Marijuana has been legal in Canada and other jurisdictions in North America for a few years now. There has been some concern that this would lead to increased usage, and consequently a higher level of auto insurance accident claims. A new research project, jointly sponsored by the CIA and Casualty Actuarial Society, has an answer to that question, and author Dr. Vyacheslav Lyubchich joins us on this episode to discuss the findings. (In English)

Welcome to Seeing Beyond Risk, a podcast series from the Canadian Institute of Actuaries. I’m Ping-Teng Lin, your host for today’s episode.

Since 2018, the Government of Canada has legalized the production, distribution and possession of cannabis for nonmedical purposes.

One of the questions since then has been if the impairment effects of marijuana would lead to riskier driving behaviour. There have been several observational research studies on the topic, but with mixed results.

I have here joining us today, Vyacheslav (or Slava) Lyubchich, who’s an associate professor at the University of Maryland. He’s very experienced in applying statistical and machine learning techniques to insurance topics of interest, such as climate change.

And today, he’s here to share the findings from his recent paper, jointly requested by the CIA and the CAS, on assessing the impact of marijuana decriminalization on vehicle accident experience.

Slava, thank you very much for joining us today.

Thank you, Ping.

Why don’t we start off by discussing what were the motivations behind this research project and what were the objectives that you were hoping to achieve?

Sure. This project addresses a request by the CIA and CAS for analysis of the impact of marijuana decriminalization on vehicle accident experience.

And as you said, there have been mixed reports about the implications of decriminalization of marijuana.

So the goal of this project was to conduct a new study with more comprehensive datasets, methods and especially with more observations after actual decriminalization.

The results of this project help insurance companies assess the associated changes, price the risk and, if needed, adjust the cost of insurance.

So, what is the current state of marijuana decriminalization in Canada and the US, and how did that affect your data collection process?

Bill C-45 legalized access to recreational cannabis in Canada starting on October 17, 2018. It happened on the same date in the whole country, so the data collected for Canada represent only the periods before and after the bill came into effect.

In the United States, some states have decriminalized the use of marijuana, while others have not. It happened at different times for each state, and to a different extent.

For example, some states fully legalized marijuana for medicinal and recreational use, such as California and Nevada. Some legalized only medicinal marijuana, and there are states in which the use of marijuana is mostly illegal, such as South Carolina and Wyoming.

With a great diversity of this situation across the state, data in the United States could be collected not just for the periods before and after, but also for concurrent observations in a state that did not change its legalization regarding marijuana during the study period.

Also, because each of the states is relatively small compared to a country like Canada, it was possible to relate other information, like weather patterns in the state, to the number of car accidents.

Very interesting. And what were the major findings from your research about the impact of marijuana on vehicle accident rates, and how would they compare to the existing research that’s already out there?

First of all, I’d like to mention that researchers often find that marijuana use affects a person’s ability to drive. These are typically controlled studies that show that the participating volunteers who smoked marijuana drive slower and maintain longer following distances.

Possibly it happens because the drivers are aware of the impairment and are more cautious.

If we look at aggregate results such as the accident rates observed in a state or a country, the reported results are quite mixed. Most studies have been done in the United States, and most often they do not find statistically significant effects of marijuana decriminalization.

The sources of this variability in the results include data periods, locations, aggregation and modelling methods applied in the studies.

For example, some studies aggregate data annually, semi-annually or monthly, collect data from cities or whole states, and consider different additional variables in the analysis, such as the general time trends, age and race of drivers.

With data separated into smaller beans, such as smaller time intervals or specific driver demographics, some studies were able to detect some short-term or demographic-specific effects of marijuana decriminalization. But again, larger-scale studies rarely detect such effects.

Same in this study, I did not detect a statistically significant change in the car accident fatality rate, insurance claim frequency or average cost per claim. The estimated statewide effects of decriminalization in the United States were mostly not statistically significant, and they also varied across states.

The estimated seasonality and dynamics in Canadian vehicle insurance statistics allowed me to accurately predict post-legalization dynamics, so suggesting the effect of decriminalization was overall negligible.

And I should also say that temporal patterns of human activities, such as yearly, weekly and daily cycles, are among the top predictors in these models for the vehicle accident experience.

And during your process, how were you able to actually separate out the impact of marijuana use on accident rates versus all of the other potential contributing factors?

The separation of different effects was a tough problem when using Canadian data because they were collected over whole provinces over a year or, in the case of Quebec, we had quarterly data.

On such a big scale, it was hard to include weather or even holiday patterns into consideration. So the implemented models included combined effects of unaccounted factors as a time trend without modelling effects of each factor.

For the United States, more granular data in space and time were available, so I could combine daily information from an airport with daily accident rates near that airport. Additionally, the models considered the effects of weekdays and holidays based on the holiday calendar.

And the partial dependence plots from the models for the United States then showed the marginal effects of each variable.

It sounds like you use quite a few different statistical techniques in your study, so could you provide an overview of some of these methods?

Yes, this study used many statistical methods to achieve robust conclusions.

For the analysis of cost per claim and frequency of collisions in Canada, I used mixed-effects models with the time trend. These models accounted for different average levels of the response variable by

province.

For inference from these models, confidence intervals for the effect of decriminalization were obtained not just with the typical statistical assumption of normally distributed data, but also using nonparametric bootstrap approaches.

Bootstrap is a method of resampling with replacement that provides data-driven results without restrictive assumptions of a certain distribution.

With the United States, several methods were used as well. First of all, to select concurrent controls for each state that legalized marijuana, the method of propensity score matching was used.

This method attempts to create homogeneous groups to potentially avoid confounding and select the states for comparisons.

For example, in this study, the following state statistics were used to select the control states: urbanization, population, number of vehicle miles per licensed driver, number of road miles and the number of vehicles registered per thousand people.

After the selection of states for the comparison, the machine learning method called random forest was implemented.

Random forest is a decision-tree-based method that delivers competitive performance compared to other models and techniques. At the same time, it has only a few parameters that the user has to set.

It also uses bootstrapping and resampling to obtain robust results and study the effects of different covariates. So I would say the main methods used in this study included mixed-effects models and random forest.

So what do you hope researchers, actuaries and insurance companies would be able to take away from this study? And where do you think the future areas of this research would lie?

I hope that researchers and actuaries will be more aware of the problems and inconsistencies that can arise when collecting and analyzing observational data, but these inconsistencies from the previous studies should not discourage researchers.

Instead, they encourage us to continue the research and look for the next possible solutions, and develop, adapt and deploy new methods that will probably help answer new questions or solve problems in other domains.

I hope insurance companies take an interest in this process and support it in many possible ways, which include data sharing such as it is done in Canada, provide feedback to researchers on the industry needs, and further support the development of statistical methods that allow the analysts to draw conclusions using observational data.

Well, thank you very much for joining us today, Slava, to talk about your recent research paper.

Thank you, Ping. My pleasure.

We now have over 150 episodes in our podcast series going back over the past three years, so we encourage you all to subscribe, and you can do that through whatever platform you use to get your podcasts.

And we’d like to hear from you. So, if you have any suggestions or episode ideas, you can send them to podcasts@cia-ica.ca.

And we’ve recently launched a new site covering all the exciting changes we are making to the CIA qualification requirements. You can find out more at https://education.cia-ica.ca/.

Until next time. I’m Ping-Teng Lin and thank you for tuning in to this episode of the Seeing Beyond Risk podcast.