The study that went looking for cannabis and found the seat belt

In 2021, two economists, Richard Fowles and Peter Loeb, published a paper in the Journal of Transport & Health with a straightforward title: the association between marijuana and motor vehicle crashes. Their raw material was a balanced panel of the 50 US states plus Washington DC across the years 2010 to 2016, which works out to 357 state-years of data. The thing they were trying to explain was each state's road-death rate, expressed the sensible way: fatalities per 100 million vehicle-miles travelled, so a big state with more driving is not automatically counted as more dangerous. Against that they lined up the usual suspects, alcohol consumption, posted speed limits, seat belt use, mobile phones, the age of the car fleet, unemployment, and their headline variable, marijuana use.

Here is the first thing to hold on to, because it shapes everything after. Their marijuana measure was not impaired drivers at the wheel. It was the percentage of adults in each state who said they had used cannabis at all in the past year, taken from a national health survey. That is an ecological measure: a population-level figure that says nothing about whether the person who used cannabis in March was the person who crashed in November, or whether they were even the same kind of person. Ecological data can hint at patterns, but it cannot pin a crash on an individual. The same caution, in a milder form, applies to their mobile-phone variable, which was really phone subscriptions per head, a proxy the authors admit may just be standing in for distraction in general. Keep that in your pocket. It matters later.

They ran that stress test under three different attitudes, which the paper calls priors and which you can read as three moods. The wide prior is the most demanding, an agnostic "show me". The pessimistic prior assumes the variables probably do not explain much, so a finding has to fight to survive. The optimistic prior assumes the variables do matter and lets the data speak more freely, so findings survive more easily under it. A finding that only holds up under the optimistic mood is real but conditional; a finding that holds up under all three is about as solid as this kind of data gets. That single distinction is the spine of the whole story.

This is the heart of it. The paper's own Table 3 reports, for each variable, whether it stayed sturdy under the wide, pessimistic and optimistic priors. Line them up by that count and you get a ladder. At the top sits the one variable in perfect conformity, the same sign and a strong result every single way you look at it. At the bottom sits the recreational-legalisation law, which held up nowhere. In between, tellingly, the frightening headline variables cluster on the bottom rung of "only under the friendliest assumptions".

Read the ladder as a whole and the shape of it is the lesson. The variable with the most alarming press, the one the study was built around, sits on the weakest rung it can occupy while still counting at all. The variable nobody writes a headline about, the seat belt, sits alone at the top, sturdy under the demanding wide prior, the sceptical pessimistic prior and the generous optimistic prior alike. And the specific legal change that generates the most heat, whether a state has legalised recreational cannabis, does nothing detectable to the death rate on its own. A study designed to measure the frightening thing ended up handing its cleanest result to the boring thing.

First, the single most important distinction in the whole piece: sturdy is not the same as proven, and association is not the same as cause. When we say the seat belt was the sturdiest variable, we mean something narrow and specific. We mean that in this one panel of US data, the statistical link between higher belt-wearing and fewer deaths held its sign and its strength across every reasonable model the authors tried. That is a strong statement about the robustness of an association. It is not, on its own, proof that belts save lives. This study did not, and could not, prove that. What makes us confident that belts genuinely save lives is a different and stronger body of evidence, built from individual crash records rather than state averages, which we come to in the next section. Hold both thoughts at once: the panel study shows the belt as the most stubbornly reliable association in the data, and separate, better-designed research supplies the causal proof. The article rests on the second, and uses the first only to make its point about certainty.

Second trap: fragility hides inside a confident-sounding sentence. The marijuana finding is the case study. Under the optimistic prior it is sturdy, with an s-value of 1.08, just over the line, and the paper leans on that to make firm claims. But loosen the mood one notch to the wide prior and the picture changes: the estimated effect runs from below zero to above zero. In plain terms, the assumptions you bring decide whether the effect even points the way you expected. A finding that depends that heavily on the analyst's starting mood is not worthless, but it is conditional, and it deserves a hedged sentence, not a headline. The belt needs no such hedge, because it never crossed zero under any mood.

Put the three together and the belt's victory stops being a surprise. It was the best-measured variable, it never flipped sign, and it survived the harshest assumptions. That combination is what statistical certainty actually looks like, and it is why the unglamorous fundamental beat the frightening headline. The frightening thing was not disproven. It was simply shown to be a less certain finding than the boring thing sitting quietly next to it in the same table.

If we stopped at Fowles and Loeb, a fair critic could say: interesting, but a sturdy association across US state averages is still just an association. Quite right. So here is the causal backbone, from the kind of research designed to establish cause, individual occupants in real crashes, matched and compared. It is settled, and it is not close.

The primary figure comes from the US National Highway Traffic Safety Administration. Analysing decades of real crash records, NHTSA's technical work reconfirmed that a standard three-point lap-and-shoulder belt reduces the risk of a fatal injury to a front-seat car occupant by about 45%, and by about 60% for occupants of light trucks and vans. For serious but non-fatal harm the same belt roughly halves the risk of moderate-to-critical injury in a car, and cuts it by around 65% in an SUV, van or pickup. These are individual-level, causal-grade numbers, and they are the ones NHTSA still relies on.

The World Health Organisation puts the same message in round global terms: wearing a seat belt can cut an occupant's risk of death by up to half, with the standard differentiated figure being a 45 to 50% reduction for drivers and front-seat occupants and roughly a quarter for rear-seat occupants. Different bodies, different datasets, the same direction and roughly the same magnitude. That convergence is what a genuinely causal, well-established finding looks like, and it is the opposite of a fragile one.

This is the ground the article actually stands on. The panel study is a lens, a way of seeing that even a dataset built to chase a frightening variable kept pointing back at the belt. The reason to wear one, and to insist your passengers do, is the causal evidence above: individual crash records, consistent across agencies and decades, showing that this one cheap habit removes something on the order of half of your risk of dying. No fragile headline comes anywhere near that combination of size and certainty.

Start with the law, because it is simple and it is stricter than many drivers assume. Where seat belts are fitted, they must be worn by the driver and every passenger, front and rear. Children under 150cm or under 36kg must use a suitable child restraint, and a rear-facing seat must never go in front of an active airbag. Responsibility is split by age: the driver is legally responsible for making sure any passenger under 17 is belted or properly restrained, and it is the driver who collects the penalty if they are not; passengers aged 17 and over answer for themselves.

The penalty is the same whether it is your own belt or an under-17 you failed to restrain: a fixed-charge notice of €120 with 3 penalty points if paid within 28 days, rising to €180 if paid between days 29 and 56, and 5 penalty points on a court conviction. That €120 figure is current: it doubled from €60 on 27 October 2022, when fines for sixteen road-safety offences were raised together. It is worth naming, because older material, including a couple of pages on this very site that we are correcting, still shows the pre-2022 figures.

Notice what that table is. It is the boring-basics trio from the research, transplanted into Irish law almost one for one. The seat belt sat at the top of the study's robustness ladder; speed was the next most Bayesian-robust variable; the mobile phone stood in for distraction, which the wider evidence treats as a genuine killer. These are not this year's frightening headline. They are the dull, well-established fundamentals, and the Irish penalty-points system is, in effect, betting its enforcement on the same three levers the statistics keep pointing at.

One forward pointer, and then we leave it, because it is a separate subject: the study's headline variable was cannabis, and Irish law does treat drug-driving seriously, with its own testing regime and penalties. That belongs in its own guide and we will give it one. It is not the point here. The point here is that when you sort the same evidence by certainty rather than by drama, the levers that rise to the top are the three plain ones you already control every time you turn the key.