Introduction

The lifetime prevalence of psychotic disorders is around 3% [1]. The associated individual, familial, social, and economic costs are vast. Psychotic disorders cause considerable distress to sufferers and their families and often lead to marked social dysfunction and exclusion. The economic costs are huge: in Europe, an estimated €94 billion per year [2], of which over half is due to the indirect costs of unemployment, lost productivity, and informal care [3]. The World Health Organisation estimated that in Western countries, the treatment and care of patients with a psychotic disorder range from 1.6 to 2.6% of total healthcare expenditures [4]. Further, individuals with a psychotic disorder are far more likely to have a physical health problem [5] and to die younger, by as much as 20 years on average, compared with the general population [6].

Our knowledge of the distribution and determinants of psychotic disorders has increased in recent years. The incidence varies by area (e.g. higher in some urban versus rural areas) [7, 8] and social group (e.g. higher in some minority ethnic groups) [9, 10] and, in addition to well-established genetic and neurodevelopmental risk factors [11, 12], there is now substantial evidence implicating several environmental risk factors [13], such as childhood adversity [14, 15] and cannabis use [16]. Pooled relative risks for these risk factors range between two and four, and population attributable risk fractions range between 20 and 35% [17, 18]. Further, there is accumulating evidence that these myriad risk factors interact in complex ways to increase risk of psychosis via effects on the dopaminergic system, dysregulation of which may be the biological process that underpins the formation of psychotic experiences.

However, there remain many gaps, inconsistencies, and unanswered questions, and recent work hints at different patterns of risk in different settings. For example, recent evidence has failed to show a universal association between city living and psychosis [19, 20]. To further add to this conundrum, Colodro-Conde et al. [21] found that the high prevalence of psychosis in some urban areas may be due to gene–environment selection, such that individuals with higher genetic loading for psychosis live in more densely populated areas. More generally, this points to a major limitation to our current knowledge of psychotic disorders: we know that environments affect onset and outcomes, but research so far has been conducted—with some important exceptions—in a remarkably small number of settings (i.e. select centres in the US, UK, and Australasia). Combined, these points emphasise the need for research in more diverse contexts to examine more nuanced hypotheses on the complex interplay between biology and environments in the aetiology of psychotic disorders.

Our knowledge of psychotic disorders is limited, in part, because of heterogeneity in methods, which limits our ability to compare findings across populations [22]. For example, differences in study design (i.e. case-register, versus cohort-based designs, versus first-contact studies), the age structures of populations at risk, case-identification procedures, diagnostic criteria, definitions and measurement of environmental factors, and analytic strategies have made cross-country comparisons difficult and likely obscured important clues to aetiology [23]. The only large-scale international comparative studies conducted to date are the World Health Organisation’s multi-country projects of the 1970s and 1980s, which compared the incidence and clinical and social characteristics of treated cases of psychoses from twelve diverse settings in ten countries using a standardised procedure for case identification and data collection [24]. However, since this landmark programme, there have been far-reaching economic and social changes (e.g. migration patterns, cannabis availability and use, and distribution of social risks) with conceivable impacts on the social epidemiology and aetiology of the psychoses. Moreover, studies of environment–gene interactions in psychotic disorders are rare and have typically involved small samples, with limited phenotyping and limited assessment of environmental factors [21, 25, 26].

The EU-GEI programme was established to address these gaps and limitations [27]. EU-GEI is a multi-national research collaboration that was funded for 5 years (1 May 2010–30 April 2015). It consisted of 11 Work Packages (see Supplementary Table S1). This paper profiles the incidence and case–control programme of work (Work Package 2), which comprises the largest multi-site study of psychotic disorders ever conducted. In this paper, we describe the objectives and main aspects of the study.

Objectives

The overall goal of the present work package was to investigate the role of multiple environmental and genetic risk factors, and their interactions, in the development of psychotic disorders. Specifically, our aims were (1) to investigate the impact of hypothesized environmental exposures, measured at individual and area levels, on (a) risk of psychotic disorders, and (b) high rates of disorder in urban areas and in migrant and minority ethnic groups; and (2) to examine hypothesized (a) gene × environment interactions (GxE), and (b) environment × environment interactions (ExE) across the life course.

Methods

Study design

The data resource comprises a multi-site population-based incidence and case–control sample of cases with a first episode of psychosis [International Classification of Diseases (ICD)-10 diagnoses F20–29 and F30–33] and controls drawn from tightly defined catchment areas in 17 sites in 6 countries (England, The Netherlands, France, Spain, Italy, and Brazil; see Fig. 1). The sites were purposefully selected to include a mix of urban and rural areas, with varying proportions from minority ethnic groups (see Table 1).

Fig. 1
figure 1

Map of EU-GEI settings for the incidence and case–control Work Package

Full size image
Table 1 Recruitment period and duration, and number of incidence and consented cases and controls, per site
Full size table

Sample

Recruitment and data collection were conducted over a 5-year period between 2010 and 2015 (Table 1). We also added data from the Veneto region, Italy, collected as part of an earlier study [the Psychosis Incident Cohort Outcome Study (PICOS); 2005–2007], but with sufficiently similar methods to be pooled with that collected for this study. The incidence sample comprised 2774 individuals with a first episode of psychosis. Of these, 1519 were approached, and 1130 were consented and assessed (41% of the total incidence sample). Reasons for non-participation among cases who were approached were refusal to participate, language barriers, and exclusion after consenting as they did not meet the age inclusion criteria. In addition, 1497 controls were recruited and assessed.

Statistical power

Our sample of 1130 cases and 1497 controls has high statistical power to test our primary study hypotheses, even after accounting for missing data and for the current necessity of restricting genetic analyses to individuals of non-African ancestry. For example, in a restricted sample of cases 1031 and 1438 controls, we have greater than 80% power to detect an interaction odds ratio of 1.2 at p ≤ 0.05, assuming an odds ratio of 2.0 for an environmental exposure and of 1.2 for each unit increase in polygenic score [assuming N (0.1) distribution].

Case ascertainment and recruitment

All cases presenting to one of the 17 participating centres in 6 countries with a suspected first episode of psychosis were potentially eligible for inclusion in the study. The inclusion criteria for cases were (a) presence of at least one positive psychotic symptom for at least 1 day duration or two negative psychotic symptoms (for at least 6 months duration) within the timeframe of the study; (b) aged between 18 and 64 years (inclusive); and (c) resident within a clearly defined catchment area at the time of their first presentation. Residence was defined as a minimum of a one night stay at a residential address within the catchment areas. Exclusion criteria were (a) previous contact with specialist mental health services for psychotic symptoms outside of the study period at each site; (b) evidence of psychotic symptoms precipitated by an organic cause (ICD-10: F09); (c) transient psychotic symptoms resulting from acute intoxication (F1X.5); (d) severe learning disabilities, defined by an IQ less than 50 or diagnosis of intellectual disability (F70–F79); and, for the case–control part only, (e) insufficient fluency of the primary language at each site to complete assessments.

Case identification procedures involved teams of researchers regularly screening both general adult and specialist mental health services (both in- and out-patients). The screening process involved researchers regularly liaising with clinical staff and checking clinical records to identity potential cases. The researchers only included those individuals who they could be sure met the criteria based on the symptoms reported in the clinical notes. Potential cases were then approached when considered appropriate by clinical staff and informed consent sought.

Control recruitment

Inclusion criteria for controls were (a) aged between 18 and 64 years; (b) resident within a clearly defined catchment area at the time of consent into the study; (c) sufficient command of the primary language at each site to complete assessments; and (d) no current or past psychotic disorder. To select a population-based sample of controls broadly representative of local populations in relation to age, gender, and ethnicity, a mixture of random and quota sampling was used. Quotas for control recruitment were based on the most accurate local demographic data available. Quotas were then filled using a variety of recruitment methods, including (1) random sampling from lists of all postal addresses (e.g. in London); (2) stratified random sampling via GP lists (e.g. in London and Cambridge) from randomly selected surgeries; and (3) ad hoc approaches (e.g. internet and newspaper adverts, leaflets at local stations, shops, and job centers). In some sites (e.g. London), some groups (e.g. black African and black Caribbean) were oversampled to enable subsequent sub-group analyses. To deal with this in subsequent analyses, weights were generated, based on the most accurate local demographic data available, to minimize any resulting bias in estimating the prevalence of exposures among controls.

Individuals who agreed to take part were screened for a history of psychosis. Those who reported previous or current treatment for psychosis were excluded. Those who responded positively to any question in the screening instrument, indicating a possible psychotic experience, were interviewed further with standardised interviews to assess symptoms and to establish the presence or otherwise of a psychotic disorder. On this basis, no potential controls were found to have a past or current psychotic disorder.

Data contents

We collected data on an extensive range of exposures and outcomes across multiple domains using previously validated questionnaires, tasks, and procedures: demographic, clinical, social, psychological, cognitive, and biological (Table 2). All environmental exposures and cognitive and psychological tests were measured using previously validated questionnaires and tasks.

Table 2 EU-GEI study battery summary for the case–control study
Full size table

Genetic risk was assessed both indirectly, using a familial liability score for psychosis [28], and directly, using DNA extracted from two 9 ml non-fasting venous blood samples and/or via saliva samples (Oragene). Samples were genotyped using custom Illumina HumanCoreExome-24 BeadChip genotyping arrays containing probes for 570,038 genetic variants (Illumina Inc., San Diego, CA, USA). Genotype data were called using the GenomeStudio package, transferred into PLINK format for further analysis, and underwent quality control based on genotype variants and samples.

Quality assurance and control

Prior to and during data collection, annual multi-site meetings were arranged to bring together principal investigators and core researchers to ensure that standardised procedures were being implemented, to provide training, to discuss issues with data collection, and to conduct inter-rater reliability exercises. The study was designed to ensure comparable procedures and methods across settings, with some local adaptation to allow for variations in healthcare provision and health service contact points. The primary deviation from protocol was in the Veneto region, Italy, where data were derived from a previous study which used comparable methods [29], but had a lower upper-age limit of 54.

Training of researchers who were responsible for administrating the assessments was performed at the outset and throughout the study. This was organised by a technical working committee of the overall EU-GEI study (Work Package 11). An online resource was made available with taped interviews, samples of recordings, and written summaries for staff training purposes. Inter-rater reliability was assessed annually. Researchers were required to attain and maintain a minimum threshold of correct ratings before being allowed to administer the core assessments. Sufficient levels of inter-rater reliability for the core measurements, ranging from 0.70 to 0.91, were achieved, and are shown in Table 3.

Table 3 Inter-rater reliability scores of 115 core researchers
Full size table

Data management

Data were collected on paper and, for some cognitive tasks (e.g. the White Noise Task), on laptops and securely stored at each of the participating centres, and was entered locally using an encrypted web-based system, using commercial software (4D) that was adapted specifically for EU-GEI purposes. Data were entered once with field codes restricted to logical values where possible, to minimise data entry errors. Blood or saliva samples were taken at approved clinical research facilities by an experienced researcher and were fully anonymized and identified by bar code, and sent to the Institute of Psychological Medicine and Clinical Neurology at Cardiff University for genotyping. The data resource has undergone a rigorous period of validation checks and cleaning by a small number of experienced researchers. This has involved checks of missing data and corroboration of these against the paper files at each of the 17 sites.

Ethical approval

All participants who agreed to take part in the study provided informed, written consent following full explanation of the study. Ethical approval for the study was provided by relevant research ethics committees in each of the study sites [30].

Results

Sample representativeness and characteristics

There were similar proportions from minority ethnic groups among consented and non-consented cases (43% vs. 40%). However, the proportion of men and the proportions in younger age groups were higher among consented, compared with non-consented, cases (men: 62% vs. 57%; aged 18–34 years: 69% vs. 60%) (Table 4a). Compared with the general population, controls were more likely to belong to a minority ethnic group (controls: 28%, population at-risk: 23%) and were younger (aged between 18 and 34 years, controls: 56%, population at-risk: 38%) (Table 4b). The greater proportion of controls who were from minority ethnic groups reflects oversampling in some sites (e.g. London) to enable subsequent sub-group analyses.

Table 4 Representativeness of (a) the consented case sample compared with the incidence sample, and (b) the control sample compared with the population-at-risk
Full size table

Cases were younger than controls {median age of cases was 29 years [interquartile range (IQR) 22–37], and controls 33 years old [IQR 26–47]}. Compared with controls, a greater proportion of the cases were men (62% vs. 49%), migrants (28% vs. 22%), and left school without any qualifications (16% versus 6%); a smaller proportion was of white ethnicity (63% versus 73%) (see Supplementary Table S2).

Discussion

This study was conducted in a diverse range of settings across Europe and one setting in Brazil, selected to ensure a mix of urban and rural areas with large migrant and minority ethnic populations. This maximises its applicability to and importance for public health initiatives, with potential implications for both prevention and intervention, particularly among minority ethnic groups, and in urban areas, and in relation to cannabis and other substance use and developmental adversity. Our primary hypotheses centre on examining variations in incidence and symptoms, environmental risk factors, and the interplay between environment and genetic factors in the development of psychotic disorders.

Incidence and symptoms

We have already published findings of the overall variations in incidence of psychoses by site [30]. Our findings suggest marked geographical differences in the incidence of psychotic disorders, with around an eightfold variation among study sites after accounting for age, sex, and minority ethnic status. At an area level, initial analyses suggest that some of this variation may be related to the proportion of owner-occupied homes in an area (a tentative proxy for social cohesion or socioeconomic deprivation), i.e. areas with more owner-occupied homes had, on average, lower rates of psychotic disorder. Analyses of variations in incidence by ethnic group are ongoing. Analyses of symptom data on incident cases, collated using the OPCRIT, have examined the validity of a transdiagnostic dimensional structure of psychopathology and, in doing so, have challenged the common binary categorisation of psychoses into non-affective and affective disorders [31]. Our findings suggest that a bifactor model of psychopathology, comprising one general factor and five dimensions (positive, negative, manic, disorganised, and depressive symptoms), best represents the structure of symptoms among those with a psychotic disorder. We further found, compared with majority populations, cases in minority ethnic groups scored higher on the positive psychotic symptom dimension; and, compared with rural areas, cases in urban areas scored higher on the general symptom dimension.

Environmental risk

The initial focus of analyses of our case–control data resource is the associations and population impact of putative environmental risk factors, including childhood adversity and abuse, adult adversity, discrimination, and cannabis use. In analyses of cannabis use data, for example, we found that, compared with those who did not use cannabis, the odds of psychosis were (1) around three times higher among those who used cannabis daily; (2) around two times higher among those who spent more than 20 Euros a week on cannabis; and (3) around 50% higher among those who used cannabis high in THC [32]. In addition, we found variations in population attributable fractions for daily cannabis use on psychosis [32], with population attributable fractions (i.e. the proportion of psychosis, assuming causality, attributable to daily use) ranging from 1 (in Puy-de-Dôme, France) to 44% (in Amsterdam). Similar analyses examining childhood and adult adversity are ongoing, focusing on type, severity, and age of exposure (Morgan et al., in preparation).

These analyses will be further extended to examine environment–environment and gene–environment interactions and to more clearly elucidate the pathogenic processes underpinning observed variations in incidence across study sites [30] and high rates of psychotic disorders in urban areas [7, 8], and in migrant and minority ethnic groups [9, 10].

Strengths and weaknesses

To our knowledge, this is the most extensive multi-site incidence and case–control study of first-episode psychosis ever conducted, with comprehensive data on a variety of environmental, psychological, and genetic risk factors. The primary strength of the EU-GEI study is its potential to provide ground-breaking and important information about the development of psychoses, by investigating the complex interrelationships between candidate environmental, psychological, and biological (genetic) factors and psychotic disorders, including the mechanisms through which they increase risk. In addition, given that our study was carried out in major urban and rural sites with heterogeneous populations suggests that our external validity may extend to other centres with similar population profiles. The combined incidence and case–control methodology allows for precise identification of, and ability to account for, any potential selection biases amongst the recruited and assessed cases. The richness of the exposure information available will allow for more nuanced analyses and a more fine-grained understanding of their impact on psychotic disorder than has been possible to date. Importantly, the inclusion of only cases with a first episode of psychosis (rather than individuals with long-standing disorder) allows inferences to be made about causal connections and processes.

The primary limitation of these data resource relates to case identification. As in all previous studies, we relied on first contact with mental health services as a proxy for first onset. While it is likely most individuals who develop a psychotic disorder do present to services, at least in sites with well-developed public health systems, some who do not present will be missed and this may introduce selection biases. Any rate estimates should, therefore, be considered as treated incidence. Further, variations in referral procedures of patients with psychosis from primary to secondary mental health care settings and in the organization of secondary mental health care services across catchment areas may have influenced the identification of cases, and may explain some of the variation in estimates of incidence across study sites and countries. For example, unlike in other settings, patients in Madrid are not constrained to using mental health services in their residential catchment areas [33]. However, as highlighted by Jongsma et al. [30], the divergences in service provision and cultural context are unlikely to fully explain the eightfold variation in incidence across sites.

There are also several limitations that are inherent to case–control designs. First, while substantial efforts were made at the outset to reduce the potential biases in the identification of cases (e.g. recruitment of participants from a number of sources using a variety of methods, including inpatient wards and community teams) and controls (e.g. use of mixture of random and quota sampling), we were not entirely successful; our cases are not fully representative of the sample identified in the incidence study, and our controls not of the population-at-risk. For example, reliance in some sites on recruitment of controls through ad hoc methods, such as newspaper advertisements, may have biased samples. Interpretations of estimated effects (odds ratios) should be considered with this in mind.

Second, there is the potential for both recall and observer bias. To minimise these, and validate environmental exposures, several steps were taken. For core environmental exposures (e.g. childhood adversity and cannabis use), we used extensive, well-validated measures, that drew on life course methods to anchor memories and improve recall. All researchers administering these assessments went through intensive training, with regular top-ups. Further, where possible, we drew on corroborative sources of information in the assessment of exposure to childhood and adulthood adversity [e.g. clinical records, interviews with siblings of a subsample of cases (n = 272)].

Third, measurement of exposure occurred after onset of disorder, making causal inferences problematic. To establish the temporal ordering of exposure and outcome, we carefully established the date of onset of disorder and, for measures of exposures in childhood and adulthood, ensured that all assessments related to the period pre-onset.

Finally, given the large battery of tests and interviews conducted with our participants, data were missing for some assessments, particularly towards the end of the study battery. Where appropriate, a standardised procedure for multiple imputation will be used to minimise the loss of precision or selection biases which may otherwise be introduced in complete case analyses.

Data resource access

The EU-GEI WP2 principal investigators (contact: craig.morgan@kcl.ac.uk) welcome formal requests for access to the data, biological samples, and/or collaborative projects. Researchers will be required to complete an EU-GEI WP2 data interest form to state their intended hypotheses and analysis plan, which will be reviewed by the PIs to determine whether the proposal can be addressed by this data resource, does not duplicate on-going or completed analyses with this dataset, and lies within the scope of current ethical approvals. More information about the study can be found on the study website (https://www.eu-gei.eu/).