This originally appeared at https://pubmed.ncbi.nlm.nih.gov/37139565/
José Ignacio Nazif-Munoz 1 , Karen A Domínguez-Cancino 1 , Marie Claude Ouimet 1 , Thomas G Brown 1
- PMID: 37139565
- DOI: 10.1111/dar.13678
- First published: 03 May 2023
Abstract
Introduction
In the past decade, a group of studies has begun to explore the association between cannabis recreational use policies and traffic crashes. After these policies are set in place, several factors may affect cannabis consumption, including the number of cannabis stores (NCS) per capita. This study examines the association between the enactment of Canada’s Cannabis Act (CCA) (18 October 2018) and the NCS (allowed to function from 1 April 2019) with traffic injuries in Toronto.
Methods
We explored the association of the CCA and the NCS with traffic crashes. We applied two methods: hybrid difference-in-difference (DID) and hybrid-fuzzy DID. We used generalised linear models using CCA and the NCS per capita as the main variables of interest. We adjusted for precipitation, temperature and snow. Information is gathered from Toronto Police Service, Alcohol and Gaming Commission of Ontario, and Environment Canada. The period of analysis was from 1 January 2016 to 31 December 2019.
Results
Regardless of the outcome, neither the CCA nor the NCS is associated with concomitant changes in the outcomes. In hybrid DID models, the CCA is associated with non-significant decreases of 9% (incidence rate ratio 0.91, 95% confidence interval 0.74,1.11) in traffic crashes and in the hybrid-fuzzy DID models, the NCS are associated with nonsignificant decreases of 3% (95% confidence interval − 9%, 4%) in the same outcome.
Discussion and Conclusions
This study observes that more research is needed to better understand the short-term effects (April to December 2019) of NCS in Toronto on road safety outcomes.
1 INTRODUCTION
The effects of cannabis use in motor-vehicle driving are well documented [1, 2]. Cannabis consumption is associated with short-term effects of euphoria and relaxation, time distortion, increased distraction and changes in perceptual–motor coordination and motor performance [3]. Nevertheless, the association between cannabis use and traffic crashes variation is unclear. A meta-analysis by Asbridge et al. [4] found that acute cannabis consumption is associated with doubled probabilities of being involved in a fatal crash. In contrast, another review of the same studies [5] obtained a lower risk, with an increase of 24% in the probability of a fatal collision.
The consequences of cannabis recreational use laws (CRUL) for traffic outcomes are also mixed. US studies have indicated increases in traffic fatalities by 2.1% following the implementation of CRULs in Colorado, Washington, Oregon and Alaska [6]. Studies focusing on cannabis-involved fatal crashes have also found significant increases following the implementation of CRULs, including a 63% increase in Colorado [7]. These changes, according to Aydelotte et al. [8], may be partially explained by the opening and consolidation of commercial stores. Similarly in Uruguay, although with a different transport system than the USA, research has pointed out that CRUL may be associated with increased traffic fatalities in urban settings when access to commercial stores increases over time [9]. Specifically, in Montevideo, an increase of 6% in driver fatality rates was observed following the implementation of its CRUL. At the same time, inconclusive changes have been reported in the US and Canada. In Oregon and Washington, 3 years after their respective CRULs were passed, no changes in traffic fatalities were observed compared to non-CRUL states where road traffic pre-conditions were similar [10, 11]. In the Canadian provinces of Ontario and Alberta, enactment of CRULs was not found to be associated with significant changes in traffic injuries [12].
In Canada, on 17 October 2018, Canada’s Cannabis Act (CCA) was enacted [13]. With the CCA, authorisations, prohibition and offences of cannabis use were legalised, but provincial and territorial jurisdictions had autonomy to regulate the sale and distribution of recreational cannabis products. This Act thus allows persons aged 18 years or older (depending on provincial regulations) to possess and/or consume cannabis, and households are allowed to grow cannabis plants following specific restrictions established by each province. For instance, while in Ontario the minimum age to legally consume and buy cannabis for recreational purposes is 19, in Quebec the minimum age is 21. Within this Act, cannabis stores are also allowed following specific regulations determined by each province or territory autonomously. In terms of sales model, Ontario allows for private and public sellers, whereas in Quebec only public is allowed. In Ontario, a second important event was the authorisation of cannabis stores that should follow specific regulations to properly sell cannabis. More specifically, on 1 April 2019, this province regulates that cannabis retailers could sell products once they were properly registered. After the CCA, between November and December 2018, there were a total of 1129 legal (registered in official government websites) and illegal (registered in Weedmpas—a platform used to facilitate illegal commerce between providers and consumers) cannabis retailers across Canada. In that period, Ontario was the Canadian jurisdiction with the highest number of commercial dispensaries with 385, of which 84 were storefronts and 301 delivery-only services [14], and these stores could only legally sell cannabis products after 1 April 2019.
Currently, the recently available evidence from the USA, Uruguay and Canada does not support a clear stance concerning the impact of CRUL on traffic-related outcomes at the ecological level, particularly with studies which focus on traffic outcomes without explicitly considering the presence of cannabis in road users [9, 15–17]. More work is thus required to understand how CRUL might affect different jurisdictions, particularly where restrictions vary over time, and in jurisdictions where drug-impaired driving may have been increased [18]. This is particularly pressing in Canada since 13.2% of reported surveyed people who have declared cannabis use in the previous 3 months have driven within 2 h of consuming cannabis. The current study aims to assess the impact of CRUL by evaluating its association with traffic-crashes, road victims and severe injuries in Toronto.
We purposefully focus on Toronto because it comprises an urban setting in which access to cannabis has varied over time and where traffic crash trends for fatalities, injuries and collisions from 2016 to 2018 have remained stable [19]. In this city, in December 2017, there were 70 illegal retailers registered (storefronts and storefront with delivery services) [20], whereas in December 2018, this number reached 168. Recent studies have also suggested that cannabis consumption is diverse within Toronto. Cannabis past 12-month consumption for ages 18–34 is about 45% in Old Toronto, Toronto-East York and York, between 35% and 40% in North York and Scarborough, and lower than 29% in Etobicoke [21]. We expect that following the CCA, a change in the opportunity to consume cannabis (and then to drive under its influence) was associated with increases in traffic crashes and number of road victims, including type of injuries.
2 METHODS
Study design
We evaluated the association of the CCA in Toronto applying a hybrid-fuzzy difference-in-difference (DID) study design, using three outcomes with a population per district as denominator: (i) all crashes; (ii) all road users’ victims (pedestrians and motor-vehicle occupants involved in a crash); and (iii) all road users killed or severely injured (KSI). Thus, the unit of analysis was daily number of crashes, victims or KSI per 1 M population. Alternatively, we also used two other denominators: number of vehicles and number of trips per district. Results with these two denominators are available in Table S1 (Supporting Information).
We applied this quasi-experimental design DID as randomisation at the individual level was not possible. DID study designs have been used in road safety and other policy intervention studies [22, 23]. Traditionally, DID has four distinguishing characteristics: first, there are two distinctive groups, the intervention group and the control group; second, the former is exposed to a policy, whereas the latter receives no policy; third, before a given policy, no significant differences between control and intervention groups should be observed (parallel trend assumption); and last, the same time and date are used when comparing both groups, that is, the date in which the intervention group is exposed to the policy, is also used to define that the control group was not exposed to it. To properly apply a traditional DID approach, at least one city in Canada should have been fully unexposed to the CCA, and the intervention should have been homogenous over time. Thus, to tackle the challenge associated with the absence of a proper control group (a territory without cannabis stores at the time the CCA was enacted), required in a conventional DID approach, we applied hybrid DID study design [28–30]. Whereby it assumed the intervention was homogenous over time (pseudo intervention group), and a control group (pseudo-control group) is created by considering an exact duration of time where the intervention was inexistent. This allows us to extract the difference between the pseudo-control and intervention groups and estimate the association between the intervention and the selected outcomes.
Population
Our analyses included the entire city of Toronto (population 2,928,017) and individual analysis for four districts of Toronto, as per defined by the Police of Toronto: (i) Old Toronto, Toronto-East York (population 865,149); (ii) North York (population 644,685); (iii) Scarborough (population 631,890); and (iv) Etobicoke (population 589,860). This city represents 19.7% of Ontario.
Dependent variable
We obtained data on all crashes registered from the Toronto Police Service Public Safety Data Portal [24]. To calculate the dependent variables, we used population estimates provided by the City of Toronto Portal [25] as offset and created: (i) crashes; (ii) road victims; and (iii) KSI per 1 M population. The same procedure was adopted when using number of vehicles and number of trips, which were also obtained from the City of Toronto Portal [30].
Control variables
Since daily crashes in urban centres are highly correlated with meteorological variables [26], we constructed a series of variables representing climate variation. We accessed weather information through the Federal Government of Canada portal [27]. We selected the stations located in Toronto that contained daily information for the period of analysis. The following variables were created: (i) daily mean temperature (°C), defined as the average of the maximum and minimum temperature, in degrees Celsius, at a location for a specified time interval; (ii) daily total rain (mm), defined as the total rainfall, or amount of all liquid precipitation in millimetres (mm) such as rain, drizzle, freezing rain and hail, observed at the location during a specified time interval; and (iii) daily total snow (cm), defined as the total snowfall or amount of frozen (solid) precipitation in centimetres (cm), such as snow and ice pellets, observed at the location during a specified time interval.
Interventions
We used two different points of time to build the intervention. First, the day in which the CCA was enacted (17 October 2018), and second, the day in which Ontario allowed legal cannabis stores to sell cannabis for recreational purposes (1 April 2019). With these dates, we created two intervention variables, one in which the law was for the very first time enacted and a second one which considered the number of stores over time per population for Ontario and the districts respectively. To obtain the number of stores we systematically called each registered cannabis legal store (N = 289), from 2 January to 31 March 2022, and requested information regarding its date of opening. A total of 193 stores provided information, eight of which initiated their sales between 17 October 2018 and 31 December 2019. We included stores which included onsite sale only or onsite sale and delivery options.
Period of analysis, parallel trend assumption and model selection
The period of analysis is from 1 January 2016 to 31 December 2019. This corresponds to a total of 1460 days, allowing us to have proper statistical power to carry out this approach. More specifically by considering more than 100 units of time, it is possible to detect seasonal and abrupt changes in a given trend that may affect the outcomes after an intervention has occurred in the series [28, 29]. We restricted our analyses to a period in which the COVID-19 pandemic did not affect travel patterns within this city [17]; otherwise, other methods should have been applied to accommodate these substantial changes. To explore the parallel trend assumption, we visualised differences across pseudo intervention and control groups using as reference 1 April (see Figures S2). To determine each model per outcome, several controls, including variables representing trends, seasonality and weather, were introduced. The best models were those whose Bayesian information criterion was the lowest and where error terms showed signs of not being autocorrelated.
Poisson regression
Due to the non-negative integer nature of road traffic crash count data, we used generalised linear models, particularly Poisson regression. These models do not assume a linear relation between outcome and independent variables [30]. Furthermore, this particular approach allows to represent an approximation of underlying motor vehicle crash processes, in which it may be expected multiple zeros [31]. Complementary, we also used negative binomial regression and fixed effects Poisson models as these can account for different Poisson distributions across districts in Toronto, and the latter more specifically can deal with the ‘excess’ of zeros that may violate the Poisson distributional assumption [32] (see Tables S36–S41 for Negative binomial outcomes and Table S42 for fixed effects Poisson models). We applied two different statistical designs: (i) hybrid DID; and (ii) hybrid-fuzzy DID. Equations representing each design are in the Supporting Information. All analyses were carried out with STATA 17, and the functions were Poisson and Negative binomial. We also applied interrupted time series and results are available in the Supporting Information (Figure S1 and Tables S3–S20).
2.7.1 Hybrid DID design
We applied a hybrid DID design [33], whereby the intervention and control groups correspond to two different periods. The pseudo intervention group is the number of traffic crashes and injuries that occurred between 1 January 2018 and 31 December 2019 (a total of 730 days). Within this period, the date in which Ontario allowed cannabis to be legal was 17 October 2018, and the date in which stores were allowed to sell cannabis for recreational purposes was 1 April 2019. The pseudo–control group is the number of crashes and injuries that occurred between 1 January 2016 and 31 December 2017 (a total of 730 days), in which cannabis stores could not sell cannabis for recreational purposes. For this, we assumed that the pseudo-control group accurately represents the trend in rates of traffic crashes between 1 January 2016 and 17 October 2016, if, counter to the fact, cannabis for recreational purpose was not legal in the same period. The pseudo-control group was also assumed to accurately represent the trend of traffic crashes between 1 January 2016 and 1 April 2017 if cannabis legal stores were not functioning during that period.
2.7.2 Hybrid-fuzzy DID design
An extension of the hybrid DID approach is the hybrid-fuzzy DID [34] version, in which an increase in the intervention could be higher in some groups than in others. For instance, the enactment of the CCA could have resulted in proportionally more stores in one district than in others. We thus applied a hybrid-fuzzy DID design. Because the intervention is not randomly assigned, two potential sets of unobserved variables may bias the hybrid-fuzzy DID estimator: (i) time-varying factors which affect each group differently such as the presence of heat waves or snowstorms in the intervention period [35]; and (ii) group attributes may change over time, for instance, in the intervention period group relative to the control period, such as an increase or decrease of alcohol consumption which in turn may be related to changes in traffic outcomes.
If we observe n crashes in two different periods (P = 0.X), zero signals the period in which the CCA was not observed—the preintervention—and X indicates the beginning of the CCA—the post-intervention. Under the hybrid-fuzzy design, the X may take different values. Under this study, it is the number of cannabis’ stores per capita which, after the introduction of the CCA, changed over time.
In Figure 1, we depict dates with relevant events, identification of pseudo intervention and control groups, number of days, outcomes and methods applied.
Incidence rate ratios
In all tables and figures with results of the analyses, we report incidence rate ratio (IRR). This outcome estimates the change in crash, road victims and KSI rates in the post-intervention period to the pre-intervention period for the pseudo intervention group, adjusted for the change in crash, road victims and KSI rates from the post-intervention period to the pre-intervention period in the pseudo-control group, and the control variables described above.
3 RESULTS
In Table 1, we observe the daily average of each rate according to jurisdiction, pseudo-control and pseudo intervention groups, and pre- and post-intervention periods, using 1 April as the main reference. In parenthesis, we included minimum and maximum values of each rate. In “All Toronto” and “North York,” regardless of each rate, averages in the pseudo-control group are slightly higher in the post-intervention period relative to any other period. In Old Toronto, Toronto East York and York, we observe heterogenous patterns. Whereas averages of crash rates and KSI are lower in the pseudo intervention group in both periods relative to the pseudo-control group rates, we observe a higher increment of KSI in the post-intervention period for the pseudo intervention group when this average is compared to the pre-intervention period, and an inverse pattern is observed for the pseudo-control group. In Scarborough, we observe that the post-intervention period for the pseudo intervention group present higher averages in each outcome. Lastly, in Etobicoke, we observe that averages of crashes are stable, whereas in the post-intervention period for the pseudo-control group road victims’ and KSI’s rates are lower relative to any other period. TABLE 1. Crashes, road victims and killed or severely injured per 1,000,000 population according to jurisdiction, pseudo-control and pseudo-intervention groups, and pre- and post-intervention periods (1 January 2016–31 December 2019).
| Jursidiction | Outcome | Pseudo-control group | Pseudo-intervention group | ||
|---|---|---|---|---|---|
| Pre- intervention period | Post-intervention period | Pre- intervention period | Post-intervention period | ||
| 1 January 2016 – 30 March 2017 | 1 April 2017 – 31 December 2017 | 1 January 2018 – 30 March 2019 | 1 April 2019 – 31 December 2019 | ||
| All Toronto | Crashes | 0.37 (0–2.19) | 0.41 (0–1.81) | 0.38 (0–2.15) | 0.38 (0–1.78) |
| Road victims | 2.17 (0–16.84) | 2.23 (0–11.61) | 2.18 (0–12.23) | 2.17 (0–14.27) | |
| KSI | 0.91 (0–6.58) | 0.96 (0–3.62) | 0.93 (0–6.47) | 0.93 (0–5.35) | |
| Old Toronto, Toronto-East York and York | Crashes | 0.37 (0–5.77) | 0.42 (0–3.40) | 0.32 (0–3.35) | 0.31 (0–3.35) |
| Road victims | 2.23 (0–53.17) | 2.05 (0–25.04) | 1.71 (0–23.47) | 1.91 (0–30.79) | |
| KSI | 0.93 (0–15.02) | 0.92 (0–9.09) | 0.77 (0–11.17) | 0.79 (0–14.29) | |
| North York | Crashes | 0.30 (0–3.10) | 0.39 (0–4.61) | 0.39 (0–3.05) | 0.34 (0–3.02) |
| Road victims | 1.77 (0–62.04) | 2.46 (0–50.77) | 2.17 (0–36.62) | 1.86 (0–27.25) | |
| KSI | 0.72 (0–15.51) | 0.93 (0–12.30) | 0.92 (0–13.73) | 0.83 (0–10.59) | |
| Scarborough | Crashes | 0.36 (0–3.16) | 0.41 (0–4.73) | 0.40 (0–4.72) | 0.45 (0–3.14) |
| Road victims | 2.17 (0–31.58) | 2.25 (0–31.58) | 2.33 (0–34.67) | 2.96 (0–42.47) | |
| KSI | 0.89 (0–12.63) | 0.95 (0–12.63) | 1.01 (0–17.33) | 1.16 (0–14.15) | |
| Etobicoke | Crashes | 0.42 (0–6.78) | 0.43 (0–5.08) | 0.45 (0–5.08) | 0.44 (0–5.08) |
| Road victims | 2.54 (0–57.64) | 2.43 (0–35.60) | 2.71 (0–49.16) | 2.38 (0–27.12) | |
| KSI | 1.13 (0–25.42) | 1.04 (0–13.56) | 1.09 (0–20.34) | 1.09 (0–13 56) | |
- Note: In parenthesis are displayed minimum and maximum rates. Estimates are pooled at the day level.
- Abbreviation: KSI, killed or severely injured.
In Figure 2, we observe the cumulative number of cannabis stores (NCS) per territory from 18 October 2018 to 31 December 2021, using our survey and quarterly reports from the Ontario Cannabis Retail Corporation (information from the Ontario Cannabis Retail Corporation for each district of Toronto is only available from January 2022; information for the complete city of Toronto is available from April 2020, every 4 months). Differences between our survey and the Ontario Cannabis Retail Corporation’s reports are explained because some stores did not provide us with information regarding the date of their launching. The overall pattern is pulled by the district Old Toronto, Toronto East York and York, where commensurate increments begin in January–April 2020, after the Ontario Regulation 468/18 was repealed [36]. For the period of analysis restricted to this study, only this district registers an important change: from October 2018 up until December 2019 seven cannabis stores were observed (in average one cannabis store was opened every 2 months). However, from January 2020 to December 2021 three stores per month were opened. In North York and Scarborough, an increase is only observed in January 2021, and Etobicoke is the only district in which the NCS is low and relatively flat.
Hybrid DID design results
In Figure 3, we report results regarding the hybrid DID design. While most IRRs in both intervention periods analysed are null, two exceptions are observed. First, North York has reductions in both intervention periods for the rate of all victims (IRR = 0.66 95% CI 0.56–0.79, October 2018) (IRR = 0.64 95% CI 0.52–0.78, April 2019). Second, in the same outcome for Scarborough, when the intervention begins in October 2018, we observe an increment of 33% (95% CI 13%–57%) in all victims. When the intervention begins in April 2019, Scarborough has the highest IRR for the same rate with a nonsignificant increase of 13% (95% CI -6%–35%). In terms of IRRs for the Old Toronto, Toronto East York and York district only all victims’ rates are slightly higher than 1, yet their associations are non-significant. When we compared IRRs within districts after changing the intervention period, we observe non-significant differences (see Tables S21–S23 for the October time of implementation and S24–S32 for April time of implementation). The control variables applied in each selected model are described in Tables S21–S32.
Hybrid-fuzzy DID design results
In Figure 4, we display the IRR for the analyses in which cannabis stores per population are present in all Toronto city and Old Toronto, Toronto East York and York districts. Because we registered the first cannabis stores on 18 October 2018, this analysis considers an intervention period which begins in this date, however, the value of the intervention changes over time. The other three districts are not included in the analysis because no changes were registered in the NCS throughout the post intervention period (18 October 2018–31 December 2019). Regardless of the outcome, the intervention either in all Toronto or in the selected district is not associated with any concomitant change. The main difference with the other approaches is the magnitude of the confidence intervals which are much lower suggesting a stabler approach to measure the association between the intervention and the selected outcomes. The controls variables applied in each selected model are described in Tables S33–S35.
4 DISCUSSION
While the literature focusing on cannabis and motor-vehicle driving skills is relatively consistent in observing an association between consumption, impairment and crashes [37, 38], an emerging group of studies probing the policies of cannabis consumption for recreational purposes and road safety outcomes at the macro level remains less conclusive [39, 40]. Our study contributes to this literature by examining two ways of operationalising changes in CRUL when it is implemented in a limited urban area, but with specific attention to changes in the NCS per capita and by analysis that zooms in on different districts within the city of Toronto. During the first year of the CRUL’s implementation in Toronto, no significant changes in crashes, number of road victims and KSI were observed. Even further, in Old Toronto, Toronto-East York and York, null effects were observed systematically, regardless of the method, operationalisation of the policy and the outcomes used.
To some extent, the observed lack of variability in the estimators, except for North York, when analyses considered a more restricted time of implementation, and for Scarborough when attention was put on one rate, namely all victims, may be due to: (i) a limited time in which the CRUL was observed; (ii) low NCS per capita; (iii) the absence of information regarding whether drivers involved in the crashes had been exposed to cannabis; or (iv) a change in a trend decline associated with road safety policies introduced in Toronto [41], which may have been inverted with the CRUL. Nevertheless, our results are consistent with Santaella-Tenorio et al. [42], in which after 1 year of implementation of a cannabis law in Washington state, no changes were observed for traffic fatalities. Importantly, their null findings were associated with much higher NCS per capita than the ones observed for Toronto. Whereas in Washington state, the rate was 7.37 per 100,000 population, in our analysis this rate was 0.43, even if when we considered this rate growing over time. When we restrict our analysis to Old Toronto, Toronto-East York and York this rate is slightly higher, 0.88, but still much lower than in Washington state.
In terms of the outcomes, while ideally, one should analyse collision data for drivers in light of information on level of tetrahydrocannabinol in the bloodstream as suggested by Vingilis et al. [39], our consideration of three different outcomes (crashes, all road victims and KSI) responds partially to this limitation. Indeed, our approach attempted to create better conditions to identify potential changes that should have been detected consistently if in average drivers involved in crashes had increased their exposure to cannabis consumption. Relatedly, lack of power, which could have been a concern given by the low number of the observed outcomes (for instance, 0.38 is the daily crash average for the pre-intervention period per 1,000,000 population) was not necessarily a limitation per se. Thus, an increase of 0.50 in the daily crash average should have been observed in the intervention period for a statistical power of 80% and an alpha of 0.05 [43]. In other words, to find an association in the intervention period with these two conditions an increase from 326 to 398 crashes should have been observed. However, the information from Toronto’s police dataset did not lack power per se, rather in the intervention period an increment relative to the pre-period was not detected. Furthermore, under the assumption that 9% of crashes could be cannabis-related in Canada [44], we should have identified 29 crashes with these characteristics in the baseline period, and to observe a significant change with 80% of statistical power at an alpha of 0.05, an increase to 55 crashes should have been observed. It is also important to notice however that our crashes’ outcomes may also be subject of reporting biases since the police in Toronto may not be consistent in reporting these events, particularly when property damage may have been observed. Interestingly, one could have expected individuals under the influence of cannabis (particularly after the enactment of the law) to be less inclined to report a crash and therefore a decline in this outcome would have been captured. However, we did not observe a significant change in this outcome. Similarly, regardless of differences in how the cannabis act was operationalised, outcomes are consistently inconclusive, suggesting that this study provides no evidence to reject the null hypothesis.
While increments could have been expected in the district of Old Toronto, Toronto-East York and York, only North York and Scarborough presented significant negative and positive changes, respectively (Figure 3). These results deserve some clarification. First, the negative direction of the estimator of North York may be explained by a sudden increment in the post-implementation period associated with the pseudo-control group as we can observe in Table 1, which ultimately could have introduced biases associated with time-variant unobserved factors such as a sudden increment in employment registered in that period. In other words, this reduction is more likely explained by a factor of this characteristic rather than because the implementation of the CCA in this district was associated with, for example, a preventive campaign. Second, the increment in Scarborough by 33% in all victims, is only observed when the post-intervention period begins on 18 October 2018, nevertheless, between this date and 1 April 2019, no evidence of legal cannabis stores was found, further indicating that our suggested mechanism of policy implementation should not be associated with this increment, unless we assume the presence of a spillover effect derived from changes occurred in Old Toronto, Toronto-East York and York. However, as we suggested, a change should have been observed in the three outcomes, not in one.
Our study has several limitations. Although the use of different denominators can improve the identification of a counterfactual, as we did by including number of vehicles and number of trips—however, these denominators may not capture road exposure homogeneously since drivers for instance can own more than one vehicle, absence in other policies or factors in post-implementation periods (i.e., a strong and sustained enforcement plan) however could partially explained for instance results in Scarborough, and further increase the precision of our estimators for the other districts. Second, throughout the post-intervention period—although adequate in terms of duration, when using day as unit of analysis to carry out the three analyses—a concomitant increase in the number of legal cannabis stores per capita was not observed with out method, therefore, overall changes in cannabis consumption at the population level may not have been simply captured. This observation leaves open, however, the development of future analysis in which the implementation of the CCA, at least in Ontario considers not only lengthier times of study but rather higher thresholds of cannabis stores per capita. Thus, while in Toronto the CCA may not be regarded as having a sudden and permanent change in road safety outcomes, at least during its first year of implementation, the hypothesis of increment of crashes associated with NCS should be revisited from 2020 onwards, year in which there is a considerable growth of these stores. Indeed, as we observe in Toronto, particularly for Old Toronto, Toronto-East York and York, this rate is currently slightly higher than in the state of Colorado (24.32 vs. 23.32) [42]. Third, while we were able to capture changes in the number of legal stores, the presence of illegal stores was not considered. This could also explain an absence of change in the outcomes since the number of cannabis illegal stores may have been relatively the same for the period of analysis. Relatedly our survey may not have captured stores that had opened after the dates and closed before we initiate our data collection process. Fourth, the inclusion of the hybrid DID approach accounts for the possibility that a change in the exposure should have occurred regardless of the number of stores that existed at the time. However, there may have been unobserved changes in the pseudo control during the post-intervention period, which may have coincided with the direction of the intervention, reducing the possibility of finding an association. Fifth, lack of more frequent data on cannabis consumption at the district level, precluded us of testing if cannabis store changes were associated with traffic outcomes, assuming population level changes mediate the implementation of CCA in Toronto and traffic outcomes. Sixth, we did not examine the association of CCA with cannabis-related deaths or injuries; because Toronto police data do not indicate whether the driver was driving while impaired by cannabis when the event occurred. Last, other important variables such as alcohol outlets should have been considered. This would have helped to better adjust the studied outcomes.
Findings of this study suggest that small variations in the implementation of density of recreational cannabis stores with low thresholds of acceptance may be an intervention to explore more consistently, particularly for those jurisdictions interested in either introducing CRUL, or for those planning to augment the provision of cannabis via the existence of more stores. More research is however needed to identify both the speed in which increments of cannabis stores should be authorised and the maximum thresholds of accepted cannabis stores per capita, as well as how the presence of illegal stores could also play a role in the distribution of this type of outcomes.
