Sam Peltzman Said Safe Cars Make Reckless Drivers. Fifty Years of Toxicology Reports Say Otherwise.

According to the toxicology reports, Sam Peltzman got one very important thing wrong.

In 1975, the University of Chicago economist published a paper arguing that vehicle safety mandates produce offsetting behavioral changes: require seatbelts and crush zones, and drivers will compensate with recklessness, erasing the gains.^[1] His thesis was elegant, counterintuitive, and politically irresistible to anyone who wanted a reason to oppose airbag standards, electronic stability control mandates, or federal crash structure requirements. For fifty-one years, the "Peltzman Effect" has been the go-to citation in committee rooms where safety regulation goes to stall.^[2] Nobody tested it against the most granular vehicle-level impairment dataset on the planet.

We did.

FARS toxicology data covers every driver in a fatal crash in the United States from 2014 through 2023, with blood alcohol and drug screening results attached to each record.^[3] We cross-tabulated those impairment flags against per-vehicle fatality rates for 198 models meeting our threshold: at least 100 total deaths and 50 toxicology-screened drivers, enough volume to produce stable percentages rather than small-sample noise. Sorted into quartiles by fatality rate, Q1 averages 0.33 deaths per 100 million VMT while Q4 averages 2.89, a nearly ninefold gap in how likely a given mile of driving kills someone.

Peltzman's framework predicts that Q1 drivers, cocooned in five-star crash structures and curtain airbags, should exhibit higher impairment rates because a safer car licenses more risk-taking, and toxicology screens should catch the transaction.

Q1 impairment: 19.5 percent, with Q2 at 20.6, Q3 at 20.4, and Q4 at 19.7. A 1.1-percentage-point spread separating the safest vehicles from the deadliest, well within the noise floor of jurisdictional variation in testing protocols, not a trend, not a weak signal, just stubbornly and completely flat.

Individual outliers are fascinating in isolation because they cancel in aggregate. Toyota's Solara carries a fatality rate of 4.25 per 100M VMT with only 4.1 percent impairment: 95.9 percent of fatal-crash drivers were sober, dying at more than twelve times the rate of a Model 3, which suggests a design problem that has nothing to do with the bottle. Land Cruiser runs worse by rate at 6.27, with 8.9 percent impaired and 91.1 percent sober. Flip the axis and you find a Pontiac Vibe posting a respectable 0.54 fatality rate alongside 25.4 percent impairment, while a Cadillac CTS at 1.32 carries 25.9 percent impaired, safe-ish vehicles with drunk-ish drivers that nonetheless scatter across the plot without forming a trend line. Uncorrelated variables behave exactly like this when the sample is large enough: noise scatters, signal resolves, and the resolved signal is a flat line at roughly one in five.

On methodology and what we are actually measuring: FARS captures only fatal crashes, roughly 36,000 of the 6.7 million annual U.S. crashes. Toxicology testing rates vary by state, with some jurisdictions screening every driver and others testing fewer than half, introducing geographic bias. Our fatality rates use VMT projections from FHWA/NHTS survey data, not odometer readings, which introduces approximately ±15 percent uncertainty for low-volume models.^[4] We are measuring selection effects across vehicle populations, not behavioral changes within individuals; a Volvo buyer and a Mustang buyer differ in ways that precede the purchase. What this cross-tabulation tests is whether safer vehicles systematically aggregate higher impaired-driving rates across the entire U.S. fatal-crash population over a decade, and they do not.

Full-strength counterargument. Peltzman wrote about regulation-induced behavioral change in response to seatbelt laws, not about the type of person who gravitates toward a particular vehicle. A true Peltzman test would track identical drivers across vehicles of varying safety levels and measure whether their risk behavior shifts. FARS was never built for that study design. Our cross-tabulation cannot disentangle self-selection from behavioral adaptation. Someone who buys a CX-5 and someone who buys a Challenger were probably different people with different risk profiles long before either walked into a dealership, and the flat impairment rates might simply reflect those pre-existing differences rather than an absence of risk compensation.

We grant that distinction entirely, but the policy conclusion survives it. Peltzman's thesis has been deployed for fifty years to argue that making the fleet safer through regulation produces offsetting recklessness in the population. If that claim held at the fleet level, safer vehicles should aggregate higher impairment regardless of whether the mechanism is individual behavior change or demographic self-selection, and they don't. Whether individuals refrain from drinking more when they buy safer cars, or whether sober drivers simply prefer safer vehicles, or some combination of both, the safety dividend is real and uncompensated. Across 198 models and tens of thousands of toxicology-screened drivers, regulation worked and no impairment offset appeared.

Several limitations should constrain how far anyone stretches this finding. Impairment is one dimension of the risk compensation Peltzman described, and drivers might compensate through speeding, following distance, distraction, or seatbelt non-use without touching alcohol, channels that would not register in toxicology data at all. Newer, safer vehicles have accumulated fewer fleet-years, leaving open the possibility that behavioral adaptation operates on a longer timeline than our ten-year window captures. Non-fatal crash impairment patterns remain entirely invisible to this analysis, and a vehicle with low fatal-crash impairment could have high non-fatal impairment if its safety features keep impaired drivers alive while sober drivers in less safe vehicles die at higher rates.

Those caveats noted, the burden of proof should sit where it belongs. IIHS estimates that vehicle safety improvements have saved hundreds of thousands of lives since 1975.^[5] Anyone invoking Peltzman in a committee room to slow those improvements owes the rest of us a dataset where the effect actually shows up at the vehicle level, and we checked 198 models over ten years and it was flat.

If you're shopping for a car: ignore anyone who tells you safety features breed complacency. Check your model at iihs.org/ratings and look up fatality rates at FARS.^[3][6] Buy the best crash structure and active safety suite you can afford, because FARS toxicology says your driving habits will not mysteriously degrade when your car has curtain airbags. Buy the safer car.