A total of 99 studies published between 1971 and 2023 reported data on rates of CAPS and were included in the analysis, comprising 126,430 individuals from independent samples. Convergence of the data extracted by the two researchers (T.S. and W.B.) was high for the pooled rates on CAPS from observational studies (rate_DIFF = −0.01%, where rate_DIFF = rate_TS – rate_WB), experimental studies (rate_DIFF = 0%), and medicinal cannabis studies (rate_DIFF = 0%). More specifically, we included data from 41 observational studies (n = 92,888 cannabis users), 19 experimental studies administering THC (n = 754), and 79 studies assessing efficacy and tolerability of medicinal cannabis products containing THC (n = 32,821). In medicinal trials, the most common conditions treated with THC were pain (k = 19 (23.75%)) and cancer (k = 16 (20%)) (see Supplementary Table 1 for an overview). The age distribution of the included participants was similar in observational studies (mean age = 24.47 years, ranging from 16.6 to 34.34 years) and experimental studies (mean age = 25.1 years, ranging from 22.47 to 27.3 years). Individuals taking part in medicinal trials were substantially older (mean age = 48.16 years, ranging from 8 to 74.5 years).

As summarized in Fig. 3 and Supplementary Table 3, substantial rates of CAPS were reported by observational studies (19.4%, 95% confidence interval (CI): 14.2%, 24.6%) and THC-challenge studies (21%, 95% CI: 11.3%, 30.7%), but not medicinal cannabis studies (1.5%, 95% CI: 1.1%, 1.9%). The pooled rates estimated for different symptom profiles of CAPS (CAPS – paranoia, CAPS – hallucinations, CAPS – delusions) are displayed in Supplementary Fig. 2. All individual study estimates are listed in Supplementary Table 2.

Most models showed significant levels of heterogeneity (Supplementary Table 3), highlighting that rates of CAPS differed as a function of study-specific features. Risk of publication bias was indicated (P_Peters < 0.05) for one of the meta-analytical models combining all rates of CAPS (see funnel plots, Supplementary Fig. 2). Applying the trim-and-fill method slightly reduced the pooled rate of CAPS obtained from medicinal cannabis studies (rate_unadjusted = 1.53%; rate_adjusted = 1.18%). Finally, Fig. 4 summarizes rates of CAPS of a subset of studies where CAPS was defined as the occurrence of a full-blown cannabis-associated psychotic episode (as described in Table 1). When combined, the rate of CAPS (full episode) was 0.52% (0.42–0.62%) across the three study designs, highlighting that around one in 200 individuals experienced a severe episode of psychosis when exposed to cannabis/THC. Rates of CAPS (full episode) as reported by the individual studies showed high levels of consistency (I² = 8%, P(I²) = 0.45; Fig. 4).

Assessing predictors of CAPS, we included 103 studies published between 1976 and 2023, corresponding to 80 independent samples (n = 170,158 non-overlapping individuals). In total, we extracted 381 Cohen’s d that were pooled in 44 separate meta-analytical models. A summary of all extracted study estimates is provided in Supplementary Table 4. Comparing the P values of the individual Cohen’s d to the original P values as reported in the studies revealed a high level of concordance (r = 0.96 P = 1.1 × 10^–79), indicating that the conversion of the raw study estimates to a common metric did not result in a substantial loss of information. Comparing the results obtained from the data extracted by two researchers (T.S. and W.B.) identified virtually no inconsistencies when inspecting estimates of Cohen’s d, as obtained for severity of cannabis use on CAPS (d_DIFF = 0, where d_DIFF = d_TS –d _WB), gender (d_DIFF = 0), administration of (placebo controlled) medicinal cannabis (d_DIFF = 0.003), psychosis liability (d_DIFF = 0), and administration of a single dose of THC (d_DIFF = 0).

Figure 5 summarizes the results obtained from the meta-analytical models. We examined whether CAPS was predicted by the pharmacodynamic properties of cannabis, a person’s cannabis use history, demographic factors, mental health/personality traits, neurotransmitters, genetics, and use of other drugs: With respect to the pharmacodynamic properties of cannabis, the largest effect on CAPS severity was present for a single dose of THC (d = 0.7, 95% CI: 0.52, 0.87) as administered in experimental studies, followed by a significant dose–response effect of THC on CAPS (d = 0.42, 95% CI: 0.25, 0.59, that is, tested as moderation effects of THC dose in experimental studies). When tested in medicinal randomized controlled trials, cannabis products significantly increased symptoms of CAPS (d = 0.14, 95% CI: 0.05, 0.23), albeit by a smaller magnitude. Protective effects were present for low THC/COOH levels (d = −0.22, 95% CI: −0.39, −0.05, that is, the inactive metabolite of cannabis), but not for the THC/CBD (cannabidiol) ratio (d = −0.19, 95% CI: −0.43, 0.05, P = 0.13).

Less clear were the findings with respect to the cannabis use history of the participants and its effect on CAPS. Here, neither young age of onset of cannabis use nor high-frequency use of cannabis or the preferred type of cannabis (strains high in THC, strains high in CBD) was associated with CAPS. The only demographic factors that significantly predicted CAPS were age (d = −0.17, 95% CI: −0.292, −0.050) and gender (−0.09, 95% CI: −0.180, −0.001), indicating that younger and female cannabis users report higher levels of CAPS compared with older and male users. With respect to mental health and personality, the strongest predictors for CAPS were diagnosis of bipolar disorder (d = 0.8, 95% CI: 0.54, 1.06)) and psychosis liability (d = 0.49, 95% CI: 0.21, 0.77), followed by mood problems (anxiety d = 0.44, 95% CI: 0.03, 0.84; depression d = 0.37, 95% CI: 0.003, 0.740) and addiction liability (d = 0.26, 95% CI: 0.14, 0.38). Summarizing the evidence from studies looking at neurotransmitter functioning showed that increased dopamine activity significantly predicted CAPS (d = 0.4, 95% CI: 0.16, 0.64) (for example, reduced CAPS following administration of D2 blockers such as olanzapine²⁸ or haloperidol²⁹). By contrast, alterations in the opioid system did not reduce risk of CAPS. Similarly, none of the assessed candidate genes showed evidence of altering response to cannabis. Finally, out of 11 psychoactive substances with available data, only use histories of MDMA (3,4-methyl enedioxy methamphetamine) (d = 0.2, 95% CI: 0.03, 0.36), crack (d = 0.13, 95% CI: 0.03, 0.23), inhalants (d = 0.12, 95% CI: 0.03, 0.22), and sedatives (d = 0.12, 95% CI: 0.02, 0.22) linked to increases in CAPS.

Most of the meta-analytical models showed considerable levels of heterogeneity (I² > 80%; Supplementary Table 5), notably when summarizing findings from observational studies (for example, severity of cannabis use: I² = 98%, age of onset of cannabis use: I² = 98%), highlighting that the individual effect estimates varied substantially across studies. By contrast, lower levels of heterogeneity were present when pooling evidence from experimental and medicinal cannabis studies (for example, effects of medicinal cannabis: I² = 18%; THC dose–response effects: I² = 37%). While risk of publication bias was indicated for four of the meta-analytical models (Egger’s test P < 0.05) (Supplementary Fig. 3), an inspection of trim-and-fill adjusted estimates did not alter the conclusions for (1) administration of a single dose of THC (P_Egger < 0.0001, d_unadjusted = 0.7, d_{trim-and-fill} = 0.49), (2) CBD administration (P_Egger = 0.0001, d_unadjusted = −0.19, d_{trim-and-fill} = −0.14, both P < 0.05), psychosis liability (P_Egger = 0.025, d_unadjusted = 0.49, d_{trim-and-fill} = 0.49), and (3) diagnosis of depression (P_Egger = 0.019, d_unadjusted = 0.37, d_{trim-and-fill} = 0.54). Outliers were identified for seven meta-analytical models (Supplementary Fig. 4). Removing outliers from the models did not substantially alter the conclusions drawn from the models, as indicated for age (d = −0.18, d_corr = −0.14, both P < 0.05); anxiety (d = 0.61, d_corr = 0.47, both P < 0.05), severity of cannabis use (d = 0.19, d_corr = 0.25, both P > 0.05), depression (d = 0.41, d_corr = 0.25, both P > 0.05), gender (d = −0.09, d_corr = −0.12, both P < 0.05), psychosis liability (d = 0.49, d_corr = 0.43, both P < 0.05), and administration of a single dose of THC (d = 0.6, d_corr = 0.56, both P < 0.05). Sensitivity checks assessing whether Cohen’s d changes as a function of within-subject correlation coefficient highlighted that the results were highly concordant (Supplementary Fig. 6). Minor deviations from the main analysis were present for the effects of a single dose of THC (d_r=0.3 = 0.64 versus d_r=0.5 = 0.69 versus d_r=0.7 = 0.77) and dose–response effects of THC (d_r=0.3 = 0.45 versus d_r=0.5 = 0.42 versus d_r=0.7 = 0.39), but this did not alter the interpretation of the findings.

Finally, we assessed consistency of findings for predictors examined in more than one of the different study designs (observational, experimental, and medicinal cannabis studies), as illustrated for four meta-analytical models in Fig. 6 (see Supplementary Fig. 7 for the complete set of results). Triangulating the results highlighted that consistency with respect to the direction of effects was particularly high for age (d_Experiments = −0.14 versus d_{Observational} = −0.19 versus d_{Quasi-Experimental} = −0.16) and gender (d_Experiments = −0.09 versus d_{Observational} = −0.07 versus d_{Quasi-Experimental} = −0.25) on CAPS. By contrast, little consistency across the different study designs was present with respect to cannabis use histories, notable age of onset of cannabis use (d_{Observational} = −0.3 versus d_{Quasi-Experimental} = 0.24), and use of high-THC cannabis (d_{Observational} = 0.12 versus d_{Quasi-Experimental} = −0.13).

Summarizing the existing evidence on rates of CAPS, we find that cannabis can acutely induce CAPS in a subset of cannabis-exposed individuals, irrespective of whether it is used recreationally, administered in controlled experiments, or prescribed as a medicinal product. Importantly, rates of CAPS varied substantially across the designs. More specifically, similar rates of CAPS were reported by observational and experimental evidence (around 19% and 21% in cannabis-exposed individuals, respectively), while considerably lower rates of CAPS were documented in medicinal cannabis samples (between 1% and 2%).

A number of factors likely contribute to the apparently different rates of CAPS across the three study designs. First, rates of CAPS are not directly comparable as different, design-specific measures were used: in observational/experimental research, CAPS is typically defined as the occurrence of transient cannabis-induced psychosis-like symptoms, whereas medicinal trials screen for CAPS as the occurrence of first-rank psychotic symptoms, often resulting in treatment discontinuation.^20,30,31 As such, transient CAPS may indeed occur commonly in cannabis-exposed individuals (as evident in the higher rates in observational/experimental research), while risk of severe CAPS requiring medical attention is less frequently reported (resulting in lower reported rates in medicinal cannabis samples). This converges with our meta-analytic results, showing that severe CAPS (full psychotic episode) may occur in about 1 in 200 (0.5%) cannabis users. Another key difference between medicinal trials and experimental/observational research lies in the demographic profile of participants recruited into the studies. For example, individuals taking part in medicinal trials were substantially older (mean age: 48 years) compared with subjects taking part in observational or experimental studies (mean age: 24 and 25 years, respectively). As such, older age may have buffered some of the adverse effects reported by adolescent individuals. Finally, cannabis products used in medicinal trials contain noticeable levels of CBD (for example, Sativex, with a THC/CBD ratio of approximately 1:1), a ratio different from that typically found in street cannabis (for example, >15% THC and <1% CBD³²) and in the experimental studies included in our meta-analyses (pure THC). As such, the use of medicinal cannabis (as opposed to street cannabis) may constitute a somewhat safer option. However, the potentially protective effects of CBD in this context require further investigation as we did not find a consistent effect of CBD co-administration on THC-induced psychosis-like symptoms. While earlier experimental studies included in our work were suggestive of protective effects of CBD,^33–35 two recent studies did not replicate these findings.^36,37

Interestingly, lower but significant rates of CAPS were also observed in placebo groups assessed as part of THC-challenge studies (%_THC = 25% versus %_placebo = 11%) and medicinal cannabis trials (%_THC = 3% versus %_placebo = 1%), highlighting that psychotic symptoms occur not only in the context of cannabis exposure. This is in line with the notion that cannabis use can increase risk of psychosis but appears to be neither a sufficient nor necessary cause for the emergence of psychotic symptoms.³⁸

Summarizing evidence on predictors of CAPS, we found that individual vulnerabilities and the pharmacological properties of cannabis both appear to play an important role in modulating risk. Regarding the pharmacological properties of cannabis, evidence from experimental studies showed that the administration of THC increases risk of CAPS, both in a single-dose and dose-dependent manner. Given the nature of the experimental design, these effects are independent of potential confounders that bias estimates obtained from observational studies. More challenging to interpret are therefore findings on individual cannabis use histories (for example, frequency/severity of cannabis use, age of onset of use, preferred cannabis strain) as assessed in observational studies. Contrary to evidence linking high-frequency and early-onset cannabis use to long-term risk of psychosis,³⁹ none of these factors associated with CAPS in our study. This discrepancy may indicate that cumulative effects of THC exposure are expressed differently for long-term risk of psychosis and acute CAPS: while users accustomed to cannabis may show a more blunted acute response as a result of tolerance, they are nevertheless at a higher risk of developing the clinical manifestation of psychosis in the long run.³⁸

We also tested a number of meta-analytical models for predictors tapping into demographic and mental health dimensions. Interestingly, among the assessed demographic factors, only age and gender associated with CAPS, with younger and female individuals reporting increased levels of CAPS. Other factors often linked to mental health, such as education or socioeconomic status, were not related to CAPS. Concerning predictors indexing mental health, we found converging evidence showing that a predisposition to psychosis increased the risk of experiencing CAPS. In addition, individuals with other pre-existing mental health vulnerabilities (for example, bipolar disorder, depression, anxiety, addiction liability) also showed a higher risk of CAPS, indicating that risk may stem partly from a common vulnerability to mental health problems.

These findings align with findings from studies focusing on the biological correlates of CAPS, showing that increases in dopamine activity, a neurotransmitter implicated in the etiology of psychosis,⁴⁰ altered sensitivity to cannabis. By contrast, none of the a priori selected candidate genes (chosen mostly to index schizophrenia liability) modulated risk of CAPS. This meta-analytic finding is coherent with results from the largest available genome-wide association study on schizophrenia,⁴¹ where none of the candidate genes reached genome-wide significance (P < 5 × 10⁻⁸) (Supplementary Information). Instead, as for any complex trait, genetic risk underlying CAPS is likely to be more polygenic in nature, possibly converging on pathways as yet to be identified. As such, genetic testing companies that screen for the aforementioned genetic variants to provide their customers with an individualized risk profile (such as the Cannabis Genetic Test offered by Lobo Genetics (https://www.lobogene.com)) are unlikely to fully capture the genetic risk underlying CAPS. Similarly, genetic counseling programs targeting specifically AKT1 allele carriers in the context of cannabis use⁴² may be only of limited use when trying to reduce cannabis-associated harms.

This work has a number of implications for future research avenues. First, experimental studies administering THC constitute the most stringent available causal inference method when studying risk of CAPS. Future studies should therefore capitalize on experimental designs to advance our understanding of the acute pharmacological effects of cannabis, in terms of standard cannabis units,⁴³ dose–response risk profiles,⁴⁴ the interplay of different cannabinoids,^44,45 and building on recent work.

Despite the value of experimental studies in causal inference, observational studies are essential to identify predictors of CAPS that cannot be experimentally manipulated (for example, age, long-term/chronic exposure to cannabis) and to strengthen external validity. However, a particular challenge for inference from observational studies results from bias due to confounding and reverse causation. Triangulating and comparing findings across study designs can therefore help to identify potential sources of bias that are specific to the different study designs.⁴⁶ For example, we observed that, despite THC dosing being robustly associated with CAPS in experimental studies, we did not find an association between cannabis use patterns (for example, high-THC cannabis strain) in observational and quasi-observational studies. This apparent inconsistency may result from THC effects that are blunted by long-term, early-onset and heavy cannabis use. For other designs, reverse causation may bias the association between cannabis use patterns and CAPS: as individuals may reduce cannabis consumption as a result of adverse acute effects,⁴⁷ the interpretation of cross-sectional estimates concerning different cannabis exposure and risk of CAPS is particularly challenging. Future observational studies should therefore exploit more robust causal inference methods (for example, THC administration in naturalistic settings⁴⁸ or within-subject comparisons controlling for time-invariant confounds⁴⁹) to better approximate the experimental design. In particular, innovative designs that can provide a higher temporal resolution on cannabis exposures and related experiences (for example, experience sampling,⁵⁰ assessing daily reactivity to cannabis⁵¹) are a valuable addition to the causal inference toolbox for cannabis research. Applying genetically informed causal inferences such as Mendelian randomization analyses⁵² can further help to triangulate findings, which would be possible once genome-wide summary results for both different cannabis use patterns and CAPS become available.

With respect to medicinal trials, it is important to note that an assessment of CAPS has not been a primary research focus. Although psychotic events are recognized as a potential adverse reaction to medicinal cannabis,⁵³ data on CAPS are rarely reported by medicinal trials, considering that only about 20% of medicinal cannabis randomized controlled trials screen for psychosis as a potential adverse effects.⁵ As such, trials should systematically monitor CAPS, in addition to longer-term follow-ups assessing the risk of psychosis as a result of medicinal cannabis use. In particular, the use of validated instruments designed to capture more-subtle changes in CAPS should be included in trials to more adequately assess adverse reactions associated with medicinal cannabis products.

Second, with respect to factors associated with risk of CAPS, we find that these are similar to factors associated with onset of psychosis, notably pre-existing mental health vulnerabilities,⁵⁴ dose–response effects of cannabis,⁵⁵ and young age.¹² The key question deserving further attention is therefore whether CAPS constitutes, per se, a risk maker for long-term psychosis. Preliminary evidence found that in individuals with recent-onset psychosis, 37% reported to have experienced their first psychotic symptoms during cannabis intoxication.⁵⁶ Future longitudinal evidence building on this is required to determine whether subclinical cannabis-associated psychotic symptoms can help to identify users at high risk of developing psychosis in the long run. Follow-up research should also examine longitudinal trajectories of adverse cannabis-induced experiences and the distress associated with these experiences, given research suggesting that high levels of distress/persistence may constitute a marker of clinical relevance of psychotic-like experiences.⁵⁷ While few studies have explored this question in the context of CAPS, there is, for example, evidence suggesting that the level of distress caused by acute adverse reactions to cannabis may depend on the specific symptom dimension.⁵⁸ Here the highest levels of distress resulted from cannabis-associated paranoia and anxiety, rather than cannabis-associated hallucinations or experiences tapping into physical sensations (for example, body humming, numbness). In addition, some evidence highlights the re-occurring nature of CAPS in cannabis-exposed individuals.^22,58 Further research focusing on individuals with persisting symptoms of CAPS may therefore help to advance our knowledge concerning individual vulnerabilities underlying the development of long-term psychosis in the context of cannabis use.

Importantly, our synthesizing analysis is not immune to the sources of bias that exist for the different study designs, and our findings should therefore be considered in light of the aforementioned limitations (for example, residual confounding or reverse causation in observational studies, limited external validity in experimental studies). Nevertheless, comparing findings across the different study designs allowed us to pin down areas of inconsistency, which existed mostly with regard to cannabis-related parameters (for example, age of onset, frequency of use) and CAPS. In addition, we observed large levels of heterogeneity among most meta-analysis models, highlighting that study-specific findings may vary as a result of different sample characteristics and study methodologies. Future studies aiming to further discern potential sources of variation such as study design features (for example, treatment length in medicinal trials, route of THC administration in experimental studies), statistical modeling (for example, the type of confounding factors considered in observational research), and sample demographics (for example, age of the participants, previous experience with cannabis) are therefore essential when studying CAPS.

A systematic literature search was performed in three databases (MEDLINE, EMBASE, and PsycInfo) following the PRISMA guidelines.⁵⁹ The final search was conducted on 6 December 2023 using 26 search terms indexing cannabis/THC and 20 terms indexing psychosis-like outcomes or cannabis-intoxication experiences (see Supplementary Information for a complete list of search terms). Search terms were chosen on the basis of terminology used in studies assessing CAPS, including observational studies (self-reported cannabis-induced psychosis-like experiences), THC-challenge studies (testing change in psychosis-like symptoms following THC administration), and medicinal studies testing the efficacy and safety of medicinal cannabis products (adverse events related to medicinal cannabis). Before screening the identified studies for inclusion, we removed non-relevant article types (reviews, case reports, comments, guidelines, editorials, letters, newspaper articles, book chapters, dissertations, conference abstracts) and duplicates using the R package revtools⁶⁰. A senior researcher experienced in meta-analyses on cannabis use (T.S.) then reviewed all titles and abstracts for their relevance before conducting full-text screening. To reduce the risk of wrongful inclusion at the full-text screening stage, 10% of the articles selected for full-text screening were cross-checked for eligibility by a second researcher (E.M.).

We included all study estimates that could be used to derive rates of CAPS (the proportion of cannabis-exposed individuals reporting CAPS) or effect sizes (Cohen’s d) for factors predicting CAPS. CAPS was defined as the occurrence of hallucinations, paranoia, and/or delusions during cannabis intoxication. These symptom-level items have been identified as the most reliable self-report measures screening for psychosis when validated against clinical interview measures.⁶¹ Table 1 provides examples of CAPS as measured across the three different study designs. In brief, from observational studies, we extracted data if CAPS was assessed in cannabis-exposed individuals on the basis of self-report measures screening for subjective experiences while under the influence of cannabis. From experimental studies administering THC, CAPS was measured as the degree of psychotic symptom change in response to THC, either estimated from a between-subject (placebo groups versus THC group) or within-subject (pre-THC versus post-THC assessment) comparison. We also included data from natural experiments (referred to as quasi-experimental studies hereafter), where psychosis-like experiences were monitored in recreational cannabis users before and after they consumed their own cannabis products.^23,62 Finally, with respect to trials testing the efficacy and/or safety of medicinal cannabis products containing THC, we extracted data on adverse events, including the occurrence of psychosis, hallucinations, delusions, and/or paranoia during treatment with medicinal cannabis products. Medicinal studies that tested the effects of cannabis products not containing THC (for example, CBD only, olorinab, lenabasum) were not included.

For 10% of the included studies, data on rates and predictors of CAPS were extracted by a second researcher (W.B.), and agreement between the two extracted datasets was assessed by comparing the pooled estimates on rates and predictors of CAPS. In addition, following recommendations for improved reproducibility and transparency in meta-analytical works,⁶³ we provide all extracted data, the corresponding analytical scripts, and transformation information in the study repository.

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Nature Mental Health thanks Evangelos Vassos and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

The data are publicly available via GitHub at github.com/TabeaSchoeler/TS2023_MetaCAPS.

All analytical code used to analyze, summarize, and present the data is accessible via GitHub at github.com/TabeaSchoeler/TS2023_MetaCAPS.