A matched case-control analysis of autonomous vs human-driven vehicle accidents

The study addresses the unresolved question of how accident risks and circumstances differ between Autonomous Vehicles (AVs) and Human-Driven Vehicles (HDVs), a topic hindered by limited real-world AV crash data. Given that human error contributes to the majority of crashes, AVs are expected to reduce crashes; however, on-road testing has shown both benefits and emerging safety challenges. Prior findings (e.g., RAND) suggest substantial long-term safety benefits even with modest AV safety improvements, yet AV miles traveled and crashes in California have risen alongside testing. Against this context, the purpose of the study is to quantify and compare accident occurrence characteristics between AVs (ADS Level 4 and ADAS Level 2) and HDVs using multi-source datasets, and to identify specific conditions under which AV crash risk differs from HDV risk. The study’s importance lies in informing AV safety design, deployment policies, and targeted improvements by rigorously controlling for confounders via a matched case-control framework.

Past work has examined AV crash patterns and disengagements with relatively small or unbalanced samples, often lacking matched HDV comparisons. Studies have reported mixed conclusions: some indicate lower AV crash rates than HDVs (e.g., Google cars’ lower police-reportable crash rates), while others find higher AV crash rates under limited, less demanding conditions or higher rear-end involvement where AVs are often struck from behind. Other research explored crash severities and contributing factors for AVs and ADAS, noting influences of environment, pre-crash movement, and system level. However, earlier studies frequently lacked comprehensive road environment, pre-accident condition, and matched HDV controls, limiting causal interpretation. This paper addresses these gaps using a larger, multi-source dataset (AVOID, NHTSA, SWITRS) and a matched case-control design to isolate differences in odds of crash involvement for AVs versus HDVs across road, environment, pre-crash movement, crash type, and outcomes.

Data sources and samples: The full AV dataset comprises 2100 crashes combining ADS (SAE Level 4; 1099) and ADAS (SAE Level 2; 1001) from the AVOID dataset (CADMV and NHTSA). The HDV dataset includes 35,133 crashes from California SWITRS, aligned by year and harmonized in format. Variables span road and environmental conditions (weather, surface, lighting, special conditions), pre-accident conditions (driving mode, vehicle movement), crash type, and outcomes (severity). For the matched case-control analysis, the focus is on ADS (California) versus HDV crashes to avoid confounding from ADAS conventional/manual operation. AADT by road type from the California Traffic Census was considered as contextual exposure information.

Matched design and strata construction: Each stratum centers on one ADS crash (case) matched to HDV crashes (controls) by location (same site; if insufficient, within 5 miles for intersections/urban segments), same road type (to align geometric design), and controlled day-of-week and time-of-day to approximate similar traffic patterns. Coordinates were extracted via Google Maps API and road types identified using geospatial processing (OSMnx). Initial models varied the case:control ratio from 1:1 upward; coefficients stabilized around 1:5, which was selected (Figure 6). The matched dataset used 548 ADS cases and 2,740 HDV controls (1:5), yielding N=548 strata.

Statistical model: Conditional logistic regression for matched case-control data was employed. Let P_ij(x) be the probability that observation j in stratum i is an AV crash, with covariate vector X_ij. The logit link with linear parameters estimates the odds of AV versus HDV involvement as a function of covariates. Inference used the conditional likelihood accounting for stratification; odds ratios (OR) are exp(beta). Modeling was conducted in R (survival package). Model fit and inference: McFadden pseudo-R^2=0.810; LR, Wald, and Score tests indicated strong overall significance. A separate random-parameter logit on full data (not detailed) corroborated key signals (notably positive effects for dawn/dusk and turning; negative for rain, rear-end, broadside, proceeding straight, run-off road, backing, entering traffic lane).

Descriptive analyses: The study also compared distributions across AV (ADS+ADAS) and HDV crashes, including participant types, weather/lighting, special conditions (work zones/traffic incidents), pre-accident movements, and crash types. Additional exploration contrasted ADS vs ADAS conditions and pre-crash speed patterns (heatmaps) to contextualize differences (e.g., ADAS more highway-oriented).

Matched case-control (ADS vs HDV; Table 1):

• Road/environment:
  • Rain: OR=0.335, 95% CI [0.195, 0.577], t=-3.948, p<0.001 (lower odds for ADS than HDV).
  • Dawn/dusk: OR=5.250, 95% CI [2.552, 10.788], t=4.508, p<0.001 (higher odds for ADS than HDV).
• Crash type:
  • Rear-end: OR=0.457, 95% CI [0.317, 0.659], t=-4.186, p<0.001 (lower odds for ADS).
  • Broadside: OR=0.171, 95% CI [0.088, 0.332], t=-5.221, p<0.001 (lower odds for ADS).
• Outcomes:
  • Moderate injury: OR=0.634, 95% CI [0.406, 0.991], t=-1.998, p=0.046 (lower odds for ADS).
  • Fatal: OR=0.097, 95% CI [0.057, 0.164], t=-8.733, p<0.001 (substantially lower odds for ADS).
• Pre-accident movement:
  • Proceeding straight: OR=0.299, 95% CI [0.178, 0.504], t=-4.544, p<0.001 (lower odds for ADS).
  • Run-off road: OR=0.021, 95% CI [0.009, 0.046], t=-9.545, p<0.001 (much lower odds for ADS).
  • Entering traffic lane: OR=0.267, 95% CI [0.211, 0.338], t=-10.954, p<0.001 (lower odds for ADS).
  • Backing: OR=0.491, 95% CI [0.367, 0.656], t=-4.803, p<0.001 (lower odds for ADS).
  • Turning: OR=1.988, 95% CI [1.421, 2.781], t=4.012, p<0.001 (higher odds for ADS).

Descriptive insights (full data):

• Sample sizes: AVs=2100 (ADS 1099; ADAS 1001), HDVs=35,133. Vehicles were 80% of AV crash participants vs 63% for HDVs; pedestrians 3% (AV) vs 15% (HDV). No or minor injuries comprised ~94% of outcomes for both.
• Special conditions: ~5% of AV crashes occurred in traffic incidents/work zones vs ~1.3% for HDVs.
• Weather/lighting: Crashes mostly under clear weather; AVs had more crashes in rain (11%) than HDVs (5%). AVs had 3.5% of crashes at dawn/dusk vs 4.9% for HDVs (descriptive), but matched modeling shows significantly higher ADS odds under dawn/dusk relative to HDV.
• Crash type: Rear-end most common for both; HDV rear-end 45%, head-on 33%. AV rear-end 39%, head-on ~33%.
• ADS rear-end conditions (Fig. 3): 79% were HDV hitting ADS; 21% ADS hitting HDV. When ADS hit HDV, 72% of ADS were in conventional mode; when HDV hit ADS, 64% of ADS were in autonomous mode. Injury severity skewed toward minor for both scenarios (e.g., 82% minor when HDV hit ADS; 67% minor when ADS hit HDV).
• ADS vs ADAS contrasts: ADAS had fewer crashes under clear weather but more under rain; more in traffic events/work zones and on wet surfaces; higher share proceeding straight and fewer turning crashes than ADS; injury outcomes showed more no-injury and slightly fewer fatal injuries for ADAS relative to ADS. Pre-crash speed heatmaps reflect higher average speeds for ADAS (highway focus) vs ADS (urban focus).

The matched case-control results directly address the research question by isolating how crash odds for ADS differ from HDVs under similar spatiotemporal and roadway contexts. ADS demonstrate significantly lower odds than HDVs in many common scenarios and movements (rain, rear-end, broadside, proceeding straight, run-off road, entering traffic lane, backing) and are associated with lower odds of moderate and especially fatal injury. These reductions are consistent with capabilities such as multi-sensor perception, precise control, and longitudinal/lateral assistance (e.g., ACC) that can mitigate rear-end and lane-keeping-related risks.

However, ADS exhibit higher odds of crash involvement under dawn/dusk lighting and while turning. These conditions challenge perception and planning: rapidly changing illumination, glare, and shadows can degrade sensor performance and object recognition, and turns (especially unprotected and in mixed traffic) require complex situational awareness, trajectory planning, and interaction with other road users. The findings align with prior concerns about AV performance in complex, dynamic scenarios and suggest areas for targeted improvement (sensor fusion robustness, redundancy, advanced perception under adverse lighting, better turn-planning and interaction modeling).

Overall, results suggest ADS are generally safer than HDVs across many comparable situations but still face elevated risks in specific edge-case environments and maneuvers. This nuanced picture supports a differentiated safety strategy focused on known weaknesses while leveraging strengths to reduce common crash types and severities.

This study integrates multi-source real-world crash data and applies a matched case-control design to compare ADS (SAE Level 4) with HDVs under closely matched conditions. The analysis shows ADS are associated with lower odds of crashes across numerous scenarios (rain, rear-end, broadside, proceeding straight, run-off road, entering traffic lane, backing) and substantially lower odds of fatal outcomes, while exhibiting higher odds during dawn/dusk and turning maneuvers. These insights provide actionable guidance for improving AV safety, particularly in perception under challenging lighting and complex turning scenarios, and support the broader conclusion that ADS can enhance safety in many typical driving contexts.

Future research directions include: incorporating detailed right-of-way and control data at intersections (yield/stop/priority signals/traffic lights); distinguishing AV system levels and activated features at crash time; enhancing exposure modeling; and engaging AV experts to qualitatively assess contributors to differential safety between AVs and HDVs.

Key limitations include: limited detail on specific AV system levels and the exact activated ADAS/AV features at the time of crashes; potential residual confounding despite location/time/road-type matching; focus on ADS in California for the matched analysis (generalizability to other regions and systems may be constrained); and incomplete availability of contextual variables (e.g., detailed right-of-way control). Additional exposure data and richer contextual information would improve model precision and external validity.