33 Vehicles Hit the Same Death-Rate Cliff Between 2018 and 2021

Something happened around model year 2019. I ran a model-year death analysis across every vehicle in the FARS database with sufficient data, looking for the point where each vehicle’s per-year fatality count dropped by at least 50% relative to prior years. Thirty-three vehicles hit that cliff in a four-year window, between model years 2018 and 2021. Not one brand. Not one segment. Sedans, SUVs, pickups, muscle cars—all of them, synchronized within three model years of each other.^[1]

The Nissan Sentra dropped 84%. From 161 deaths per model-year average to 26. The Chevrolet Equinox: 81%. The Volkswagen Jetta: 80%. Even the Dodge Challenger—a 6,000-pound muscle car whose entire brand identity revolves around going fast—shed 76% of its per-model-year fatalities after 2020.

Combined, these 33 vehicles went from approximately 2,219 deaths per year in their pre-cliff model years to 797 after. That’s 1,422 people per year who didn’t die, and that’s only counting the models with enough FARS data to calculate. The real number across the full fleet is almost certainly larger.^[1]

What Changed in Those 150 Milliseconds

In March 2016, twenty automakers signed a voluntary agreement with NHTSA and the IIHS to make automatic emergency braking standard on virtually all new light vehicles by September 2022.^[2] Tesla and Volvo were already at 100% by 2018. Toyota rolled Safety Sense across its lineup that same year. By 2019, Honda, Nissan, Subaru, and Mazda had followed. By 2020, over 90% of new vehicles shipped with AEB as standard equipment.^[3]

The IIHS estimates AEB reduces rear-end crashes by roughly 50%.^[4] Forward collision warning alone cuts them by 27%. But our FARS data suggests the combined effect—AEB plus improved crash structures, side curtain airbags becoming universal, updated IIHS testing protocols forcing structural redesigns—is producing reductions far exceeding any single technology’s contribution.

The cliff isn’t about one feature. It’s about a generation of vehicles that were designed, from the ground up, around standards that didn’t exist five years prior.

The Class Breakdown

This isn’t a sedan story or an SUV story. It’s everything:

Sedans: Nissan Sentra (−84%), Volkswagen Jetta (−80%), Chevrolet Malibu (−65%), Hyundai Elantra (−61%), Honda Accord (−57%), Toyota Camry (−56%), Honda Civic (−53%).

SUVs/Crossovers: Chevrolet Trax (−83%), Chevrolet Equinox (−81%), Jeep Wrangler (−72%), Nissan Rogue (−61%), Toyota RAV4 (−60%), Subaru Forester (−54%).

Muscle/Sport: Dodge Challenger (−76%), Dodge Charger (−66%), Chevrolet Camaro (−59%).

Trucks: Ram 2500 (−55%), Ram 1500 (−53%).

When a Jeep Wrangler—a vehicle whose occupants historically treat seatbelts as suggestions—drops 72% in fatalities at the same moment a Nissan Sentra drops 84%, you’re not looking at model-specific redesigns. You’re looking at a systemic shift in how cars interact with the moments before a crash.

The Earlier Cliff

This isn’t the first time the FARS data shows a synchronized drop. Between model years 2004 and 2008, sixty-seven vehicles show a similar cliff—correlating precisely with the rollout and eventual federal mandate of Electronic Stability Control.^[5] NHTSA finalized the ESC rule in 2007, requiring full compliance by 2012. The death rates didn’t wait for the mandate; they fell as soon as manufacturers started installing the systems, years ahead of the legal requirement.

The 2019 cliff is the ESC story replayed with a bigger cast. Twenty automakers instead of a federal rule. AEB and FCW instead of stability control. And a death reduction—64% average across 33 models—that rivals or exceeds the ESC generation.

What This Means for Used Car Buyers

A 2017 Nissan Sentra and a 2020 Nissan Sentra are, by the badge, the same car. By the FARS data, they are radically different objects. The 2020 model year kills at roughly one-sixth the rate of the pre-cliff versions. Similar gaps exist across nearly every nameplate on the used market.

If you’re shopping used and choosing between a well-maintained 2015 and a higher-mileage 2020, the data overwhelmingly favors the newer vehicle. The model year premium isn’t cosmetic. It’s structural, electronic, and measured in body bags.

Limitations

Model-year exposure bias is real and significant. A 2020 model year vehicle has only 3–4 years of exposure in the 2014–2023 FARS window, versus 9+ years for a 2005. This mechanically depresses newer model-year death counts. Our analysis uses 3-year rolling averages to partially control for this, comparing adjacent cohorts rather than endpoints, but the bias cannot be fully eliminated without per-year registration-weighted normalization, which FARS alone cannot provide.^[1]

Fleet composition also shifted during this period: sedan sales collapsed while SUV and crossover sales surged. Since SUVs have lower average death rates in the FARS data (0.95 per 100M VMT vs. 1.45 for sedans), part of the overall improvement is compositional rather than technological. But the cliff appears within both classes independently—the Camry fell 56% and the RAV4 fell 60%—which argues against a pure mix-shift explanation.

Finally, FARS records only fatal crashes. A vehicle that prevents fatalities but still produces serious injuries would appear safer in this data than it functionally is. AEB, by design, reduces impact speed rather than preventing all collisions—which may convert some fatalities into serious injuries rather than eliminating crashes entirely.

Strongest Counter

The exposure bias argument deserves its full weight: of course 2020 models have fewer deaths than 2005 models in a 2014–2023 dataset. There are simply fewer 2020 vehicles on the road during more of those years. The 3-year rolling average helps—comparing 2017/2018/2019 to 2020/2021/2022—but cannot perfectly control for the fact that a 2017 vehicle had six years of FARS exposure while a 2021 vehicle had two.

A proper counterfactual would require annual registration data cross-tabulated with FARS—deaths per registered vehicle per year, by model year—which neither FARS nor publicly available registration databases readily provide at this granularity. Until that analysis exists, the 64% figure should be understood as an upper bound on the true improvement. The synchronized timing across 33 models, however, is real regardless of the magnitude. Thirty-three independent engineering programs don’t produce the same inflection point by coincidence. The system changed. The vehicles measured it.