Introduction
Cannabis is used by over 200 million people worldwide, and the prevalence of use has increased in many countries in recent years, as has the potency of the cannabis available and the number of people seeking treatment for cannabis-related problems (European Monitoring Centre for Drugs and Drug and Addiction, 2016; Freeman et al., Reference Freeman, Groshkova, Cunningham, Sedefov, Griffiths and Lynskey2019; Grucza, Agrawal, Krauss, Cavazos-Rehg, & Bierut, Reference Grucza, Agrawal, Krauss, Cavazos-Rehg and Bierut2016; Manthey, Reference Manthey2019). Prospective epidemiological and biological studies (Gage, Hickman, & Zammit, Reference Gage, Hickman and Zammit2016; Murray et al., Reference Murray, Englund, Abi-Dargham, Lewis, Di Forti, Davies and D'Souza2017) suggest a causal link between cannabis use and psychotic disorder, and evidence supports a dose–response association (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019; Marconi, Di Forti, Lewis, Murray, & Vassos, Reference Marconi, Di Forti, Lewis, Murray and Vassos2016). Frequent cannabis use and use of high-potency types have been linked to variations in the incidence of psychotic disorder across Europe (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019; Gonçalves-Pinho, Bragança, & Freitas, Reference Gonçalves-Pinho, Bragança and Freitas2020; Hjorthøj, Posselt, & Nordentoft, Reference Hjorthøj, Posselt and Nordentoft2021; Rognli et al., Reference Rognli, Taipale, Hjorthøj, Mittendorfer-Rutz, Bramness, Heiberg and Niemelä2023), North America (Callaghan et al., Reference Callaghan, Sanches, Murray, Konefal, Maloney-Hall and Kish2022; Moran, Tsang, Ongur, Hsu, & Choi, Reference Moran, Tsang, Ongur, Hsu and Choi2022), and in the Global South (Lee Pow et al., Reference Lee Pow, Donald, di Forti, Roberts, Weiss, Ayinde and Hutchinson2023). These findings have raised the important issue of whether increased consumption particularly of high-potency cannabis will lead to an increase in the incidence of psychosis (Murray & Hall, Reference Murray and Hall2020).
Patterns of cannabis use such as lifetime cannabis use (never/ever used) and cannabis use disorder (CUD) are influenced by genetic factors (Agrawal & Lynskey, Reference Agrawal and Lynskey2006; Pasman et al., Reference Pasman, Verweij, Gerring, Stringer, Sanchez-Roige, Treur and Ong2018). Heritability from twin studies is approximately 45% for lifetime cannabis use and between 51% and 70% for CUD (Kendler et al., Reference Kendler, Ohlsson, Maes, Sundquist, Lichtenstein and Sundquist2015; Verweij et al., Reference Verweij, Zietsch, Lynskey, Medland, Neale, Martin and Vink2010). Narrow-sense heritability, based on estimates from single-nucleotide polymorphisms (SNPs) only, is estimated at 11% for lifetime cannabis use and 12% for CUD (Johnson et al., Reference Johnson, Demontis, Thorgeirsson, Walters, Polimanti, Hatoum and Clarke2020; Pasman et al., Reference Pasman, Verweij, Gerring, Stringer, Sanchez-Roige, Treur and Ong2018). Genome-wide association studies (GWASs) have also shown a significant genetic correlation between lifetime cannabis use or CUD and schizophrenia (Demontis et al., Reference Demontis, Rajagopal, Thorgeirsson, Als, Grove, Leppälä and Reginsson2019). Moreover, polygenic risk scores (PRSs) for schizophrenia have been reported to explain a small but significant proportion of the variance in lifetime cannabis use, quantity of cannabis used (Power et al., Reference Power, Verweij, Zuhair, Montgomery, Henders, Heath and Martin2014), and CUD (Demontis et al., Reference Demontis, Rajagopal, Thorgeirsson, Als, Grove, Leppälä and Reginsson2019). Previous studies have shown that individuals at high risk for psychotic disorder (Vadhan, Corcoran, Bedi, Keilp, & Haney, Reference Vadhan, Corcoran, Bedi, Keilp and Haney2017) and/or with a known family for psychosis (Henquet, Murray, Linszen, & van Os, Reference Henquet, Murray, Linszen and van Os2005), are more vulnerable to the psychotogenic effect of cannabis use (Verweij et al., Reference Verweij, Abdellaoui, Nivard, Cort, Ligthart, Draisma and Hottenga2017).
We have used data on patterns of cannabis use (frequency of use and, where available, potency of the type used) along with genotype data from two studies: the European Network of National Schizophrenia Networks Studying Gene–Environment Interactions (EU-GEI) case-control study and the UK Biobank. We used both datasets to investigate the following questions: (1) is schizophrenia liability, as measured by the PRS, associated with lifetime cannabis use and/or patterns of cannabis use in population controls and in subjects with a diagnosis of psychotic disorder? (2) What are the independent and combined effects of schizophrenia PRS and cannabis use on odds of psychotic disorder? (3) To what extent does adding schizophrenia PRS data to information on patterns of cannabis use improve the identification of those heavy cannabis users who will develop psychotic disorder?
Methods
Samples
We analyzed data from two independent studies: first, we used a first-episode psychosis case-control sample, the EU-GEI, a multi-center case-control study of the genetic and environmental determinants of psychotic disorders. First-episode psychosis patients (FEPp) and population-based controls were recruited between May 2010 and April 2015 in 17 catchment areas in England, France, the Netherlands, Italy, Spain, and Brazil (Jongsma et al., Reference Jongsma, Gayer-Anderson, Lasalvia, Quattrone, Mulè, Szöke and Tarricone2018). Ethical approval was provided by relevant research ethics committees in each of the study sites.
Second, we conducted a comparative analysis using data from UK Biobank, a population-based study, including over 500 000 UK-based participants. The UK Biobank study was approved by the North-West Research Ethics Committee (ref 06/MREC08/65) in accordance with the Helsinki Declaration of 1975. All participants of both UK Biobank and EU-GEI provided written informed consent.
Participants
EU-GEI study
FEPp were included if (a) aged 18–64 years and (b) resident within the study areas at the time of their first presentation, and received a diagnosis of psychosis (ICD-10 F20-29); further details are provided in the online Supplementary methods and previous publications (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019; Quattrone et al., Reference Quattrone, Reininghaus, Richards, Tripoli, Ferraro, Quattrone and Gayer-Anderson2021). All cases interviewed received a research-based diagnosis (McGuffin, Farmer, & Harvey, Reference McGuffin, Farmer and Harvey1991; Quattrone et al., Reference Quattrone, Di Forti, Gayer-Anderson, Ferraro, Jongsma, Tripoli and Berardi2019). FEPp were excluded if (a) previously treated for psychosis, (b) they met criteria for organic psychosis (ICD-10: F09), or for a diagnosis of transient psychotic symptoms resulting from acute intoxication (ICD-10: F1X.5). Controls were excluded if they had received a diagnosis of, and/or treatment for, psychotic disorder.
UK Biobank
Subjects aged 40–70 years were recruited from 22 UK assessment centers. This study has been described in detail previously, see Bycroft et al. (Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018) and Bycroft et al. (Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018). All necessary demographic, medical, and genetic data were downloaded from UK Biobank. Cases were defined as any participant with either a recorded diagnosis of psychotic disorder, identified through recorded ICD-10 data (codes F20–F29), or who self-reported ‘schizophrenia’ or ‘Any other type of psychosis or psychotic illness’ as part of the online Mental Health Questionnaire (MHQ) (in response to the question: ‘Have you been diagnosed with one or more of the following mental health problems by a professional, even if you don't have it currently?’). Participants without psychosis were defined as any UK Biobank participant who had no reported psychotic disorder or previous treatment with an antipsychotic. We compared the baseline demographic data as well as information on the prescription of antipsychotics to consider the differences between cases defined by ICD-10 criteria and through self-report (see online Supplementary materials).
Sociodemographic variables
EU-GEI study
Data on age and sex were collected using the Medical Research Council Sociodemographic Schedule modified version (Mallett, Leff, Bhugra, Pang, & Zhao, Reference Mallett, Leff, Bhugra, Pang and Zhao2002).
UK Biobank
Information on age and sex was collected at recruitment when participants provided sociodemographic details (Bycroft et al., Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018).
Measures of cannabis use
Data on patterns of cannabis use were collected from the EU-GEI study using the modified Cannabis Experience Questionnaire further updated (CEQEU-GEI) (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019). The following measures of cannabis use were recorded: (1) age at first use of cannabis; (2) lifetime frequency of use, and (3) the potency of the cannabis used (Di Forti et al., Reference Di Forti, Marconi, Carra, Fraietta, Trotta, Bonomo and Russo2015). Potency was estimated as described in Di Forti et al. (Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019) using published data on the types of cannabis available and its potency from each of the sites included in this paper (see online Supplementary materials for a detailed discussion on this variable) (Brisacier et al., Reference Brisacier, Cadet-Taïrou, Díaz Gómez, Gandilhon, Le Nézet and Lemenier-Jeannet2015; de Oliveira, Voloch, Sztulman, Neto, & Yonamine, Reference de Oliveira, Voloch, Sztulman, Neto and Yonamine2008; European Monitoring Centre for Drugs and Drug and Addiction, 2016; Niesink, Rigter, Koeter, & Brunt, Reference Niesink, Rigter, Koeter and Brunt2015; Potter, Clark, & Brown, Reference Potter, Clark and Brown2008; Potter, Hammond, Tuffnell, Walker, & Di Forti, Reference Potter, Hammond, Tuffnell, Walker and Di Forti2018; Zamengo, Frison, Bettin, & Sciarrone, Reference Zamengo, Frison, Bettin and Sciarrone2014). We used the lifetime frequency of use and the cannabis potency variables to build the ‘frequency-type composite cannabis use measure’ that we previously found (Di Forti et al., Reference Di Forti, Marconi, Carra, Fraietta, Trotta, Bonomo and Russo2015) and replicated (Murray et al., Reference Murray, Englund, Abi-Dargham, Lewis, Di Forti, Davies and D'Souza2017) to be a strong predictor of psychotic disorder independently of other drugs of abuse, age, gender, ethnicity, site, and level of education. Study participants reported in their language the name of the type of cannabis used. Low-potency cannabis was defined as cannabis with a tetrahydrocannabinol (THC) concentration of less than 10% (THC < 10%), and high-potency cannabis was defined as THC concentration as greater or equal to 10% (THC ⩾ 10%).
In the UK Biobank sample, three specific questions on cannabis use were recorded as part of the online MHQ. We used these data to identify those subjects that had (a) never used cannabis, (b) used cannabis at least once, (c) used cannabis weekly at some stage, and (d) used cannabis daily at some stage. Data on potency or age of first use were not captured.
Choice of primary outcome measure
The primary outcome measure is defined as case status. In both cohorts, cases were included if they had a diagnosis of psychosis defined as ICD-10 codes F20–29. In the UK Biobank replications sample, we also included participants who self-reported a psychosis diagnosis. A comparison between those cases with a defined ICD-10 diagnosis and self-report only is provided in the online Supplementary materials. The measures of cannabis use as described above were chosen based on available data in our two cohorts.
Genotyping
Genotyping and imputation of EU-GEI and UK Biobank subjects has been described previously (Bycroft et al., Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018; Quattrone et al., Reference Quattrone, Reininghaus, Richards, Tripoli, Ferraro, Quattrone and Gayer-Anderson2021). Briefly, EU-GEI samples were genotyped at the MRC Centre for Neuropsychiatric Genetics and Genomics in Cardiff (UK) using a custom Illumina HumanCoreExome-24 BeadChip genotyping array covering 570 038 genetic variants (Quattrone et al., Reference Quattrone, Reininghaus, Richards, Tripoli, Ferraro, Quattrone and Gayer-Anderson2021). Genotyping for UK Biobank participants was undertaken using the Affymetrix UK BiLEVE Axiom array (used for the first ~50 000 participants) and the Affymetrix UK Biobank Axiom Array (~450 000 participants) (Bycroft et al., Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018). The Haplotype Reference Consortium (The Haplotype Reference Consortium, 2016) and the UK10K consortium (Huang et al., Reference Huang, Howie, McCarthy, Memari, Walter, Min and Zheng2015; UK10K Consortium, 2015) were used as imputation panels. Relatedness between participants was assessed using kinship scores provided by UK Biobank. One of each related pair (KING r 2 > 0.044) (Manichaikul et al., Reference Manichaikul, Mychaleckyj, Rich, Daly, Sale and Chen2010) was removed using the GreedyRelated algorithm, which prioritizes including cases (Choi, Reference Choi2020). For both samples, we calculated genetic principal components to assess population stratification and used these to assign genetic ancestry using the Genopred pipeline, using 1000 Genomes data as the reference populations (1000 Genomes Project Consortium, 2010; Reference PainPain; Pain et al., Reference Pain, Glanville, Hagenaars, Selzam, Fürtjes, Gaspar and Lewis2021). Participants who were not assigned to an ancestry group were excluded from our analyses (see online Supplementary materials).
PRS calculation
PRSs were calculated separately for different ancestry groups to account for population-specific differences in linkage disequilibrium and allele frequency. PRSs for schizophrenia were generated for participants of European ancestry (EUR) and East Asian ancestry (EAS) using PRS-CS (Ge, Chen, Ni, Feng, & Smoller, Reference Ge, Chen, Ni, Feng and Smoller2019) and Plink (Purcell et al., Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira, Bender and Daly2007), based on GWAS summary statistics from the Schizophrenia Working Group of the Psychiatric Genomics Consortium (PGC) wave three (with EU-GEI samples excluded) (Trubetskoy et al., Reference Trubetskoy, Pardiñas, Qi, Panagiotaropoulou, Awasthi, Bigdeli and Hall2022). For individuals of African ancestry PRSs were generated using PRS-CSx (Lee, Goddard, Wray, & Visscher, Reference Lee, Goddard, Wray and Visscher2012; Lewis & Vassos, Reference Lewis and Vassos2017), which works in the same way as PRS-CS but allows for inclusion of multiple sets of summary statistics from different populations to improve predictive power. To calculate the AFR PRS, we used the EUR summary statistics described previously in combination with summary statistics from the Genomic Psychiatry Cohort (Bigdeli et al., Reference Bigdeli, Genovese, Georgakopoulos, Meyers, Peterson, Iyegbe and Pato2020). PRS for CUD were calculated using EUR and AFR GWAS summary statistics (Levey et al., Reference Levey, Galimberti, Deak, Wendt, Bhattacharya, Koller and Gupta2023) using PRS-CS and PRS-CSx respectively, as described above. Once calculated, these ancestry-specific scores were combined into a single column and all groups analyzed together. Due to the large dominance of the EUR ancestry participants in both cohorts, we were unable to perform population-specific analyses in the EAS or AFR group. Each PRS was standardized to mean of 0 and standard deviation of 1 (Lewis & Vassos, Reference Lewis and Vassos2017).
Statistical analysis
Adjusted logistic regression models were run to estimate: (1) odds of cannabis use for each unit increase in schizophrenia PRS and (2) the independent and combined effect of the selected measures of cannabis use and the schizophrenia PRS on the odds ratio (OR) for psychotic disorder. We fitted multiplicative interaction terms to the logistic models to test if schizophrenia PRS modified the effect of cannabis use on the OR for psychotic disorder. Interaction analyses were conducted in EUR participants only, due to limited numbers of non-EUR participants. All regression models were adjusted for: the first 10 principal components, recruitment site, age, sex, and tobacco smoking (as defined in our previous publication; Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019). The latter was added due to the clinical and genetic overlap among schizophrenia, tobacco use, and cannabis use (Johnson et al., Reference Johnson, Demontis, Thorgeirsson, Walters, Polimanti, Hatoum and Clarke2020). We conducted additional analyses adjusting for CUD PRS, to consider the impact of underlying genetic risk for CUD on patterns of cannabis use and schizophrenia case status. For each model performed, we calculated Nagelkerke's R 2 to consider model fit and converted this observed scale R 2 measure to a liability scale measure (Lee et al., Reference Lee, Goddard, Wray and Visscher2012). We used 0.1 as an estimate for the population level lifetime risk for psychosis. To mitigate potential statistical confounding in the interaction models, we carried out additional analyses adding covariate × environment and covariate × gene interaction terms to the model, as recommended by previous publication (see online Supplementary materials section) (Keller, Reference Keller2014). We calculated the positive-predictive value (PPV) for both the schizophrenia and CUD PRS, using the pROC package in R (Robin et al., Reference Robin, Turck, Hainard, Tiberti, Lisacek, Sanchez and Müller2011). All analyses were conducted using R version 4.1.1 (R Core Team., 2021).
Results
Baseline characteristics of study participants
EU-GEI
In total, 1130 FEPp and 1497 population controls consented to take part. The total sample with available genetic data and data on cannabis use was 945 controls and 647 cases (total 1592). We defined the following ancestry groups to build the schizophrenia PRS by ancestry: N EUR = 1262, N SAS = 35, N AFR = 192, N EAS = 18, N-undefined = 409; the latter were excluded. All analyses reported here are the results of the full sample. Given the differences in predictive power of polygenic scores across ancestry groups, we report a EUR-only sensitivity analysis in the online Supplementary materials.
The final EU-GEI sample consisted of 405 FEPp (cases) and 693 controls (see recruitment flow chart in the online Supplementary materials). As highlighted in Table 1, cases were younger and more likely to be men than controls. Cases were also more likely to have tried cannabis, to have first used it at age 15 years old or younger, and to have used it daily. Cases were also more likely to have used more potent types and to have used them daily than controls.
Table 1. Differences between cases and controls in across sociodemographic factors and patterns of cannabis use
df, degrees of freedom; s.d., standard deviation; THC, tetrahydrocannabinol.
a Cases for EU-GEI study defined as first-episode psychosis patients; cases for UK Biobank defined as either bSchizophrenia or psychosis based on self-report and/or ICD-10 code or cany major psychiatric disorder defined by ICD-10 code.
d EU-GEI study recorded data on age at first use, UK Biobank recorded data on age at last use.
e In EU-GEI data, eight cases (1.98%) did not provide this information. In UK Biobank data 864 (0.63%) controls and six (0.81%) cases did not provide this information.
f Data on potency not recorded for UK Biobank study.
g 23 controls (3.32%) and 31 (7.65%) cases did not provide this data.
h 24 controls (3.46%) and 33 (8.15%) cases did not provide this data.
UK Biobank
Data for a total of 455 538 UK Biobank participants with high-quality genetic data were downloaded. Of these, approximately 32% also had responded to the MHQ and thus provided data on previous cannabis use, giving us a final working sample of 145 244 (N EUR = 143 600, N AFR = 1177, N EAS = 527). This final sample consisted of 743 psychosis cases, as defined by a combination of ICD-10 data and self-report, and 142 857 participants without psychosis (additional detail on case ascertainment for the UK Biobank replication study provided in the online Supplementary materials).
PRS distribution
The schizophrenia PRS was on average higher in FEPp than in controls (Fig. 1): EU-GEI case mean schizophrenia PRS = 0.40, s.d. = 1.05; controls mean schizophrenia PRS = −0.17, s.d. = 1.00; t = −10.86, df = 1349.3; p = 2.2 × 10−26. There were more controls in the schizophrenia PRS quintile 1 compared to cases, while the opposite was true in quintile 5: controls quintile 1 = 251/319 (78.68%); cases quintile 1 = 68/319 (21.32%); controls quintile 5 = 136/318 (42.77%); cases quintile 5 = 182/318 (57.23%) (χ2 = 127.33, df = 4, p = 1.45 × 10−26). The CUD PRS was also higher in EU-GEI first-episode psychosis cases compared to controls: EU-GEI case mean PRS = 0.31 ± 1.1, controls mean CUD PRS = −0.1 ± 0.98, p diff = 7.63 × 10−14). We observed similar patterns in UK Biobank data (see online Supplementary materials).
Figure 1. Distribution of schizophrenia PRSs shows an overall increase in PRS for cases compared to controls in the EU-GEI cohort (mean PRS cases = 0.40 ± 1.05, mean PRS controls = 0.17 ± 1.00, p = 2.2 × 10−26).
Variance explained and PPV of schizophrenia and CUD PRS
We calculated pseudo-R 2 statistics (Nagelkerke and liability scale, adjusted for the sample prevalence) by each of our predictors. In the EU-GEI cohort, a model including schizophrenia PRS, site, sex, and 10 principal components explained 25.9% of the variation in case-control status (12.8% on the liability scale). When tobacco smoking (more or less than 10 cigarettes per day) was included, the variance explained was 34.2% on the observed and 18.0% on the liability scale. Adding daily cannabis use increased this to 50.3% on the observed and 29.8% on the liability scale. This was not significantly increased by adding age at first use (R 2obs = 50.3%, R 2liab = 30.1%) but was increased by adding the use of high-potency cannabis (THC > 10%) (R 2obs = 55.5%, R 2liab = 34.7%). Adding the CUD PRS did not improve the model (R 2obs = 56.0%, R 2liab = 34.7%) (online Supplementary S-Fig. 10).
In the UKB cohort, a model including schizophrenia PRS, site, sex, age, and 10 principal components explained 1.3% (2.2% on the liability scale). When tobacco smoking (more or less than 10 cigarettes per day) was included, the variance explained was 1.7% on the observed and 2.9% on the liability scale. Adding daily cannabis use increased this to 29.3% on the observed and 48.3% on the liability scale (online Supplementary S-Fig. 11). CUD PRS was not associated with schizophrenia status and did not increase the variance explained 29.3% (48.3% on the liability scale).
In our EU-GEI control sample alone, schizophrenia PRS and 10 principal components explained a small but non-significant proportion of the variance between those who never used cannabis and (a) those who had tried at least once (lifetime use R 2 = 7.86%; p schizophrenia PRS = 0.08), (b) those who had started at age 15 or younger (R 2 = 3.89%; p schizophrenia PRS = 0.3), (c) having used it daily (R 2 = 3.14%; p schizophrenia PRS = 0.21), and (d) using high-potency types (R 2 = 18.38%; p = 0.41). In our EU-GEI control sample alone, CUD PRS and 10 principal components explained a significant proportion of the variance between those who never used cannabis and (a) having tried it at least once (lifetime use R 2 = 8.31%; p schizophrenia PRS = 0.03), and (b) having used it daily (R 2 = 4.52%; p schizophrenia PRS = 0.01) but was not associated with use of high-potency types (R 2 = 18.56%; p = 0.62) or having started at age 15 or younger (R 2 = 4.08%; p schizophrenia PRS = 0.96).
In the EU-GEI cohort, we calculated the PPV for assigning psychosis case/control status to be 0.65 for schizophrenia PRS and 0.63 for CUD PRS.
Does schizophrenia PRS predict cannabis initiation and/or patterns of cannabis use?
Regression adjusted for age, sex, tobacco smoking, recruitment site, and for the 10 principal components showed that schizophrenia PRS was not associated with cannabis initiation (lifetime cannabis use yes/no) among cases or controls from EU-GEI (EU-GEI cases: OR = 0.9; 95% confidence interval [CI] 0.63–1.3; p = 0.58; EU-GEI controls: OR = 1.14; 95% CI 0.9–1.44; p = 0.28). In addition, schizophrenia PRS did not explain how frequently either cases or controls used cannabis, including when we specifically compared never use with daily use: (EU-GEI cases: OR = 0.78; 95% CI 0.53–1.15; p = 0.22; EU-GEI controls: OR = 1.25; 95% CI 0.82–1.94; p = 0.31). Schizophrenia PRS also did not predict age at first use among either cases or controls: (EU-GEI cases: β = 0.08; s.e. = 0.36, p = 0.83; EU-GEI controls: β = −0.24; s.e. = 0.34, p = 0.48) (Table 2).
Table 2. ORs for varying measures of cannabis use in EU-GEI and UK Biobank cohorts
All models compared to never users as reference group, and adjusted for age, sex, recruitment site, tobacco smoking, and 10 principal components.
Regression adjusted for age, sex, tobacco smoking, recruitment site, and the 10 principal components showed that CUD PRS did predict cannabis initiation (lifetime cannabis use yes/no) among cases or controls from EU-GEI (EU-GEI cases: OR = 1.57; 95% CI 1.03–2.4; p = 0.03; EU-GEI controls: OR = 1.36; 95% CI 1–1.86; p = 0.05). In addition, CUD PRS was significantly associated with weekly use among controls only, and daily use among cases only: EU-GEI cases OR (weekly use) = 2.22; 95% CI 1.31–3.85; p = 0.004; EU-GEI controls OR (weekly use) = 0.91; 95% CI 0.49–1.68; p = 0.76; EU-GEI cases OR (daily use) = 1.26; 95% CI 0.8–1.98; p = 0.31; EU-GEI controls OR (daily use) = 1.84; 95% CI 1.05–3.28; p = 0.04. CUD PRS did not predict age at first use among cases or controls: (EU-GEI cases: β = 0; s.e. = 0.42, p = 1; EU-GEI controls: β = −0.12; s.e. = 0.43, p = 0.77) (online Supplementary S-Tables 12, 13, S-Fig. 13).
We saw similar patterns among UK Biobank cases. However, we found an association of small magnitude between schizophrenia PRS and both lifetime cannabis use and frequency of use in participants without psychosis (UK Biobank participants without psychosis lifetime use OR = 1.08; 95% CI 1.07–1.09; p = 9.72 × 10−23; UK Biobank participants without psychosis weekly use OR = 1.09, 95% CI 1.06–1.13 p = 6.95 × 10−7; UK Biobank participants without psychosis daily users OR = 1.14, 95% CI 1.07–1.20, p = 1.61 × 10−5) (Table 2). Including the CUD PRS in these models led to a reduction in the magnitude of this effect (online Supplementary S-Tables 13, 14). For lifetime use, the effect size for schizophrenia PRS reduced by 39%, but the association with lifetime use remained significant. For daily use, the effect size for schizophrenia PRS reduced by 82% and was no longer significantly associated with lifetime cannabis use. By comparison, a model including only CUD PRS was less impacted by the addition of schizophrenia PRS, with a reduction of effect size for CUD PRS of 23 and 3% for lifetime and daily use respectively (online Supplementary S-Tables 14, 15, S-Fig. 14).
The independent and combined effect of schizophrenia PRS and pattern of cannabis use on the OR for psychotic disorder
In the EU-GEI sample, both adjusted and unadjusted regression for schizophrenia PRS showed that lifetime cannabis use was associated with an increased risk for psychotic disorder (adjusted OR (inc. schizophrenia PRS) = 1.63; 95% CI 1.26–2.12; p = 2.32 × 10−4; unadjusted OR = 1.67; 95% CI 1.3–2.15; p = 6.53 × 10−5). Weekly cannabis use was also significantly associated with increased odds of psychosis (adjusted OR = 2.31; 95% CI 1.52–3.51; p = 8.72 × 10−5; unadjusted OR = 2.42; 95% CI 1.62–3.63; p = 1.67 × 10−5), and the strongest association was with daily cannabis (OR = 3.7; 95% CI 2.59–5.35; p = 1.53 × 10−12; unadjusted OR = 3.7; 95% CI 2.62–5.28; p = 2.61 × 10−12) (Table 3). In models additionally adjusted for CUD PRS, we demonstrate that CUD PRS is also associated with case-control status in the EU-GEI cohort, with little evidence that the CUD PRS confounds the results for the schizophrenia PRS (difference in schizophrenia PRS OR for models with and without CUD PRS <10% in all cases) (online Supplementary S-Table 17, S-Fig. 15). These results were replicated in the UK Biobank sample (Table 3, although in all analyses the CUD PRS was not associated with psychosis status) (online Supplementary S-Table 18). We fitted interaction terms to the logistic models to consider the putatively modifying effect of schizophrenia PRS on the impact of cannabis use on the OR for psychotic disorder and found no evidence of an interaction (Table 3, online Supplementary S-Table 9; for additional discussion of interaction models in both cohorts see online Supplementary materials and S-Tables 10, 11).
Table 3. Independent and combined effects of schizophrenia PRS and cannabis use measures on risk for psychosis in the EU-GEI and UK Biobank
Panel A illustrates the main effect of schizophrenia PRS and frequency of cannabis use, independent of each other, on the risk of psychosis. Panel B illustrates the results for models including schizophrenia PRS and frequency of cannabis use in interaction. All models adjusted for age, sex, recruitment site, tobacco smoking, and 10 principal components.
We observed that those who used either high- or low-potency cannabis on a daily basis had an increase in the risk for psychotic disorder compared to never users, independently of their schizophrenia PRS and after adjusting for age, sex, site, and 10 principal components, with the greatest magnitude of effect observed in high-potency daily users (low-potency daily OR = 3.02; 95% CI 1.80–5.08; p = 3.13 × 10−5; high-potency daily OR = 5.09; 95% CI 3.08–8.43; p = 3.21 × 10−10) (online Supplementary S-Tables 5, 6). We fitted interaction terms to the logistic models and observed no evidence of a modifying effect of schizophrenia PRS and cannabis use on the OR for psychotic disorder, although there did appear to be a trend increase in psychosis risk across all levels of use (Fig. 2).
Figure 2. Probability of psychosis case status within EU-GEI cohort as schizophrenia PRS increases, across seven levels of cannabis potency–frequency composite measure.
Discussion
This study is the first to provide estimates of risk for psychotic disorder by the joint modeling of cannabis use (frequency) and common variant liability to schizophrenia. In keeping with previous studies (Di Forti et al., Reference Di Forti, Marconi, Carra, Fraietta, Trotta, Bonomo and Russo2015, . Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019; Marconi et al., Reference Marconi, Di Forti, Lewis, Murray and Vassos2016), our analyses reveal an association between case-control status in both cohorts. Lifetime cannabis use was associated with increased odds of psychosis, and the magnitude of this effect was greater when considering those users who consumed cannabis more regularly (weekly and daily use). These results remained consistent when we adjusted for the schizophrenia PRS, demonstrating that genetic risk is independent of the environmental risk factor that is cannabis use. In the EU-GEI sample, we also see independent effects of high-potency cannabis use when adjusting for schizophrenia PRS. When we fitted interaction terms to these models, we found little evidence for a modifying effect of schizophrenia PRS.
In the EU-GEI cohort, we found that schizophrenia PRS was not associated with an individual's propensity to try cannabis or, among users, with the frequency, or other patterns of use. These findings are consistent with a recent cross-sectional study of patients with established psychosis (using a different EU-GEI cohort of chronic schizophrenia patients), which showed no evidence of correlation between schizophrenia PRS and regular cannabis use (Guloksuz et al., Reference Guloksuz, Pries, Delespaul, Kenis, Luykx, Lin and van Os2019).
Among the UK Biobank sample of participants without psychosis we found that schizophrenia PRS was associated with patterns of cannabis use (explaining less than 1% of the variance in lifetime cannabis use and daily use). When we included the CUD PRS in these models the effect size for schizophrenia PRS was reduced by 39% and 82% for lifetime and daily use, respectively, indicative of substantial confounding by the CUD PRS. We also observed evidence of confounding by the schizophrenia PRS on the CUD PRS for lifetime use, but not daily use. Given more frequent cannabis use confers greater risk for psychosis, arguments of reverse causality (schizophrenia leading to heavy cannabis use) would be supported by evidence of an association between schizophrenia PRS and cannabis use patterns. Here, we show that while a higher genetic risk for schizophrenia may increase the chance of ever trying cannabis, it may have less bearing on the likelihood of becoming a heavy user. Nonetheless, there does seem to be some independent effects of schizophrenia genetic risk on cannabis use, which should be interrogated in other population-based cohorts.
Future research should implement robust causal inference methods to interrogate this reported association more thoroughly. Our study, which utilized an SNP-inclusive method to calculate polygenic scores, is not designed to confirm or exclude the possibility of a causal association. Previous studies investigating this relationship using genetic data have found conflicting results (Elkrief et al., Reference Elkrief, Lin, Marchi, Afzali, Banaschewski and Bokde2023; Gillespie & Kendler, Reference Gillespie and Kendler2021; Hjorthoj et al., Reference Hjorthoj, Uddin, Wimberley, Dalsgaard, Hougaard, Borglum and Nordentoft2021; Jones et al., Reference Jones, Hammerton, McCloud, Hines, Wright, Gage and Zammit2022; Verweij et al., Reference Verweij, Abdellaoui, Nivard, Cort, Ligthart, Draisma and Hottenga2017; Wainberg, Jacobs, di Forti, & Tripathy, Reference Wainberg, Jacobs, di Forti and Tripathy2021).
More recently, Mendelian randomization analyses have taken advantage of the available genetic data on both schizophrenia and cannabis use, but so far they have produced contradictory findings, perhaps due to variation in the cannabis use instrumental variables used (Cheng et al., Reference Cheng, Parker, Karadag, Koch, Hindley, Icick and Bahrami2023; Gage et al., Reference Gage, Jones, Burgess, Bowden, Davey Smith, Zammit and Munafo2017; Pasman et al., Reference Pasman, Verweij, Gerring, Stringer, Sanchez-Roige, Treur and Ong2018; Vaucher et al., Reference Vaucher, Keating, Lasserre, Gan, Lyall, Ward and Paré2018). The recent well-powered GWAS on CUD found evidence of a bi-directional association between schizophrenia and CUD, with a larger magnitude of effect from CUD to schizophrenia (Levey et al., Reference Levey, Galimberti, Deak, Wendt, Bhattacharya, Koller and Gupta2023).
In a commentary, Gillespie and Kendler (Reference Gillespie and Kendler2021) discuss three possible scenarios, which might explain the complexity of the relationship between cannabis use and psychotic disorders. They point out that, for instance, cannabis use might be partly causal and partly confounded by genetic/familial effects and/or reverse causation, suggesting that ‘a causal role’ of cannabis use in psychotic disorders might not exclude a non-causal genetic confounding and vice versa. They indicate that more clarity might come from studies investigating changes in incidence rates of schizophrenia in places where rates of cannabis use have changed because of its legalization. These data are beginning to be published and further support a causal role of cannabis use in psychotic disorders (Callaghan et al., Reference Callaghan, Sanches, Murray, Konefal, Maloney-Hall and Kish2022; Gonçalves-Pinho et al., Reference Gonçalves-Pinho, Bragança and Freitas2020; Hjorthøj et al., Reference Hjorthøj, Posselt and Nordentoft2021).
Another important issue to consider is the relative importance of the genetic v. environmental contribution to both cannabis use and schizophrenia. This appears to vary across samples and is likely to depend on (a) the study sample size; a larger sample size is more likely to enable the detection of small effects and (b) the availability of cannabis and acceptability of its use during the period of data collection. For example, a recent study from the United States used data from twin pairs discordant for residential address with one twin living in a state where cannabis use was legal and the other twin living in a state where it was illegal. Their findings indicate that genetic correlations on frequency of cannabis use were significantly lower where cannabis use was legal compared to where it was illegal, suggesting the important influence of the social context (Zellers et al., Reference Zellers, Ross, Saunders, Ellingson, Anderson, Corley and McGue2023).
The EU-GEI sample is a unique sample with self-reported details on the type of cannabis used (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019), which allowed us to explore for the first time the relationship between schizophrenia PRS and use of high-potency cannabis; the availability of which is increasing worldwide and is associated with high rates of psychosis across Europe (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019). In our previous paper, we described a probabilistic sensitivity analysis, which showed that selection bias is unlikely to explain the reported findings on the observed magnitude of effect for daily cannabis use and use of high potency on the risk of psychotic disorder (Di Forti et al., Reference Di Forti, Quattrone, Freeman, Tripoli, Gayer-Anderson, Quigley and La Cascia2019). The incorporation of frequency and potency of cannabis use alongside schizophrenia PRS results in an improvement in R 2 values, indicating a significant improvement in the model's explanatory power. The delta R 2, while positive, was modest, perhaps due to the multifactorial risk profile for psychotic disorders.
It should be noted that the estimates of cannabis potency cannot account for differences in the THC concentration in individual samples. Our dichotomous measure of THC above or below 10% is conservative and likely to have resulted in an underestimate of the effects of cannabis potency on the risk for psychotic disorder. We were not able to assess potency in UK Biobank. The people in UK Biobank were aged at least 40 years old at recruitment which was from 2006 to 2010, and most participants reported their last use of cannabis was well before recruitment (online Supplementary Fig. 6). There has been a significant change in the type of cannabis available in the UK in recent decades, and consequently, it is plausible that the middle-aged cannabis users within UK Biobank had mainly used lower potency cannabis (Potter et al., Reference Potter, Clark and Brown2008, . Reference Potter, Hammond, Tuffnell, Walker and Di Forti2018). While this lack of potency data is a limitation, it is worth noting that we still observe strong effects for frequency of use in the UK Biobank sample. If the cannabis used by many of these participants was indeed lower potency, it would follow that we might observe even stronger effects where the same analyses to be conducted with participants recruited today.
This study must be considered in the context of some limitations. First, the cannabis measures were based on retrospectively collected self-reported information. In both cohorts, the data on cannabis use were collected as part of a questionnaire which did not refer to the association between cannabis use and psychosis. Biological data on potency and accurate levels of THC can be obtained from blood samples and there is a validation of biological data only when measuring current use up to few weeks rather than lifetime use, which is the measure of exposure we use in our analyses. Furthermore, previous research has clearly reported the reliability of self-reported measures on the frequency and potency of the cannabis used (Bharat et al., Reference Bharat, Webb, Wilkinson, McKetin, Grebely, Farrell and Clark2023; Buchan, Dennis, Tims, & Diamond, Reference Buchan, Dennis, Tims and Diamond2002; Di Forti et al., Reference Di Forti, Morgan, Dazzan, Pariante, Mondelli, Marques and Paparelli2009; Freeman et al., Reference Freeman, Morgan, Hindocha, Schafer, Das and Curran2014).
In addition, individuals of European ancestry are over-represented in these analyses. We opted to utilize PRS-CSx, which allows the inclusion of multiple GWAS summary statistics and linkage disequilibrium (LD) reference panels to improve prediction in non-EUR populations. However, the predictive power of these PRS is limited by the available discovery datasets. This is a common problem in genetic studies, and one that must be rapidly addressed to ensure scientific discoveries, especially those with potential clinical implications, are relevant to all populations (Duncan et al., Reference Duncan, Shen, Gelaye, Meijsen, Ressler, Feldman and Domingue2019; Fatumo et al., Reference Fatumo, Chikowore, Choudhury, Ayub, Martin and Kuchenbaecker2022; Peterson et al., Reference Peterson, Kuchenbaecker, Walters, Chen, Popejoy, Periyasamy and Duncan2019). The two samples included are very different, both in terms of recruitment strategy (case-control v. healthy volunteer) and the data collected, meaning that we could not carry out a direct replication across the two cohorts (Bycroft et al., Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and O'Connell2018; Gayer-Anderson et al., Reference Gayer-Anderson, Jongsma, Di Forti, Quattrone, Velthorst, De Haan and Tortelli2020). A recent study investigated the impact of selection bias in the UK Biobank and found evidence that it can impact genetic correlation and Mendelian randomization results between several traits (Schoeler et al., Reference Schoeler, Speed, Porcu, Pirastu, Pingault and Kutalik2023). This again indicates that replication of our findings is essential to draw firm conclusions applicable to a wider clinical and healthy control population.
Our analyses were adjusted for a range of demographic and genetic factors, aiming to account for potential confounding variables that could influence the association between schizophrenia PRS and cannabis use. One putatively confounding factor is the underlying genetic risk for CUD, which might drive cannabis use patterns and schizophrenia risk. We therefore additionally adjusted all models for a CUD PRS, to investigate evidence of confounding. In the EU-GEI sample, we show that CUD PRS is largely independently associated with schizophrenia risk and that the addition of this variable does not greatly impact the effect size for the schizophrenia PRS. This suggests that despite the known genetic overlap between the two traits, there is a degree of specificity in their association with schizophrenia risk. One speculative interpretation could be that a higher CUD PRS increases the likelihood of using cannabis, which in turn increases the risk for psychosis. If this were the case, we might expect to see evidence that adjusting for the CUD PRS reduces the effect of the measure of cannabis use frequency or potency, which was in fact not what we observed. Thus, we cannot rule out the possibility that some of the genetic factors that confer risk for CUD have a pleiotropic effect on schizophrenia pathogenesis.
Our work, along with multiple previous studies, confirms that cannabis use is much more common among people with psychosis than controls. We can, therefore, assume that a larger proportion of cases in the PGC and Genomic Psychiatry Cohort (GPC) GWAS on schizophrenia are cannabis users, relative to controls (Elkrief et al., Reference Elkrief, Lin, Marchi, Afzali, Banaschewski and Bokde2023). This remains a limitation of genetic studies of cannabis and schizophrenia and it could potentially lead to an inflated estimate of the true shared genetic liability between CUD and schizophrenia. Future analyses would be improved by tracking down and accounting for comorbid cases when building schizophrenia GWASs (Colbert & Johnson, Reference Colbert, Johnson, D'Souza, Castle and Murray2023). Finally, psychosis is a multifactorial disease with a wide number of established risk factors, not all of which can be accounted for in any single model. It remains possible that some of the findings detailed here could be explained, in part, by other factors such as comorbid disease, trauma, sociodemographic factors, or underlying genetic risk for other traits.
In conclusion, our findings indicate (a) heavy cannabis use remains a strong risk factor independent of schizophrenia genetic load and (b) as available samples increase for more well-powered and diverse GWASs on schizophrenia, schizophrenia PRS may be useful in identifying individuals most at risk for cannabis-associated psychosis (Pain & Lewis, Reference Pain and Lewis2022). Currently, schizophrenia PRS risk can only explain a small proportion of the risk (Power et al., Reference Power, Verweij, Zuhair, Montgomery, Henders, Heath and Martin2014). While this study did not set out to prove causality or specifically address the above-mentioned methodological controversies around genetic confounding, it clearly shows that cannabis users at all levels of schizophrenia PRS are more likely to belong to the FEP group compared with non-cannabis-using individuals. This is consistent with the epidemiological evidence which shows that cannabis is an important and modifiable risk factor for psychosis, as it has recently been outlined by the World Federation Society of Biological Psychiatry (D'Souza et al., Reference D'Souza, DiForti, Ganesh, George, Hall, Hjorthøj and Nguyen2022). Therefore, our findings provide information that public education campaigns could use toward the prevention of an increase in the rates of psychotic disorders (Murray & Hall, Reference Murray and Hall2020).
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291724002058.
Funding statement
The EU-GEI project was funded by the European Community's Seventh Framework Programme under grant agreement No. HEALTH-F2-2009-241909 (Project EU-GEI). The Brazilian study was funded by the São Paulo Research Foundation under grant number 2012/0417-0, Medical Research Council (MRC), National Institute for Health Research (NIHR) Biomedical Research Centre (BRC) at South London and Maudsley NHS Foundation Trust and King's College London and the NIHR BRC at University College London. Wellcome Trust (grant 101272/Z/12/Z). M. D. F., I. A.-Z., G. T., and E. S. were supported by MRC SRF Fellowship (MRC MR/T007818/1). E. S. was also supported by Lord Leverhulme's Charitable Trust and the Velvet Foundation and by the Medical Research Council (MRC) (MR/T007818/1). Dr Arango receives support from the Spanish Ministry of Science and Innovation, Instituto de Salud Carlos III (ISCIII), co-financed by the European Union, ERDF Funds from the European Commission, ‘A way of making Europe’, financed by the European Union – NextGenerationEU (PMP21/00051), PI19/01024. CIBERSAM, Madrid Regional Government (B2017/BMD-3740 AGES-CM-2), European Union Structural Funds, European Union Seventh Framework Program, European Union H2020 Program under the Innovative Medicines Initiative 2 Joint Undertaking: Project PRISM-2 (grant agreement No. 101034377), Project AIMS-2-TRIALS (grant agreement No. 777394), Horizon Europe, the National Institute of Mental Health of the National Institutes of Health under Award Number 1U01MH124639-01 (Project ProNET) and Award Number 5P50MH115846-03 (project FEP-CAUSAL), Fundación Familia Alonso, and Fundación Alicia Koplowitz. Funders were not involved in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript, and decision to submit the manuscript for publication.
Competing interests
M. Di Forti reports personal fees from Janssen, outside the submitted work. R. M. Murray reports personal fees from Janssen, Lundbeck, Sunovion, and Otsuka, outside of the submitted work. Dr Arango has been a consultant to or has received honoraria or grants from Acadia, Angelini, Biogen, Boehringer, Gedeon Richter, Janssen Cilag, Lundbeck, Medscape, Menarini, Minerva, Otsuka, Pfizer, Roche, Sage, Servier, Shire, Schering Plough, Sumitomo Dainippon Pharma, Sunovion, and Takeda. All the remaining authors have declared that there are no competing interests in relation to the subject of this study.
Open access
