Background Cannabis use has been linked to impaired driving and fatal accidents. Prior evidence suggests the potential for population-wide effects of the annual cannabis celebration on April 20th (‘4/20’), but evidence to date is limited.
Methods We used data from the Fatal Analysis Reporting System for the years 1975–2016 to estimate the impact of ‘4/20’ on drivers involved in fatal traffic crashes occurring between 16:20 and 23:59 hours in the USA. We compared the effects of 4/20 with those for other major holidays, and evaluated whether the impact of ‘4/20’ had changed in recent years.
Results Between 1992 and 2016, ‘4/20’ was associated with an increase in the number of drivers involved in fatal crashes (IRR 1.12, 95% CI 0.97 to 1.28) relative to control days 1 week before and after, but not when compared with control days 1 and 2 weeks before and after (IRR 1.05, 95% CI 0.92 to 1.28) or all other days of the year (IRR 0.98, 95% CI 0.88 to 1.10). Across all years we found little evidence to distinguish excess drivers involved in fatal crashes on 4/20 from routine daily variations.
Conclusions There is little evidence to suggest population-wide effects of the annual cannabis holiday on the number of drivers involved in fatal traffic crashes.
- public health
- descriptive epidemiology
- Motor vehicle Occupant
Statistics from Altmetric.com
The potential role that cannabis use and impairment may play in traffic crashes is the subject of limited evidence. Given legislation to legalise recreational cannabis use in some US states and legalisation nationwide Canada in 2018, understanding the potential impact of cannabis consumption on traffic crash risks is needed. Traffic crashes already create a substantial burden of death and injury, so reported increases in drugged driving1 2 are concerning. Prior studies have suggested that, like alcohol, cannabis use impairs driving in simulated environments.3
Another strand of research has focused on population-level risks, and systematic reviews4 have suggested that cannabis use is associated with an approximate doubling of fatal crash risk, though there is important heterogeneity by study design and quality.5 On the other hand, two separate studies found that US states that legalised medical cannabis had reduced rates of fatal traffic crashes after passing legislation.6 7 These results have been interpreted as supporting the idea that marijuana and alcohol may be substitutes,8 particularly for younger people, or perhaps that cannabis users are more likely to consume it at home and are therefore less likely to drive.
However, most studies of cannabis-impaired driving have either been conducted in a laboratory setting or in local field trials, so population-level evidence is limited. Recently, however, Staples and Redelmeier reported that the annual ‘4/20’ cannabis celebration that occurs on April 20th each year was associated with a 12% increase in the number of drivers involved in fatal traffic crashes in the USA.9 In this paper, we aim to further this evidence by comparing the impact of 4/20 with other risky days identified in prior work,10–12 as well as evaluating whether the impact of 4/20 has changed over time.
We used publicly available data from the Fatal Analysis Reporting System,13 14 which covers all crashes in which at least one person died (within a month of the crash) in the USA. The data cover the years 1975–2016, but we primarily focus on the years 1992–2016 since 4/20 is a recent phenomenon. Similar to prior work, we restricted our analysis to drivers, who are in control of the vehicle, rather than all passengers.9 Since prior work suggested an association between mass cannabis consumption at 16:20 hours and fatal crashes, we also restricted our analysis to the period from 16:20 to 23:59 hours.
The primary challenge in isolating the effect of any specific day on fatal crashes is noise in daily crash rates, which are affected by large-scale secular trends, economic conditions, safety legislation, road and vehicle engineering, seasonal weather patterns, weekends and holidays. Prior studies have used crash days in the same year 1 week before and after as control days.9 12 15 This usefully accounts for day of the week and potentially seasonal effects, but also requires accounting for the matched/clustered nature of the data (days clustered within a given year) to obtain valid SEs. We used several regression-based approaches to account for extra-Poisson variation in daily fatal crash rates16: negative binomial regression, rescaling SEs using the Pearson χ2 statistic, cluster-robust variance (clustered by year) and bootstrapping.
Although using control days 1 week before and 1 week after helps to adjust for potential confounding by day of the week and seasonal factors, the choice of 2 days a week apart is arbitrary. To assess robustness we also estimated the increase in potential fatal crashes on 4/20 relative to control days 1 and 2 weeks before and after, as well as to all other days of the year. For comparative purposes, we also conducted the same analyses for a day with established excess risk: US Independence Day (July 4th).10–12 To investigate heterogeneity in daily drivers involved in fatal crashes, we also repeated the same analysis for each day of the year in a separate negative binomial regression using two control days 1 week before and after each day. As a sensitivity analysis, we also repeated these analyses for the number of fatal crashes rather than the total number of drivers involved in fatal crashes.
Since public celebrations of 4/20 are a recent phenomenon, to see whether we could detect increasing excess numbers of fatal crashes over time that may be attributable to 4/20 we also fit a multilevel Poisson model using all of the Fatality Analysis Reporting System data from 1975 to 2016. We included random effects for each year as well as random slopes for 4/20 and a handful of other notable days: July 4th, the Saturday before Labor Day, the day before Thanksgiving, Christmas Day and New Year’s Day. We fit this model using weakly informative priors17 and used 4000 posterior samples to generate model predictions of the excess number of drivers involved in fatal crashes on each day, year by year. More details on the models are given in the online supplementary appendix 1. We used Stata (V.14), R (V.3.4.3) and Stan software18 to conduct our analysis. The data and code to replicate our findings are posted in a public repository (https://osf.io/qnrg6/).
Summary statistics showing average number of drivers involved in fatal crashes by day of the week, month and year over the entire sample are given in Online supplementary appendix tables 1-3. The number of drivers involved in fatal crashes between 16:20 and 23:59 hours has decreased from around 66 per day in 1975 to 56 in 2016. Figure 1 shows the estimates from our series of regression models for 4/20 and July 4th for the years 1992–2016. The first estimate in the upper panel shows a 12% increase in the number of drivers involved in fatal crashes on 20 April relative to control days 1 week before and after (IRR 1.12, 95% CI 1.04 to 1.19). The additional estimates in the upper panel demonstrate that the Poisson model underestimates uncertainty by ignoring extra-Poisson variation in daily crashes. The middle panel of figure 1 shows that comparing drivers in fatal crashes on 4/20 and July 4th to the number occurring on the same day of the week, 1 and 2 weeks before and after, attenuates the effect for 4/20 but not July 4th, and the bottom panel shows evidence consistent with null effects for 4/20 when using fatal crashes on all other days as the reference group. The number of drivers involved in fatal crashes increases by roughly 40% on July 4th, regardless of which control period was used. Estimates from each regression are provided in online supplementary appendix tables 4 and 5. Online supplementary appendix tables 6 and 7 show similar results when the outcome is the number of crashes rather than the number of drivers involved in crashes.
Figure 2 shows the estimates from separate negative binomial regressions for each day of the year relative to control days 1 week before and 1 week after, pooled over the period 1992–2016. Again, there is some evidence for elevated relative rates on 4/20 (shown in red), but the magnitude of the estimate is consistent with the bulk of daily variations for many other days. In online supplementary appendix table 8, we list 16 days with higher rates and 15 days with lower rates relative to the population average that are also inconsistent with the null. However, some more notable days stand out as particularly risky (July 4th, New Year’s Eve).
Figure 3 shows the average number of excess drivers involved in fatal crashes each year between 16:20 and 23:59 hours on 4/20 as well as several other notable days (averages across the entire period and 95% posterior intervals are given in online supplementary appendix table 9, and traceplots from the Bayesian multi-level models are given in online supplementary appendix figures 1-6). There is little evidence of any additional drivers involved in fatal crashes on 4/20 across the entire range of data (average annual excess crashes of −3.5, 95% posterior interval −8.0, 1.4), or in recent years. However, there is a consistent excess every year on July 4th and the days before Labor Day and Thanksgiving, as well as fewer drivers involved in fatal crashes on Christmas Day and New Year’s Day.
We found little evidence that the number of drivers involved in fatal traffic crashes is elevated on 20 April relative to any control period. This is not because daily crash rates are too noisy to detect any signal. On the contrary, we find important, systematic and meaningful variation in the daily number of drivers involved in fatal traffic crashes across the period from 1975 to 2016. We find consistent evidence for increases in the number of drivers involved in fatal crashes on 4 July and the days prior to Labor Day and American Thanksgiving, as well as systematically fewer drivers involved in fatal crashes on Christmas Day and New Year’s Day. Similar patterns have also been observed in Canada, with the riskiest periods being weekends before major holidays such as Canada Day, Thanksgiving and Victoria Day.19
With respect to prior studies, our results are consistent with existing work that has evaluated daily fluctuations in fatal crashes, including elevated risks on 4 July, weekends before Labor Day and Christmas Day, and New Year’s Eve.10–12 We were also able to reproduce the previously reported9 12% increase in the national rate of fatal crashes among drivers on 4/20. However, we found that this estimate was sensitive to methods for calculating its variance, as well as to how the control group was constructed, with diminished and largely null associations when crashes on 4/20 are compared with control days 1 and 2 weeks before or after, or to the rest of the days of the year. In contrast, the excess number of drivers in fatal crashes on July 4th was robust to these alternative control periods. We also found little variation in the annual impact of 4/20 over time. If recent celebrations of 4/20 were generating excess fatal crashes we would expect to see a greater excess in recent years.
To generate a reliable increase in the national number of drivers in fatal crashes on any given day requires either the presence of a substantially increased RR, population-wide exposure to a moderately increased risk, or some combination of both. In the case of alcohol, there is evidence that increasing impairment increases the RR of a fatal crash exponentially,20 as well as demonstrated increases in population-wide exposure to alcohol and drunk driving around major holidays. This likely generates the observed excess risk and is reflected in the higher proportion of fatal crashes involving alcohol on those days.10
Less is known about how widely 4/20 is celebrated, or about drugged driving more generally. Both American and Canadian reports suggest a similar prevalence (~12%) of past-year cannabis use among adults.21 22 Past month use in the USA is closer to 8%, and daily use nearer to 2.5%.23 Estimates of the proportion of the population that may engage in drugged driving are less common. Roadside tests in Canada indicate 4%–6% of drivers reported having used cannabis, and roughly 20% of the 12% of past-year cannabis users admitting to driving within 2 hours of cannabis use.24 Given the localised celebration of 4/20, the small fraction of individuals who report driving after using cannabis, as well as the moderate RR of a fatal crash for cannabis use,5 it seems unlikely that enough of the population is exposed to drugged driving on 4/20 to generate substantial population-wide excesses in fatal crashes.
Our results also have implications for other analyses of study designs focused on transient risks that have used the ‘double-pair control’ design.15 Conceptually, it seems reasonable to consider a set of control days 1 week apart from the index date since this provides good control for any day of the week or (less clearly) seasonal factors. However, because the data generating processes resulting in daily counts are noisy and tend to reflect both strong stochastic and deterministic components, failing to model or incorporate this variation leads to degraded inference.
Some limitations of this analysis should be noted. We do not measure changes in cannabis consumption on 4/20 or any other days, so clearly any estimates of the impact of 4/20 on fatal traffic crashes among drivers must be considered a misclassified estimate of the population impact of cannabis consumption on driving-related deaths. We compared the impact on 4/20 with that on 4 July and other days but those choices are subjective and it would be worthwhile to conduct more detailed investigations of systematic variations in daily fatal traffic crashes. In particular, for comparison with prior research we restricted our analysis to focus on the period from 16:00 to 11:59 hours to capture the potential for ‘mass consumption’ of cannabis to have an impact on 4/20. Because the period with the highest crash risk often overlaps 2 days (from 21:00 to 03:00 hours), using alternative hours could produce different risky days (eg, New Year’s Day was protective in our analysis but would likely be risky if the hours between midnight and 03:00 hours were included).
Population-based evidence on both the prevalence and the RR of marijuana-impaired driving is weak, so increased efforts to understand the potential impact that decriminalising or legalising marijuana may have on driving risks would be welcome. We conclude that there is limited evidence of any population-wide effects of the annual cannabis holiday on fatal traffic crashes. The policy landscape for addressing marijuana-impaired driving is rapidly changing, and efforts to combat drugged driving are likely to require new initiatives. Given the considerable resources already devoted to heightened enforcement campaigns to reduce the impact of impaired driving,25 there does not seem to be sufficient evidence for including 4/20 among ‘risky’ days.
What is already known on the subject
Cannabis use is associated with moderate excess MVC risk.
There are reported increases in ‘drugged driving’.
Prior work has suggested an increase in crashes on 20 April, the annual ‘cannabis holiday’.
What this study adds
Comparison of cannabis holiday with other known risky days for MVCs.
Evaluation of time trends in the impact of the cannabis holiday on drivers involved in fatal crashes.
There is little evidence of excess numbers of drivers involved in fatal crashes on 20 April relative to normal daily variations.
Contributors SH and AP conceptualised and designed the study. SH acquired the data and conducted the analysis. SH and AP interpreted the findings and wrote the paper. Both authors read and approved the final version of the manuscript.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Patient consent for publication Not required.
Ethics approval This study used publicly available data and did not require ethics review.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement The data and code to replicate our findings are posted in a public repository (https://osf.io/qnrg6/).
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.