Table 5. Analysis of risk of bias for included studies.

RISK OF BIAS	SUMMARY
Random sequence generation	Random sequence generation methods were found to be low risk of bias in fifteen studies as a computerized random generator [48,49,51,56,57,58,61,63] random number permutations [52], Latin square design [62], William’s square [47,60], and web-based number generator [53,54,55] were used. An unclear risk of bias was identified in two of the studies as they stated the goal was randomization but the strategy for randomization was not specifically stated [50,59].
Allocation concealment	Allocation concealment was identified as low risk in fifteen trials. The interventions were prepared and packaged by a third party in identical packaging in seven [47,50,52,57,60,61,63], key study assignments were withheld in one 56, sealed envelopes were used in four [48,49,51,58] and the allocation was kept concealed in the pharmacy in three [53,54,55] trials. The risk of bias was unclear in two studies. One paper 62 stated "We have shown the feasibility of a single-dose delivery method for smoked cannabis, and that blinding participants to treatment allocation is possible using this method", but does not describe how. High risk of bias was given to one trial [59] for having no description of allocation concealment.
Blinding of participants/personnel	Blinding of participants and personnel was identified as low risk in four studies [57,58,60,61]. Placebo capsules were identical and randomization, labeling, and packaging in high-density polyethylene bottles and dispensed under blinded conditions in one study [57]. Study spray medication and placebo were taste- and color-matched with peppermint oil and coloring in three studies [58,60,61]. An unclear risk assessment was assigned to eleven studies [47,48,63,49,50,51,52,53,54,55,62]. Although there were identical pre-rolled cigarettes in one study, participants were required to have previous experience smoking cannabis63. Seven studies did not give enough information to determine level of blinding [47,48,49,50,53,55,62]. In one study although medication and placebo were in identical containers 67% of participants correctly identified active medication [51]. Similarly 89% of the subjects correctly identified the active medication in another study [54]. One study was assigned unclear risk of bias because of psychoactive effects of both placebo and treatment [52]. A high risk of bias for participants and personnel was assigned to two trials [56,59]. In one study all but one participant correctly identified the active treatment [56]. One study was assigned a high risk of bias because no blinding description was given [59].
Blinding assessors/statistician	Blinding of assessors and statisticians was identified as low risk in five studies with the key for study assignment withheld from investigation until analysis was completed [47,51,52,55,56]. Blinding of assessors and statisticians was assigned unclear risk bias for ten studies as they stated the trial was "double-blinded" but gave no description of blinding methods for assessors and/or statisticians [48,49,50,53,54,57,60,61,62,63]. Two studies were identified as high risk of bias for assessors and statisticians [58,59]. In one study the sponsor of the study participated in the analysis [58]. Another study did not indicate that blinding of assessors and/or statisticians was performed [59].
Incomplete outcome data	Twelve of the studies were assigned low riskof bias for incomplete outcome data; these studies had no missing outcome data reported [47,49,62,63,50,51,52,53,55,56,57,58]. Fourof the studies were identified as unclear risk of bias due to incomplete outcome data reporting [54,59,60]. Two studies lacked information on intent-to-treat analysis adherence [59,60]. One study had participants who were not included in the analysis, although reasons for non-inclusion were stated [54]. One study had patients drop out during the study that were not excluded from analysis and no information was given on point of study during which they withdrew [48]. One study was identified as high risk of bias of incomplete outcome data reported due to a withdrawal rate of 29.7% for the study [61].
Selective reporting	A low risk of bias was assigned to all studies [47,48,57,58,59,60,61,62,63,49,50,51,52,53,54,55,56] because all outcomes were described and presented as pre-specified.
Other potential sources of bias	An unclear risk of bias was assigned to eleven [48,50,63,51,52,53,54,55,56,59,62] studies. Nine had co-interventions (patients used concomitant medications for pain) [48,50,52,53,54,55,56,62,63] One studydid not state the co-interventions [59] and one trial [51] was funded by the drug company that manufactures the intervention, however the statistical analyses were blinded. A high risk of bias was identified in six studies [47,49,57,58,60,61] as the authors received funding by the intervention manufacturer with a proprietary interest in the medications used.
Overall bias	Overall, the risk of bias was unclear in nine of the seventeen RCTs (52.9%) [48,50,51,52,56,54,55,62,63], and a high risk of bias was found in eight of the seventeen RCTs (47.1%) [47,49,56,57,58,59,60,61].

Abbreviations: RCT, randomized controlled trial.