Cannabis intoxication and fatal road crashes: case-control study

Are drivers under the influence of cannabis at higher risk of being responsible for road crashes? Peter Leman explains how a population based case-control study was used to investigate a possible association

This month's paper is Laumon B, Gadegbeku B, Martin J-L, Biecheler M-B. Cannabis intoxication and fatal road crashes in France: population based case-control study. BMJ 2005;331:1371-4 (10 December).

Why do the study?

This study was sponsored by the French government to obtain data on drug and alcohol consumption and road crashes in France over a two year period (2001-3). In particular, the authors wanted to explore the association, or lack thereof, between cannabis use and alcohol use and their relation to road crashes.

The authors used a case-control study, with the population of France as the potential population at risk. It is important when reading any case-control study to ensure that the potential cases and controls are not skewed in some way (for example, patients attending a specialist clinic who may not represent the usual spectrum of disease). In this study, countrywide data were therefore used as a reasonably valid sampling frame.

The authors limited the cases reviewed to fatal road crashes only—this could be the death of a driver, passenger, or pedestrian. The police took any drivers involved to a hospital for urine testing or blood tests if urine couldn't be obtained. This included drivers who had died in the accident. If the urine screen was positive, blood tests were taken anyway (although this is not clear), although some alcohol readings are based on a breathalyser test. Most samples were collected within four hours, but clear timings are not provided.

In nearly half of the crashes, data were missing (9653/20 401; 47% of fatal crashes). It has to be asked how these drivers differed from those where sampling did occur—were they more or less likely to have taken drugs? Did they leave the scene of the accident? The authors did not assess them but compared the rates of alcohol use in the sampled fatal crashes with data from non-fatal creashes and found similar distributions—so the subset studied may be representative. In this situation you have to ask, however, whether assessing only half of the potential case-controls has missed something important.

The authors decided to work out if a particular driver was to be held responsible for the accident or not. Of the 10 748 drivers with blood or urine samples that they had, they determined that about 63% were responsible. The police assignment of responsibility was only one factor used, and much is made throughout the paper of the method proposed by Robertson and Drummer, which references a paper from 1994 on responsibility analysis. In this paper, which was based only on alcohol related fatalities, the result could also be termed “contributory” (it was not clear whether the driver or the circumstances had the larger role in the accident). The authors in the French paper decided to be dichotomous: you were either responsible for the crash or you weren't at all. This has the potential for misclassification bias, but from the data presented it is uncertain whether more or fewer drivers are being held responsible for the crashes seen.

It would have taken quite some time to analyse nearly 10 000 accident reports individually, and the group also used a subset of analyses of multiple vehicle crashes by other “experts” to standardise their outcomes. The internal measure used to determine how reliably the authors assigned responsibility had a κ score of just 0.71. This score measures the extent of agreement (or, more correctly, the proportion of agreement over that which would have been expected by chance alone) between the different assessors on the assignment of responsibility. A score closer to 1.0 represents total agreement between assessors. A score of 0.71 is quite reasonable, but in this study, where you were either totally innocent or totally guilty, it would have been preferable if agreement had been higher (at least 0.85). How these differential classifications were resolved isn't clearly stated; this is an important oversight.

So, having used the 63% of assumed responsible drivers as cases, the authors used the non-responsible drivers as controls. The validity of this clearly depends on whether the assignment of responsibility is accurate, and whether it can be truly dichotomous. Also they removed 976 drivers who were held non-responsible (this is a quarter of their controls) and who died. They did this because they thought that the concentrations of drugs in these drivers might be higher than in the overall population of French drivers—that is, you might not be guilty of causing the crash, but you were at higher risk of dying because you had used drugs (for example, by not wearing a seatbelt, etc). This means that, among the controls, many who might have used drugs were removed from analysis, but this didn't happen with the cases (they were dead and responsible). This is going to bias the final analysis towards more drug users being among the cases than the controls.

The authors determined the relative risk of the responsible drivers being exposed to any of the four drugs measured (cannabis, amphetamines, opiates, and cocaine), compared with the non-responsible drivers. They used cut-offs for exposure, which may allow for second hand exposure (other people were smoking but not that driver), cross reaction in the assay (for example, codeine and opiates), or lack of acute exposure when the driver smoked cannabis some hours ago—although this is an unclear science. They also introduced a dose-response analysis for higher concentrations of Δ9 tetrahydrocannabinol in urine or blood. In some analyses the authors then used all cannabis concentrations detected, and in others they used only drivers with a cannabis concentration of 0.12 ng/ml or higher in blood, which seems to suit their purpose when describing dose-response associations but not when discussing absolute exposure and risk.

The authors claimed that they needed to use attributable risk rather than odds ratios because the event was not rare. An odds ratio is an indirect determinant of the true relative risk of an event, which is valid only if the event of interest is rare. It is the odds in the exposed subjects divided by the odds in the non-exposed subjects, whereas the relative risk is the direct ratio of events in each group. In this study, the event (being held responsible for the crash) was very common, and so odds ratios shouldn't have been used. Yet all the results and the tables provided in this paper show odds ratios. One could argue which is best used in a case-control study, yet to argue against odds ratios and then use them seems difficult to reconcile.

In the poorly described validity study of non-fatal crashes, some standardisation seems to have occurred, but it is unclear how this happened, as data on drug exposure were not collected, only data on alcohol exposure. This paragraph is not developed, and so it leaves the reader uncertain how this standardisation occurred, as it was not built into the multivariate analysis.

In road crashes, many other confounders need to be taken account of. A good example is vehicle design and active and passive safety features. Few of these were analysed by the authors. It might well be that cannabis users were poorer than non-users and drove cheaper or less safe cars. For any given accident the cannabis user may, therefore, have been in a car with poorer brakes or have no airbags; this may have resulted in a greater likelihood of fatality. Or the driver may have been on a motorbike (table 4 in the paper), although this was measured and allowed for.

Throughout France, over two years, only 7% of the drivers for whom the authors actually obtained a blood sample tested positive for cannabis alone, another 2.9% for both cannabis and alcohol, and many more tested positive for alcohol (21.4%). An odds ratio of 3.32 for exposure to cannabis (≥1 ng/ml in blood) over-represents the relative risk of around 3.11 slightly. This means that the odds or risk of being held responsible for the crash was just over three times higher in drivers who had detectable cannabis in their blood and urine. A much larger association with alcohol (>0.5 g/l) was seen (odds ratio 15.5). So the consumption of alcohol sufficient to raise your blood concentration of ethanol to ≥0.5 g/l meant that you were over 15 times more likely to be held responsible for the fatal crash. However, when coingestion of alcohol is taken into account (table 3 in the paper) the odds of being the responsible driver falls over all detected concentrations of Δ9 tetrahydrocannabinol. In the multivariate model (also allowing for vehicle type, age, time of crash, but nothing else) it also falls considerably more. Thus, although the odds of any tetrahydrocannabinol on board for being held responsible is 3.17, by the time alcohol use is taken into account this falls to 2.37, and by the time other measured factors relating to the type of accident are included it falls further to 1.78 (95% confidence interval 1.40 to 2.25). This still exceeds 1, and the confidence interval doesn't include 1 either, so the odds reached significance. But this odds ratio is now quite small, and when the other biases and unmeasured confounders are considered you have to wonder how much true association is left (or if the real association might be even stronger).

In any case-control study, unmeasured confounding needs to be explained. When large variations in the reported odds occur after adjustment for just some measured confounders, one always wonders if further analysis may reduce the odds even further. Furthermore, the artificial removal of a quarter of the control group is potentially harmful to the validity of the study, let alone the missing 50% of the potential study population.

It is difficult to base strong conclusions on case-control studies, and the level of evidence drawn from these studies is quite low.

But this study is very large and comprehensive, and the clear and significant exposure ratios seen with alcohol, which are well supported by the literature, make the observations about cannabis dose and risk seem potentially plausible.

So, although alcohol is clearly a major factor in fatal road traffic accidents, cannabis seems likely to be the next largest contributor when illicit drug use is considered.