estimator.
The contrast becomes less secure when the exposure data are revised. In one official national series, the FHWA 2026 travel account reports 3,309 billion vehicle-miles for 2024. In a separate publication, NHTSA's 2024 crash overview reports a 2024 FARS fatality count of 39,254 and a fatality rate of 1.19 per 100 million vehicle-miles, while its table uses 3,294,031 million vehicle-miles. These related official records should be retained with their publication date, revisions, statistical series, and units rather than presumed interchangeable. In particular, the 2024 FARS fatality number is a count of persons, not an annual collision-event count or insurer claim number; the denominator is nationwide estimated travel, not insured travel.
The distinction between preliminary estimates and finalized counts matters in time series. An early release may use statistical projections from incomplete reporting; a subsequent annual FARS file may revise counts. The 2025 early estimates and 2024 annual release explicitly separates an estimated 2025 fatality number from the published 2024 annual count. Modeling the earliest release of each year produces a different information set from retroactively using the latest revised historical observations. A legitimate forecast backtest must state whether information is frozen at the forecast origin or revised data are allowed and, if allowed, why.
Absolute differences and relative differences should also remain visible. In the synthetic two-year example, the reported count rises by 18, whereas the rate falls by 0.30 events per million vehicle-kilometers. The rate ratio of 0.8 indicates a 20% relative decline in distance-normalized incidence. Each claim is arithmetically consistent with the same data, and each expresses a different denominator and counterfactual interpretation. A regulatory or underwriting claim that road users became safer requires further causal evidence on who traveled, where, under which conditions, and with what reporting intensity.
Exposure composition and transportability
Cross-jurisdictional standardization requires the analyst to choose a coherent target population. A national rate is a mixture over mode of travel, region, driving time, road network, household and commercial vehicle classes, driver age, injury ascertainment and law. The IRTAD Road Safety Database documents a common set of fatality, injury, road-class and exposure series collected from participating national providers, while the Road Safety Annual Report 2025 presents comparative results from its stated reporting group. The fact that national providers use harmonized indicator definitions does not mean that all countries measure traffic volume, injury severity, reporting completeness, or insurance coverage with identical quality and methods.
Direct standardization can be presented with a transparent synthetic example. Suppose jurisdiction A has 20% of its travel in a high-risk road class with a rate of 4.0 and 80% in a lower-risk class with a rate of 1.0, both measured as events per million vehicle-kilometers. Its crude rate is 1.6. Jurisdiction B has 80% of its travel in the same high-risk class with a rate of 3.0 and 20% in the lower-risk class with a rate of 0.8. Its crude rate is 2.56. Although B has a lower class-specific rate in both classes, its aggregate is higher because the travel composition differs. Under a reference mixture of 50% high-risk and 50% lower-risk road travel, A has a standardized rate of 2.5 and B has a standardized rate of 1.9. This is a mixture-composition reversal, not evidence of an actual named country's road safety.
The example illustrates the basic estimand question: which mixture is relevant to the insurance portfolio? If a carrier underwrites mostly commercial highway traffic, a population-weighted national average may have less predictive relevance than a portfolio-weighted measure. If a national regulator evaluates overall safety, choosing one insurer's exposure distribution as reference would be inappropriate. A rigorous study reports both the unstandardized rates and the target weights, checks overlap or positivity among classes, and describes sensitivity to plausible alternative weight systems. It also recognizes that broad categories such as urban and rural can conceal materially different roads, vehicle technologies and driving conditions.
Comparison of temporal cohorts requires a common calendar convention and a known observation lag. Insurance policy earning may occur daily, crash occurrence may be logged immediately, a police injury classification may later change, and a claim can be reported after policy expiration or even following prolonged litigation. The loss occurrence date, claim notification date, administrative extraction date and accounting valuation date must be retained separately. Under different time windows, a calendar-year fall in claim counts can reflect delayed reporting rather than a change in collision risk. Conversely, improvements in digital reporting can increase recorded frequency while actual collisions remain stable.
Causal design, practical validation, and limits of inference
Observational evaluations of speed limits, traffic enforcement, roadway treatments and safety technology face endogeneity. Authorities may deploy an intervention to locations where recent crashes were unusually high; subsequent reduction may partly reflect regression toward a longer-term mean. Hauer's observational before–after research and the FHWA evaluation protocols emphasize the importance of study design, comparison units and model assumptions. An insurer using such estimates should distinguish a documented road-safety modification from the separate contractual question of whether, and at what lag, the modification affects covered claims.
A conceptual design compares treated and untreated locations over before and after periods while measuring travel and exposure characteristics in every cell. Its identifying comparison can be expressed as the change in treated exposure-normalized rate minus the change in control exposure-normalized rate. This difference-in-differences-style estimand measures the causal intervention effect only under credible assumptions on untreated trends, stable measurement and the absence of differential concurrent interventions or spillovers. If drivers respond to enforcement by shifting traffic to neighboring roads, an untreated adjacent road can become an affected comparison and obscure the effect. If an intervention changes driving distance or route selection, exposure is also an outcome, and treating it as a fixed control may distort the total effect on insured losses.
The validation plan for a road-to-insurance model has three distinct levels. Measurement validation checks the event identity, reporting and linkage sensitivity, route/vehicle coverage, time origin, exposure units, data revisions, duplicate claims, and changing legal thresholds. Statistical validation checks calibration and distributional fit by severity, geography, year, vehicle type and relevant protected or underwriting classes; it evaluates uncertainty under heterogeneity and tests performance in later time windows. Institutional validation checks whether the insurance product covers the measured event, whether policy tenure and rating classification are correctly interpreted, and whether a published risk metric is used within the permitted legal and contractual framework. Success at one level does not establish success at the other two.
Useful robustness exercises include alternative reporting windows, different exposure denominators, severity-stratified models, late-claim development factors, comparison of reported versus eventually revised road statistics, case-definition sensitivity, exclusion of anomalous periods, and validation on regions not used for estimation. Analysts should retain the full provenance of public and licensed data, including collection agency, dataset version, geographic resolution, extraction date, and transformation code. A machine-learning model achieving high within-sample discrimination while failing when the data source or reporting rule changes is unsuitable for an unsupported forward-looking actuarial conclusion.
The scientific conclusion is therefore conditional and operational: road-safety data can improve insurance loss analysis when the event, exposure, selection mechanism, claims relation, and target population are separately defined and validated. A defensible international synthesis can compare statistical and institutional mechanisms without asserting that a national crash rate directly determines a private insurer's loss ratio. This boundary is central to the appropriate use of road exposure in actuarial pricing, claims forecasting, loss prevention and public-risk evaluation.
Methodological disputes and empirical reporting evidence
There is a substantive statistical debate about the appropriate distribution of rare crash counts. Lord, Washington, and Ivan (2005) compare Poisson, Poisson-gamma and zero-inflated approaches while emphasizing that statistical fit alone is insufficient to determine the correct accident-generating mechanism. In a road safety file, a road segment with zero recorded collisions may simply have little exposure, an unusually short observation period, or substantial reporting omissions. A structural extra-zero component should represent a defensible latent process; fitting it to every excess zero can impose unsupported conclusions on the crash population. A Poisson-gamma mixture can represent unobserved heterogeneity without declaring a separate population permanently incapable of collisions.
The broader methodological review by Lord and Mannering (2010) identifies challenges involving low mean counts, overdispersion and underdispersion, omitted variables, variable functional forms, and temporal heterogeneity. These are not interchangeable statistical nuisances. Small counts create finite-sample uncertainty, incomplete exposure or covariates create identification risk, and model misspecification can bias both forecasts and intervention estimates. An insurer can improve conditional calibration by using more flexible models while simultaneously degrading causal interpretation if variable selection draws upon post-crash treatment, report timing or settlement outcomes. Reporting prediction and intervention impact as distinct scientific targets is therefore essential.
Independent evidence establishes that crash underreporting is heterogeneous rather than a single national adjustment coefficient. In their study of linked Danish police and hospital records, Janstrup and coauthors (2016) analyze crashes in Funen during 2003–2007 and report marked differences in police capture by road-user category and injury severity. This finding supports severity- and mode-specific sensitivity analysis; it does not imply that their observed rates apply universally to other periods, police systems or private motor-insurance portfolios. Hospital admission also represents an injury population, whereas insurer collision claims include many property-only losses and a different set of economically significant outcomes.
A disciplined quantitative analysis should therefore compare a minimum of four estimands: police-reported crashes per roadway-travel exposure; estimated underlying crashes per travel exposure under an explicit ascertainment model; insurer covered-claim events per earned insured vehicle-year; and paid losses per earned insured vehicle-year or per relevant financial exposure. The first is a descriptive administrative rate, the second is a model-dependent estimate, the third expresses insured contract experience, and the fourth introduces payment severity and financial development. Conversion from one estimand to another requires empirically identified selection, coverage and claims processes. Without them, a national safety statistic is contextual evidence rather than a substitute underwriting estimate.
Independent computational checks of the illustrative examples
The numerical examples are independently reproducible from the defined hypothetical inputs. For the two-year crash comparison: 150 / 100 = 1.5 and 168 / 140 = 1.2 crashes per million vehicle-kilometers, with 168 / 150 = 1.12 for the raw-count ratio and 1.2 / 1.5 = 0.8 for the rate ratio. The standard error of the log rate ratio, assuming independent Poisson counts, equals the square root of 1/150 + 1/168, approximately 0.1122. The conventional unadjusted 95% log-rate interval yields an approximate rate-ratio interval of 0.64–1.00. This interval should not be transferred to real data with clustered crash records or uncertain exposure without corresponding variance adjustments.
For the insurance exposure example: 800 covered claims divided by 10,000 vehicle-years produces 0.08 claims per vehicle-year. The same 800 claims divided by 120 million vehicle-kilometers gives 6.6667 claims per million vehicle-kilometers, compared with 5.0 under 160 million kilometers. The two measurements answer distinct questions because driving distance per vehicle-year differs. For the synthetic reporting illustration, inverse inclusion probabilities reconstruct 80 / 0.8 = 100 severe and 40 / 0.1 = 400 minor events; observed severity share is 80 / 120, while the reconstructed share is 100 / 500. The exercise depends on known reporting probabilities, an assumption clearly separated from ordinary unobserved collision datasets.
Finally, standardizing the synthetic two-jurisdiction example to equal high- and low-risk travel shares yields A = 0.5(4.0) + 0.5(1.0) = 2.5 and B = 0.5(3.0) + 0.5(0.8) = 1.9. With their actual travel mixes, crude rates remain A = 0.2(4.0) + 0.8(1.0) = 1.6 and B = 0.8(3.0) + 0.2(0.8) = 2.56. The resulting aggregate reversal illustrates compositional confounding and the necessity of prespecifying the reference exposure mix. None of these calculated examples is represented as an estimate from insurer records or a newly collected national road dataset.
Conclusion
Road traffic crash frequency is defined by the event-counting rule and the exposure mechanism together. Vehicle-kilometers, trips, driver-years, vehicle-years, population and road-conflict events measure different opportunities for harm and cannot be exchanged without an empirical or analytical bridge. Police records, fatality systems, specialized crash investigations and insurance claims apply different ascertainment and eligibility rules; each can serve valid research purposes within its defined population. A defensible insurer-facing analysis first reconstructs unique crash occurrence and exposure, then models reporting, contractual coverage and conditional claim severity under the relevant policy and legal environment.
The methodological synthesis identifies three practical priorities: maintain source-level unit and version provenance, model numerator and denominator error separately, and validate claims forecasts in genuine forward time. The official WHO, ITF, NHTSA and FHWA series provide a substantial comparative foundation, while published research on overdispersion, encounter exposure and before–after evaluation establishes important methodological constraints. Further actuarial inference requires portfolio-specific evidence and explicit mapping from physical events to covered claims. These conditions support internationally comparable research without attributing insurer-specific outcomes to public statistics.
References
World Health Organization. (2023). Global Status Report on Road Safety 2023. https://www.who.int/publications/i/item/9789240086517
World Health Organization. (2025). Definition of a Road Traffic Death by WHO and the Health Sector. https://cdn.who.int/media/docs/default-source/documents/health-topics/road-traffic-injuries/definition-of-a-road-traffic-death-by-who-and-the-health-sector.pdf
International Transport Forum. (n.d.). IRTAD Road Safety Database. https://www.itf-oecd.org/irtad-road-safety-database/
International Transport Forum. (2026). Road Safety Annual Report 2025. https://www.itf-oecd.org/road-safety-annual-report-2026
National Highway Traffic Safety Administration. (n.d.). Fatality Analysis Reporting System. https://www.nhtsa.gov/crash-data-systems/fatality-analysis-reporting-system
National Highway Traffic Safety Administration. (n.d.). Crash Report Sampling System. https://www.nhtsa.gov/crash-data-systems/crash-report-sampling-system
National Highway Traffic Safety Administration. (n.d.). Crash Investigation Sampling System. https://www.nhtsa.gov/crash-data-systems/crash-investigation-sampling-system
Federal Highway Administration. (2026). Our Nation's Highways 2026: Driving Data. https://www.fhwa.dot.gov/policyinformation/pubs/our_nations_highways_2026/driving_data.cfm
Federal Highway Administration. (2026). Our Nation's Highways 2026: Travel. https://www.fhwa.dot.gov/policyinformation/pubs/our_nations_highways_2026/travel.cfm
Federal Highway Administration. (2026). Highway Statistics 2024: Table VM-1. https://www.fhwa.dot.gov/policyinformation/statistics/2024/vm1.cfm
Federal Highway Administration. (2026). Highway Statistics 2024: Table FI-200. https://www.fhwa.dot.gov/policyinformation/statistics/2024/fi200.cfm
Federal Highway Administration. (2026). Our Nation's Highways 2026: Safety, Condition, and Performance. https://www.fhwa.dot.gov/policyinformation/pubs/our_nations_highways_2026/safety_condition_performance.cfm
Elvik, R.. (2015). Some Implications of an Event-Based Definition of Exposure to the Risk of Road Accident. https://pubmed.ncbi.nlm.nih.gov/25560900/
Hauer, E. (2001). Overdispersion in Modelling Accidents on Road Sections and in Empirical Bayes Estimation. https://pubmed.ncbi.nlm.nih.gov/11579982/
Hauer, E. (1997). Observational Before–After Studies in Road Safety: Estimating the Effect of Highway and Traffic Engineering Measures on Road Safety. https://shop.elsevier.com/books/observational-before-after-studies-in-road-safety/hauer/978-0-08-043053-9
Federal Highway Administration. (n.d.). Recommended Protocols for Developing and Reporting Crash Modification Factors. https://cmfclearinghouse.fhwa.dot.gov/collateral/CMF_Protocols.pdf
National Highway Traffic Safety Administration. (2026). 2025 Traffic Death Estimates and 2024 FARS. https://www.nhtsa.gov/press-releases/traffic-deaths-2025-early-estimates-2024-annual
National Highway Traffic Safety Administration. (2024). Model Minimum Uniform Crash Criteria, Sixth Edition. https://www.nhtsa.gov/traffic-records/model-minimum-uniform-crash-criteria
National Highway Traffic Safety Administration. (2026). Overview of Motor Vehicle Traffic Crashes in 2024 (Traffic Safety Facts Research Note, DOT HS 813 791). https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/813791
Janstrup, K. H., Kaplan, S., Hels, T., Lauritsen, J., & Prato, C. G. (2016). Understanding traffic crash under-reporting: Linking police and medical records to individual and crash characteristics. Traffic Injury Prevention, 17(6), 580–584. https://doi.org/10.1080/15389588.2015.1128533
Lord, D., Washington, S. P., & Ivan, J. N. (2005). Poisson, Poisson-gamma and zero-inflated regression models of motor vehicle crashes: Balancing statistical fit and theory. Accident Analysis & Prevention, 37(1), 35–46. https://doi.org/10.1016/j.aap.2004.02.004
Lord, D., & Mannering, F. (2010). The statistical analysis of crash-frequency data: A review and assessment of methodological alternatives. Transportation Research Part A: Policy and Practice, 44(5), 291–305. https://doi.org/10.1016/j.tra.2010.02.001