Balancing risks and benefits of cannabis use: umbrella review of meta-analyses of randomised controlled trials and observational studies

Introduction

Cannabis contains over 100 cannabinoids, of which Δ⁹-tetrahydrocannabinol and cannabidiol are the most clinically relevant. Tetrahydrocannabinol is a partial agonist at CB1 and binds CB2 receptors. CB1 is widely expressed by central and peripheral neurones but also by immune cells and other type of cells in the brain and in the periphery, and when it binds with tetrahydrocannabinol, a so-called high is induced, which is responsible for potential misuse. CB2 receptors are also expressed by neurons, but less than CB1, and are most abundantly expressed in immune cells.123 Cannabidiol, however, does not produce the high and thus does not carry the same potential for substance misuse.4 Furthermore, cannabidiol does not seem to promote psychosis inducing effects.5 Cannabis use can evolve into cannabis use disorder, broadly defined as an inability to quit cannabis use, continuous use despite harmful consequences (eg, cannabinoid hyperemesis syndrome6), or functional impairment.78

According to the Global Burden of Disease 2019 study, more than 23.8 million people have cannabis use disorder globally,9 and cannabis use ranks third worldwide among consumed substances of misuse, after alcohol and tobacco.10111213 Cannabis use disorder is more common in men and high income countries. The prevalence of cannabis use disorder in the USA has been estimated to be around 6.3% in a lifetime and 2.5% for 12 months, and in Europe, around 15% of people aged 15-35 years reported cannabis use in the previous year.14 Of those using cannabis, one in three developed problems related to cannabis use that impaired functioning,13 and 10% used cannabis on a daily basis.15 Cannabis use disorder can affect up to 50% of people who use cannabis daily.16

In Europe, over the past decade, self-reported use of cannabis within the past month has increased by almost 25% in people aged 15-34 years, and more than 80% in people who are 55-64 years.17 Cannabis or products containing tetrahydrocannabinol (cannabinoids) are widely available and have increasingly high tetrahydrocannabinol content.18 For instance, in Europe, tetrahydrocannabinol content increased from 6.9% to 10.6% from 2010 to 2019.17 Evidence has suggested that cannabis may be harmful, for mental1920 and physical health,21 as well as driving safety,22 across observational studies but also in experimental settings.23 Conversely, more than a decade ago, cannabidiol was proposed as a candidate drug for the treatment of neurological disorders such as treatment-resistant childhood epilepsy. Furthermore, it has been proposed that this substance might be useful for anxiety and sleep disorders, and even as an adjuvant treatment for psychosis.24 Moreover, cannabis based medications (ie, medications that contain cannabis components) have been investigated as putative treatments for several different conditions and symptoms.23

The multifarious nature of cannabis’s main active components, contrasting evidence from observational studies reporting detrimental effects of cannabis, and therapeutic findings of cannabis based medicines from interventional studies, is reflected in different legislative approaches. Thus, in most countries cannabis use is illegal, but in a small and growing number of countries and states cannabis is legally sold without the need for a medical prescription.252627

Publication of meta-analyses investigating the effects of cannabinoids on health and other outcomes have substantially increased. However, most meta-analytical findings synthesised data from observational studies and are prone to several sources of bias.2829 To date, no umbrella review has systematically evaluated the evidence around cannabis, cannabinoids and cannabis nased medicines and health outcomes in humans from meta-analyses encompassing both observational studies and randomised controlled trials. Thus, this work aimed to systematically evaluate the breadth, quality, credibility, and certainty of associations between cannabis, cannabinoids, cannabis based medicines, and human health. We aimed to use established quantitative criteria, account for several sources of bias,303132 and identify converging findings from different study designs.

Methods

Searches and inclusion criteria

We conducted an umbrella review of meta-analyses of observational studies(ie, case-control and cohort studies) and randomised controlled trials that reported on any outcome associated with cannabis and cannabinoids use in humans. We followed an a-priori protocol (PROSPERO CRD42018093045). We adhered to PRIOR and PRISMA 2020 guidelines (adapting PRISMA to the abstract of an umbrella review; supplementary tables 1-2).3334 Two of the authors independently screened literature that was retrieved systematically by searching PubMed, Embase, and PsycINFO from database inception up to 9 February 2022, without language restrictions, and extracted data into a spreadsheet. The search key is available in the supplementary methods. We also manually searched the Cochrane Library. When two or more meta-analyses examined the same association, we selected only the one that included the largest number of studies. We excluded systematic reviews without a meta-analysis, meta-analyses of risk factors for cannabinoids use, meta-analyses of cross-sectional studies only, pooled analyses of studies identified without a systematic search, and individual studies.

Outcomes

The co-primary outcomes were the efficacy and safety of cannabinoids on target symptoms (eg seizures in epilepsy) in meta-analyses of randomised controlled trials. The secondary outcomes were any outcome reported in the meta-analyses of observational studies.

Data extraction and quality assessment

Extracted information from meta-analyses and individual studies included in meta-analyses were the bibliographic identifiers of the publication (ie, PubMed-Indexed for Medline or the digital object identifier), first author, year of publication, design of included studies (ie, cohort, case-control, randomised controlled trial), number of included studies in the meta-analysis, specific population under investigation (ie, general population, pregnant women, or people with medical disorders), the exposure and comparison definitions (eg author defined marijuana use v no use or heavy use of cannabis v no use), the outcomes, and their effect size and dispersion measure (when adjusted and unadjusted effect sizes were reported, we selected the adjusted ones). We also extracted what factors analyses were adjusted for. The methodological quality of each included meta-analysis was assessed by two independent investigators using A Measurement Tool to Assess Systematic Reviews version 2 (AMSTAR 2).35

Data analysis

For each association from observational studies (ie, between exposure to cannabis or cannabinoids and outcomes), we extracted the effect sizes of individual studies reported in each meta-analysis, recalculating the pooled effect sizes and 95% confidence intervals, using random effects models. Specifically, we re-analysed each eligible association under the random effects model with DerSimonian and Laird method if included studies were equal or more than 10,36 and Hartung, Knapp, Sidik, and Jonkman if less than 10.37 We transformed the initial effect sizes or modified the direction of associations presented by the original authors to present comparable estimates (ie, equivalent odds ratio; supplementary methods).38 Heterogeneity was tested with the I² and Tau statistics.39 I² measures the proportion of the total variability due to heterogeneity, Tau measures true heterogeneity as an absolute measure of heterogeneity, instead. Moreover, 95% prediction intervals for the summary random effect sizes were computed to estimate the possible range in which the effect sizes of future studies were anticipated to fall.40 We calculated prediction intervals using both the estimated between-study heterogeneity variance given from tau2 as well as the standard error of the pooled effect. We then examined small study effect bias (ie, whether smaller studies generated larger effect sizes compared with larger studies).38414243444546 Small study effect was deemed present when both the Egger regression asymmetry test indicated publication bias (P value ≤0.10), and the random effects summary effect size was larger than the effect size of the largest study contributing to that association.42444546 Finally, we evaluated significance bias using an updated method to detect the publication selection of statistically significant findings based on observable excess statistical significance.4748 We computed the test of excess statistical significance and the proportion of statistical significance, which have adequate control for type I errors and high statistical power. The presence of excess significance bias for individual meta-analyses was considered if either excess statistical significance or proportion of statistical significance were greater than 1.645.47

All analyses were conducted in Stata/SE, version 17.0.

Assessment of the credibility of evidence

In accordance with previous umbrella reviews,49505152 eligible associations from observational studies were classified into five levels according to the strength of the evidence of potential environmental risk or protective factors: convincing (class I), highly suggestive (class II), suggestive (class III), weak (class IV), and not significant. Briefly, credibility of evidence from observational studies is rated on the basis of the number of events developing the outcome of interest, P value of the association, small study effect, excess of significance bias, prediction intervals, statistical significance of the largest study, and heterogeneity. The specific criteria are exhaustively reported in the supplementary methods. We used sensitivity analyses on all levels of evidence, removing the criterion of more than 1000 cases, and on adjusted estimates and cohort studies on class I and II evidence only (supplementary methods).

We classified evidence from meta-analyses of randomised controlled trials, updating a previously proposed framework, classifying certainty of evidence as high, moderate, low, or very low,53 based on GRADE (Grading of Recommendations, Assessment, Development and Evaluations).32 GRADE is a transparent framework that is widely used to develop and present evidence synthesis, providing a set of explicit criteria across different domains to assess level of evidence, and making clinical practice recommendations. As recommended by GRADE, the level of evidence was determined by risk of bias, inconsistency, indirectness, imprecision, and publication bias (supplementary methods).

Patient and public involvement

This study was author funded and we did not involve patients and the public in this work, but we will apply for funding to involve them in the knowledge translation of present findings. Knowledge translation activities will include, but will not be limited to, dissemination of findings via personal and institutional social media, education of health professional trainees, and continuous medical education activities for health professionals. We will involve patient and public representatives in creating a plain language summary of findings to be distributed to the clinical population with mental health disorders, and pregnant women, informing policy makers across different countries with written communications.

Results

Literature search

Starting from 6657 records after duplicate removal, we excluded 5941 studies at title and abstract screening stage, and 599 at full-text level, resulting in 101 publications included. Studies identified by manual search had already been identified from the systematic search. The list of studies excluded after full-text assessment, with reason for exclusion, is reported in supplementary table 3, and the article selection flow is reported in figure 1.33 Of the 101 articles, 50 were meta-analyses of observational studies (215 meta-analytical associations),212254555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101 and 51 were meta-analyses of randomised controlled trials2374102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151 (364 meta-analytical associations) Of note, one meta-analysis reported on both observational and randomised controlled trials (table 1, table 2, supplementary material 2).74

source: https://www.bmj.com/content/382/bmj-2022-072348

1. Marco Solmi, associate professor1234567,  
2. Marco De Toffol, psychiatrist8,  
3. Jong Yeob Kim, public health doctor9,  
4. Min Je Choi, psychiatrist9,  
5. Brendon Stubbs, advanced fellow1011,  
6. Trevor Thompson, associate professor of clinical research12,  
7. Joseph Firth, research fellow1314,  
8. Alessandro Miola, psychiatrist15,  
9. Giovanni Croatto, psychiatrist16,  
10. Francesca Baggio, psychiatrist16,  
11. Silvia Michelon, psychiatrist17,  
12. Luca Ballan, psychiatrist17,  
13. Björn Gerdle, professor18,  
14. Francesco Monaco, psychiatrist1920,  
15. Pierluigi Simonato, psychiatrist21,  
16. Paolo Scocco, psychiatrist22,  
17. Valdo Ricca, Professor23,  
18. Giovanni Castellini, associate professor23,  
19. Michele Fornaro, assistant professor24,  
20. Andrea Murru, researcher25,  
21. Eduard Vieta, professor25,  
22. Paolo Fusar-Poli, professor526,  
23. Corrado Barbui, professor27,  
24. John P A Ioannidis, professor282930,  
25. Andrè F Carvalho, senior researcher31,  
26. Joaquim Radua, associate professor32,  
27. Christoph U Correll, professor73334,  
28. Samuele Cortese, professor635363738,  
29. Robin M Murray, professor39,  
30. David Castle, professor4041,  
31. Jae Il Shin, professor4243,  
32. Elena Dragioti, associate professor1844
33. Correspondence to: M Solmi msolmi@toh.ca

• Accepted 27 June 2023