The association between DUIC and the risk of road traffic events was intensely studied before (Williams et al., 1985; Ramaekers et al., 2004; Mura et al., 2006; Turner, 2007; Romano and Voas, 2011; Pickett et al., 2012; Poulsen et al., 2012; Li et al., 2013; Rossheim et al., 2014; Terry-McElrath et al., 2014; Urfer et al., 2014). There are several methods of assessing this association – from experimental studies aimed to measure the influence of THC on driving performance, to epidemiological, culpability, or case–control studies. Although each add pieces of information, a definite answer is difficult to obtain, as these studies often had conflicting results and the research methodology was regularly prone to biases. Due to these reasons, a series of systematic reviews (e.g., Ramaekers et al., 2004) or meta-analysis (Asbridge et al., 2012; Li et al., 2012) were recently published, which strongly suggested an association between DUIC and UTEs.

When reporting a meta-analysis carried out using the random-effects model, the researchers usually present the summary effect size and its CI. These values allow us to estimate the mean effect size and precision but not the distribution of the true effects around the summary effect (Borenstein, 2009). Therefore, when exhibiting the result of a random-effect meta-analysis, we should detail three pieces of information: mean effect size, CI, and PI. The PI shows the distribution of the true effect sizes around the mean. Both meta-analyses (Asbridge et al., 2012; Li et al., 2012), as discussed above, used a random-effect model and obtained significant effect sizes. However, neither reported PIs for their analyses. By including the PI in the description of the results, we failed to obtain a significant association between DUIC or other computed parameters and UTEs, even if the OR and CI of our study were highly similar to theirs. Moreover, when analyzing various subgroups, we could not find statistical associations.

We could imagine three possible causes for the results obtained in our study: (1) the absence of any correlation between DUIC and UTEs; (2) a very high heterogeneity of the studies included in the analysis, which led to a very high dispersion of the true effect size; and (3) a lack of sensitivity of the random-effects model for our analysis. Riley et al. (2011) argued that using PIs is appropriate when the studies included in the meta-analysis have a low risk of bias, which was not the case in our study (see I² values in Figures 3, 5–8).

The heterogeneity of the results is caused both by the different methodologies used by various studies and by the technical difficulties in actually assessing impairment of driving associated with cannabis use (Fischer et al., 2006; Gorun et al., 2010; Hartman et al., 2016b). For example, it is known that the concentration in blood of THC decreases rapidly after use, but the clinical effects take longer to dissipate (Hartman et al., 2016a). This, associated with the fact that sampling is done sometimes even hours after the event, may significantly alter the results. The metabolism of cannabinoids is highly variable in different subjects; therefore, at the same concentration, a person may be under the influence, while another may have normal driving ability (Hartman et al., 2016a). There is no standardized cut-off for blood THC from which a person is considered under the influence. For more details, see National Academies of Sciences and Medicine (2017, Chapter 15).

To clarify which hypothesis is more plausible, we reanalyzed the studies, by using a newly developed approach to meta-analyses namely the IVHet model, which was developed specifically to decrease heterogeneity and increase the overall performance of the meta-analysis. Succinctly, IVHet has as key benefits over the random-effect model the lack of penalization for larger trials, a more conservative CI, and the fact that it exhibits a lesser true variance irrespective of the degree of heterogeneity (Doi et al., 2015). By using this approach, we identified a lack of significance when including in the final analysis studies that couldn’t quantify the actual DUIC using objective methods (blood tests). When DUIC was properly established using blood tests, we saw an OR suggesting a small-to-medium effect size (by also considering the CI). Case–control studies showed a significantly increased effect size; most likely the primary cause is a more frequent usage of a blood test to detect THC compared to other study designs. Also, case–control tests may misrepresent the actual situation in the population from which the cases/controls were obtained, by inadvertently pre-selecting subgroups from the population of drivers, which are more prone to risky behaviors. For example, it is possible that the control group was matched for age and sex with the case group (Mura et al., 2003), not taking into account the fact that men and younger persons tend to use cannabis more frequently. Alternatively, it is possible that the design included temporal patterns of obtaining controls (Asbridge et al., 2014), not considering that cannabis use occurs more at night or weekends (Laumon et al., 2005; Kuypers et al., 2012), when the risk of UTE is higher. Alternatively, the controls were matched geographically (Woratanarat et al., 2009), not taking into account that usually urban areas have a higher overall usage of cannabis compared to the countryside (Drummer et al., 2003; Li et al., 2012), and subsequently there are differences in the collision risk (Khorashadi et al., 2005; Zwerling et al., 2005). Some studies specifically scrutinized for this issue and adjusted the OR for confounding variables; by using this approach, they obtained inconclusive results even if they initially hinted an increased risk for cannabis users. For example, Blows et al. (2005) showed a decrease in OR from 11.4 (3.63–35.75) to 0.8 (0.2–3.1) after adjusting for confounders. By using adjusted ORs in a separate inquiry, we obtained results that were highly similar with those obtained through the analysis of unadjusted ORs.

A similar reasoning could be applied to the association between cannabis use and death. All studies that included death as the outcome were case–controls. Moreover, for the cases group, intoxication status was assessed using blood analysis; the cannabis use in control subjects was evaluated using various methods including saliva and/or urine (Li et al., 2013). Case–control studies in this field are known to have a high selection bias, caused by the fact that the two groups may be from different time periods or geographical regions, or they might include different cannabis-detection cutoffs (Hartman and Huestis, 2013).

Studies performed using other designs like retrospective cohorts or surveys often suffer from the same methodological flaws, especially preselecting participants who are more likely to use marijuana. They also rely heavily on self-reported cannabis use, that is known to under-represent the actual consumption and subsequently DUIC (Asbridge et al., 2014). Therefore, a methodology combining both types of studies would most likely obtain results that could reveal closer estimates of the true effect size of DUIC on UTEs.

We assessed publication bias both qualitatively (funnel plot) and quantitatively (Rosenthal’s Fail-Safe N and Duval and Tweety’s Trim and Fill methods). The latter were particularly useful as they allowed us to estimate the effect size after adjusting for publication bias. By analyzing the results obtained through the Trim and Fill method applied to the results of the random-effect model, we saw a sharp decrease in the effect size in most instances (except chronic cannabis use, that was usually used as a comparison/control group to acute exposure in the analyzed studies). Moreover, the results were, in most part, very close to those obtained using IVHet, suggesting that a significant proportion of the heterogeneity encountered by the random-effect model was caused by publication bias favoring studies that showed a positive association between DUIC and UTEs [see Doi and Thalib (2008) for further details].

Our analysis suggested that the effect size for unadjusted DUIC on UTEs was not statistically significant. This result might be caused by (1) methodological flaws, which are often encountered in articles on this topic [for a detailed analysis see Hartman and Huestis (2013)]; (2) the indiscriminate use of the term cannabis use (which in our study included a wide array of studies, including some in which the cut-off value for this substance was below the one known to cause a significant clinical impairment); and (3) or a true absence of a negative effect of DUIC on UTEs. Simply identifying cannabis use in a driver is not enough to justify the assumption of an increased risk for UTEs. When such a result is obtained, it should be corroborated with either quantitative data regarding cannabis use, or a clinical assessment of the driver, before establishing his fitness to drive.