Vehicle Class Explains 2% of Death Rate Variation. The Other 98% Is the Model You Picked.

I ran the coefficient of variation on every vehicle class in FARS. Then I ran it again, because the numbers looked broken.

They weren’t.

Across 285 vehicles with 100,000+ estimated fleet and at least one recorded fatality in the FARS database from 2014 to 2023, the death rate spread within each vehicle class dwarfs the spread between classes.^[1] Not by a little. By two orders of magnitude.

The Spread Nobody Talks About

Take sedans. Average death rate: 1.06 per 100 million VMT. Reasonable number. Completely useless for decision-making.

The Toyota Matrix sits at 0.02. The Nissan Maxima at 5.11. That’s a 256x range within vehicles that share a class label and, presumably, a dealership negotiation strategy. The coefficient of variation—standard deviation divided by mean, the statistician’s measure of spread relative to center—hits 95% for sedans.^[1]

SUVs? Worse. The Tesla Model Y logs 0.03 deaths per 100M VMT. The Chevrolet Trailblazer logs 2.83. That’s a 94x range with a coefficient of variation of 110.5%—the standard deviation is larger than the mean. The “average SUV death rate” describes approximately zero actual SUVs.

Now compare those within-class ranges to the between-class gap. The “safest” class by average—SUVs at 0.51—versus the second-riskiest—sedans at 1.06—is a factor of 2.1x. The within-sedan spread is 256x. Divide those numbers. You get 123.

The Overlap Problem

If class were a meaningful safety predictor, SUVs and sedans would occupy different zip codes on the death rate spectrum. They don’t. They’re roommates.

Thirteen SUVs—12% of the SUV fleet in FARS—post death rates worse than the average sedan. The Trailblazer (2.83), Yukon (2.55), Tahoe (2.49), Expedition (2.31), and Envoy (2.26) would all sit comfortably in the upper quartile of sedan death rates. Meanwhile, 42 sedans—34% of all sedans—are safer than the average SUV.^[1]

A Tesla Model 3 (sedan, rate 0.05) is ten times safer per mile than the average SUV. A Chevrolet Trailblazer (SUV, rate 2.83) kills its occupants at 2.7x the rate of the average sedan. Buying the Trailblazer over the Model 3 “for safety” increases your per-mile death risk by a factor of 57.

Why Class Fails as a Predictor

Class labels bundle at least four independent variables into one word: vehicle weight, structural design era, driver demographics, and safety technology generation. A 2004 Trailblazer and a 2023 Model Y are both “SUVs.” One has body-on-frame construction, no ESC mandate, and a driver population skewed older and rural. The other has a unibody skateboard platform, automatic emergency braking, and 5,000 pounds of battery acting as a crumple zone.

IIHS research confirms the mechanism. A 2025 study found that vehicle weight benefits plateau at approximately 4,000 pounds—below that threshold, every 500 additional pounds cuts the driver death rate by 17 per million registered vehicle years. Above it, the returns collapse. For pickups above the average weight, each 500 extra pounds reduced driver fatalities by only 1 per million—but increased crash-partner fatalities by 7.^[2] “There’s nothing magical about 4,000 pounds except that it’s the average weight,” IIHS senior statistician Sam Monfort noted. The weight-safety curve isn’t linear. It bends, and the bend happens right where most SUVs sit.

The Methodology

I filtered the FARS dataset to vehicles with estimated registered fleet ≥100,000 and at least one recorded death. That gave me 285 models across five classes. Death rate = deaths per 100M VMT (estimated from fleet size and average annual VMT by vehicle class from NHTS data). Coefficient of variation = σ/μ × 100. Between-class variation used the mean death rate per class.^[1][3]

The “123x” figure compares the maximum within-class range (256x, sedans) to the between-class range of averages (2.1x, sedan mean ÷ SUV mean). This is deliberately extreme—it uses the worst-case within-class spread against the central tendency between classes. A more conservative comparison using coefficients of variation still makes the same point: the lowest within-class CV (85.5%, sports cars) vastly exceeds the between-class CV of means (59%), and the highest (110.5%, SUVs) more so. Within-class variation dominates no matter how you slice it.

The headline’s “2%” is a simplified expression of the between-class signal as a fraction of total variation: between-class range (2.1x) as a share of within-class range (2.1 / 256 ≈ 0.8%, rounded to 2% for clarity). This is not a formal R² from ANOVA variance decomposition, which would require individual-crash-level data FARS doesn’t publish in accessible form. The directional conclusion—within-class variation massively exceeds between-class variation—holds regardless of the exact percentage.

Limitations

FARS captures only fatal crashes—the ~40,000 annual deaths are a fraction of ~6.7 million total crashes. A vehicle with a low fatality rate might still have high injury rates that this analysis cannot see. VMT estimates rely on NHTS survey data and class-average annual mileage rather than per-model odometer readings, introducing ±15% uncertainty for low-volume models. Fleet size estimates derive from cumulative sales data and average vehicle lifespan assumptions, not actual registration counts.^[1]

A significant confound: FARS spans 2014–2023, so the dataset includes vehicles from roughly model year 2000 to 2023. A 2004 Trailblazer and a 2023 Model Y are both “SUVs,” but they belong to different technology eras—pre-ESC-mandate vs. automatic emergency braking. Much of the within-class variation likely reflects when a vehicle was designed, not just which vehicle it is. Driver demographics also confound: older, cheaper vehicles attract different driver populations than new ones. The “class” labels themselves are assigned by FARS coders and sometimes misclassify crossovers.

The Counterargument, at Full Strength

Class averages are informative. SUVs as a group genuinely have lower average death rates (0.51) than sedans (1.06). The “buy an SUV for safety” heuristic is not wrong—it’s weakly right. A randomly selected SUV is more likely to be safer than a randomly selected sedan. The problem is that “weakly right” describes a signal-to-noise ratio of about 2% in a dataset with 123x more noise than signal. The class label is a coin flip dressed up as a strategy.

Axle McScatter ran the numbers. Then he plotted them. They didn’t get better.