Every Car Has the Same Drunk Drivers — So Why Do Death Rates Vary 285x?

I ran a Pearson correlation between impairment rates and fatality rates across 200 vehicle models with 100 or more deaths in the FARS database. R-squared: 0.038. Impairment explains 3.8% of why some cars kill more than others. The other 96.2%? That’s the vehicle talking.^[1]

This isn’t a finding I expected. The working assumption in crash analysis—the one that underlies drunk driving campaigns, ignition interlock laws, and a billion-dollar DUI enforcement apparatus—is that driver behavior is the dominant variable. Sober people in safe cars, drunk people in dangerous ones. The data says otherwise.

The Constant

Across 490,736 drivers involved in fatal crashes between 2014 and 2023, impairment rates barely budge. Sedans: 20.4%. Pickups: 20.1%. SUVs: 19.5%. Vans: 18.1%. Sports cars push it to 22.5%, which, fine. But the coefficient of variation across individual models is just 13.7%.^[1]

Put differently: if you pulled a driver at random from a fatal Prius crash and a fatal Cobalt crash, the odds of each being impaired are nearly identical—16.0% versus 22.4%. A six-point gap. Meanwhile, the Cobalt’s death rate is 9.3 times higher.^[2]

Same drivers. Different coffins.

The Variable

Death rates per 100 million vehicle miles traveled range from 0.03 (Tesla Model Y) to 8.54 (Hyundai Veloster). Coefficient of variation: 99.2%. That’s 7.2 times more variable than impairment rates.^[1]

The midsize sedan comparison is particularly damning. Every vehicle below sits in the same market segment, targets the same buyer demographic, parks in the same driveways. Impairment rates cluster within two percentage points of each other. Death rates don’t.

The Accord’s death rate is 5.3 times the Optima’s. The Optima has the highest impairment rate in the group. If driver behavior were the dominant predictor, the Optima should be the deadliest car on this list. It’s the safest by a factor of five.^[1]

The Backward Correlation

The correlation isn’t just weak. It’s negative. Pearson r = −0.195 across all 200 models. Vehicles with more impaired drivers tend to have lower death rates.^[1]

By class, it’s consistent:

• Sedans: r = −0.27
• SUVs: r = −0.29
• Pickups: r = −0.12
• Sports cars: r = −0.52

Why? Because newer, heavier, better-engineered vehicles attract the full spectrum of human behavior, including the impaired 20%. But those vehicles are also the ones with curtain airbags, ESC, automatic emergency braking, and crumple zones designed by people who actually cared. Old vehicles—the ones with 4x or 8x the death rate—kill their sober drivers just as efficiently as their drunk ones. A 2003 Chevrolet Tracker doesn’t care about your BAC.^[3]

What This Means (and What It Doesn’t)

Drunk driving kills people. That’s not in dispute and this analysis doesn’t change it. What the data does say: if your goal is reducing variation in fatality outcomes across vehicles, impairment is the wrong knob. It’s already roughly constant. The variable that actually moves the needle—the one creating a 285x spread in death rates—is the vehicle itself. Weight. Structure. Safety tech vintage. The year the engineers were allowed to add side curtain airbags versus the year management said it was too expensive.^[3][4]

The policy implication: every dollar spent on ignition interlocks and sobriety checkpoints is targeting a variable that explains less than 4% of the outcome variance. Accelerating fleet turnover—getting 2003 Trackers and 2005 Cobalts off the road—targets the other 96%.^[5]

Methodology

Pearson product-moment correlation computed over 200 vehicle models (FARS 2014–2023) with ≥100 deaths each. Death rates use estimated VMT from fleet size × NHTS annual mileage averages. Impairment defined as BAC > 0 or drug-positive toxicology per FARS coding. Coefficient of variation = standard deviation / mean × 100. All computations performed on the same FARS dataset underlying this site’s other 96 articles.

Limitations

FARS captures only fatal crashes—roughly 38,000 per year out of 6.7 million total crashes. A vehicle with a low fatality rate might still have astronomical injury or property-damage rates. VMT estimates use fleet-level averages, not odometer data, introducing ±15% uncertainty for low-volume models. The impairment data reflects drivers in fatal crashes, not drivers on the road—impaired drivers are overrepresented in FARS by definition. This analysis can’t separate vehicle age from vehicle design; a 2003 Tracker’s death rate includes 20 years of accumulated exposure. And Pearson r measures linear relationships—a nonlinear interaction between impairment and crash lethality wouldn’t surface here.

The Counterargument, at Full Strength

Impairment does cause crashes. What this analysis measures is variation in fatality rates, not crash incidence. An impaired Camry driver is still far more likely to crash than a sober one. The ~20% constant means that once a crash is fatal, the impairment mix is similar regardless of vehicle. But that’s exactly the point: impaired drivers are everywhere, and what separates life from death once they crash is the sheet metal and silicon around them. The counterargument and the thesis arrive at the same destination.