3. Crash test data sources
Injury risk modeling depends on data that spans two very different worlds: controlled crash tests conducted under repeatable laboratory conditions, and real-world crashes that occur with unpredictable configurations, occupants, and severities. A robust model usually needs both. This section maps the major data sources available to a safety analyst, explains what each source actually contains, and shows how to select and combine them for a specific modeling task.
Regulatory Compliance Data: FMVSS
FMVSS (Federal Motor Vehicle Safety Standards) testing is the baseline for legal vehicle certification in the United States. The most relevant standards for injury modeling are FMVSS 208 (occupant crash protection, including frontal rigid barrier and offset tests), FMVSS 214 (side impact, including the moving deformable barrier and pole tests), and FMVSS 301 (fuel system integrity). These tests use standardized dummies — Hybrid III in frontal tests, ES-2re or SID-IIs in side tests — and produce time-history channels such as head acceleration, chest deflection, femur load, and pelvis acceleration.
The key characteristic of FMVSS data is regulatory pass/fail structure. Each test has a defined injury criterion threshold, for example HIC 15 ≤ 700 for the driver in a 56 km/h frontal rigid barrier test, or chest deflection limits in side impact. Because the test matrix is fixed and the dummy fleet is standardized, FMVSS data is excellent for building baseline injury criterion distributions, but it covers a narrow band of crash severities and a single impact configuration per test type.
Consumer Rating Data: Euro NCAP and IIHS
Euro NCAP and IIHS extend beyond regulatory minimums into consumer-facing star ratings and awards. Euro NCAP's frontal offset test (50 km/h, 50% overlap against a deformable barrier), mobile progressive deformable barrier (MPDB) test, and side pole test generate detailed dummy response data plus structural deformation measurements. IIHS is known for its 40% overlap moderate overlap frontal test, the 25% overlap small overlap frontal test, side impact with a 1,500 kg moving deformable barrier, and roof strength and head restraint evaluations.
These programs are valuable for injury modeling because they push vehicles into more severe and more varied loading conditions than FMVSS. The small overlap test, for example, produces occupant compartment intrusion patterns and dummy responses that rarely appear in regulatory frontal tests. However, the data is still test-track data: the barrier is rigid or deformable in a known way, the impact speed is fixed, and the occupant is a calibrated dummy rather than a living human with variable posture and physiology.
Real-World Crash Databases: NASS-CDS and CISS
NASS-CDS (National Automotive Sampling System – Crashworthiness Data System) was the primary source of real-world crash injury data in the U.S. for decades. It used a probability sample of police-reported tow-away crashes, with trained investigators collecting vehicle damage, occupant injury (via AIS coding), and crash configuration data. NASS-CDS is now largely superseded by CISS (Crash Investigation Sampling System), which uses a similar sampling philosophy but updated protocols, including more detailed event data recorder (EDR) recovery and a revised injury coding process.
Real-world databases provide the variability that test data lacks: real impact angles, multiple impacts, occupant ages from 16 to 90, belt use and non-use, and injury outcomes coded by medical professionals. The trade-off is data quality and completeness. Injury severity is often coded from hospital records with varying detail, crash severity must be reconstructed from vehicle deformation or EDR data, and missing values are common. For injury risk modeling, this means real-world data is essential for external validity but requires careful handling of sampling weights, missingness, and confounding.
Practical guidance: Use FMVSS and consumer test data to establish the mechanical relationship between crash pulse, restraint performance, and dummy injury metrics. Use real-world databases to calibrate how those metrics translate into actual injury risk across the population. Never treat a dummy-based criterion as a direct probability of human injury without real-world validation.
Comparing the Sources
| Source | Crash Configuration | Occupant | Injury Outcome | Best Use |
|---|---|---|---|---|
| FMVSS | Fixed regulatory tests | Standard dummies | Dummy criteria only | Baseline compliance and criterion thresholds |
| Euro NCAP | Offset, MPDB, pole, side | Standard dummies | Dummy criteria + rating | Severity variation and structural performance |
| IIHS | Moderate/small overlap, side, roof | Standard dummies | Dummy criteria + rating | Realistic overlap and intrusion scenarios |
| NASS-CDS / CISS | Real-world, all configurations | Real occupants | AIS-coded injuries | Population risk, external validation |
From Source to Model Input
When building an injury risk model, the first decision is the target injury metric. If the goal is to predict AIS 3+ head injury in frontal crashes, the model needs a link between a measurable crash parameter (delta-v, barrier equivalent speed, or dummy HIC) and the probability of severe head injury. FMVSS and Euro NCAP provide the dummy HIC values under controlled conditions; CISS provides the real-world delta-v and AIS outcomes. The modeling task is to bridge them.
A common workflow is:
- Extract dummy injury criteria from test databases for the relevant crash mode.
- Extract real-world cases with matching crash mode, delta-v range, and occupant characteristics.
- Fit a statistical model (e.g., logistic regression or survival model) relating the test-based criterion to real-world injury probability.
- Validate the model on a held-out subset of real-world cases and check calibration across delta-v bins.
For example, a simplified logistic model for AIS 3+ head injury might use HIC 15 as the predictor:
logit(P(AIS3+ head)) = -9.2 + 0.008 * HIC15
With this illustrative coefficient, HIC 15 = 700 gives a predicted probability of about 0.12, while HIC 15 = 1,000 gives about 0.45. These numbers are example parameters for teaching purposes and must be re-estimated from actual data before any real application.
Common pitfall: Mixing test data and real-world data without accounting for sampling weights. NASS-CDS and CISS use complex probability sampling; unweighted analysis will bias injury rates toward the over-sampled crash types. Always apply the provided case weights when estimating population-level risk.
Data Quality and Access Considerations
FMVSS test data is available through NHTSA's public databases, though detailed time histories may require a request. Euro NCAP publishes test results and some dummy response data through its official reports. IIHS provides detailed test reports and video, but raw channel data is less accessible. NASS-CDS and CISS data are available through NHTSA's file transfer system, with documentation on sampling, variables, and imputation. For any modeling project, document the data version, extraction date, and inclusion criteria, because these databases are updated periodically and case counts change.
The most defensible injury risk models use test data to anchor the mechanical dose-response relationship and real-world data to calibrate the absolute risk level. Treating either source as sufficient on its own is a common source of overconfident predictions.
The diagram above summarizes the relationship: controlled test data anchors the mechanical dose-response, real-world data calibrates the absolute risk level, and validation closes the loop. The next section will build on this foundation by examining how crash severity is quantified and reconstructed from these sources.