Browse the corpus

Introduction Auditory hallucinations—or voices—are a common feature of schizophrenia. They also occur in other disorders and in individuals with no psychiatric history.1 Understanding of subjective experiences of hallucination—and how they vary between different populations—is a central concern of psychiatry, and can help with the development of new causal accounts of auditory hallucination and more effective therapeutic interventions.2,3 Although various resources document first-person experiences of voice-hearing,4 systematic empirical research on the phenomenology of auditory hallucinations remains scarce. Nayani and David's 1996 study5 analysed clinical interview data from 100 patients with psychosis with auditory hallucinations (61% of 100 individuals had ICD-10 schizophrenia diagnoses). The investigators concluded that auditory hallucinations in this population are typically repetitive emotive utterances that increase in number and complexity over time. In 2014, McCarthy-Jones and colleagues6 analysed auditory hallucination descriptions from 199 patients (81% of individuals had a diagnosis of DSM-III-R schizophrenia), obtained through the Mental Health Research Institute (MHRI) Unusual Perceptions Scale.7 Cluster analysis of these findings suggested four common factors: voices that were repetitive, commanding or involved running commentary (86%); voices similar to a person's own thoughts (36%); voices that were clearly reminiscent of specific memories (12%); and non-verbal auditory hallucinations (42%).6

eptions Scale.7 Cluster analysis of these findings suggested four common factors: voices that were repetitive, commanding or involved running commentary (86%); voices similar to a person's own thoughts (36%); voices that were clearly reminiscent of specific memories (12%); and non-verbal auditory hallucinations (42%).6 Although such surveys provide insight into the experience of auditory hallucinations, the focus on psychosis, particularly schizophrenia, leaves the potential cross-diagnostic features of auditory hallucinations unexplored. Additionally, the semi-structured interviews and closed-ended approaches often used make several a priori assumptions about the key features of auditory hallucinations, which prioritise some structural characteristics (eg, loudness) over others (eg, voice identity). Clinical terminology is often itself loaded and might prime or encourage participants to describe their experiences in particular ways (eg, as auditory or linguistic). From a phenomenological perspective, these approaches might constrain understanding of auditory hallucinations in potentially serious ways.8,9

dentity). Clinical terminology is often itself loaded and might prime or encourage participants to describe their experiences in particular ways (eg, as auditory or linguistic). From a phenomenological perspective, these approaches might constrain understanding of auditory hallucinations in potentially serious ways.8,9 To address these concerns, and as part of the Hearing the Voice project and Lived Experience Network, we developed a questionnaire on voices and voice-like experiences. We drew on the expertise of philosophers, psychologists, medical humanities scholars, and researchers with lived experience of auditory hallucination, in consultation with clinicians and people who hear voices, from the project's advisory group. We aimed to record a detailed and diverse collection of experiences, in the words of the people who hear voices themselves. Methods Participants We made the questionnaire available via the project website for 3 months for anonymous online completion. We invited people aged 16–84 years with experience of voice-hearing to take part via an advertisement circulated through clinical networks, hearing voices groups, and other mental health forums. We asked participants if they had ever received a psychiatric diagnosis, and if so, to report their present or most recent diagnosis. Participants consented to use of their data in the study before accessing the questionnaire and confirmed this upon completion. All procedures were approved by Durham University ethics committee.

participants if they had ever received a psychiatric diagnosis, and if so, to report their present or most recent diagnosis. Participants consented to use of their data in the study before accessing the questionnaire and confirmed this upon completion. All procedures were approved by Durham University ethics committee. Procedures Participants completed a 13 item questionnaire that was available online through Qualtrics (Provo, UT, USA; appendix). Recognising that no term is neutral or universally accepted, we chose to use the term voices because it is widely understood and used in non-clinical and clinical contexts. Many people who hear voices regard the term auditory hallucination as stigmatising because it implies that their experiences are not real.10,11 Furthermore, we did not want to restrict the study by implying that the phenomena in question are necessarily always auditory or perceptual. We designed the questions to be unbiased, non-leading, and non-hierarchising prompts that aimed to elicit phenomenologically rich data. The questionnaire combined closed-ended and open-ended questions (eg, “Please try to describe your voice(s) and/or voice-like experiences”; “How, if at all, are these experiences different from your own thoughts?”). All questions were optional and no word limit was imposed on responses.

it phenomenologically rich data. The questionnaire combined closed-ended and open-ended questions (eg, “Please try to describe your voice(s) and/or voice-like experiences”; “How, if at all, are these experiences different from your own thoughts?”). All questions were optional and no word limit was imposed on responses. Statistical analysis We analysed the data using a mixture of qualitative and quantitative methods. First, we integrated responses into single narratives. We then did an inductive thematic analysis.12,13 Each member of the research team initially coded 20 responses. Once collated, we refined and organised the lists of codes into a coding framework with inclusion and exclusion criteria noted for each code. Two independent raters (AW and NJ) then coded the data using NVivo 10 software. Once high inter-rater reliability (κ=0·85) was established for 30% of the sample, the raters divided and coded the remaining data independently. Responses were analysed as single integrated narratives that could be assigned each code a maximum of once. Any ambiguous instances were resolved through discussion and a consensus-based decision. The nature of some questions allowed for mutually exclusive categorical coding of responses (eg, codes for child, adolescent, and adult onset). However, most of the codes that we used were not mutually exclusive because participants often described a range of phenomenological and structural characteristics.

Statistical analysis We analysed the data using a mixture of qualitative and quantitative methods. First, we integrated responses into single narratives. We then did an inductive thematic analysis.12,13 Each member of the research team initially coded 20 responses. Once collated, we refined and organised the lists of codes into a coding framework with inclusion and exclusion criteria noted for each code. Two independent raters (AW and NJ) then coded the data using NVivo 10 software. Once high inter-rater reliability (κ=0·85) was established for 30% of the sample, the raters divided and coded the remaining data independently. Responses were analysed as single integrated narratives that could be assigned each code a maximum of once. Any ambiguous instances were resolved through discussion and a consensus-based decision. The nature of some questions allowed for mutually exclusive categorical coding of responses (eg, codes for child, adolescent, and adult onset). However, most of the codes that we used were not mutually exclusive because participants often described a range of phenomenological and structural characteristics. We used coded data to calculate descriptive statistics for common features of voice-hearing across the full sample. We used a mixed-methods priority-sequence model, in which we used quantitative analyses (χ2 tests) to test additional associations of selected codes that were either identified in the principal qualitative analyses or suggested by previous studies.14 We applied a false discovery rate correction15 to correct for multiple comparisons. We did not calculate any post-hoc measures of power for the study, mainly because specific hypothesis testing was not the focus of the study (as this would contradict key components of the phenomenological method), but also because of theoretical concerns about the notion of post-hoc power.

ct for multiple comparisons. We did not calculate any post-hoc measures of power for the study, mainly because specific hypothesis testing was not the focus of the study (as this would contradict key components of the phenomenological method), but also because of theoretical concerns about the notion of post-hoc power. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results 157 participants completed the survey, and we excluded four responses that did not discuss voice-hearing experiences, for a total of 153 responses. Various diagnoses were reported (table 1), the most common of which were schizoaffective disorder (24 [16%] of 153 individuals) and bipolar disorder (21 [14%] individuals). The total length of the responses ranged from 24 to 2474 words (mean 510 words, SD 432). Table 2 shows demographic details of the survey population.

arious diagnoses were reported (table 1), the most common of which were schizoaffective disorder (24 [16%] of 153 individuals) and bipolar disorder (21 [14%] individuals). The total length of the responses ranged from 24 to 2474 words (mean 510 words, SD 432). Table 2 shows demographic details of the survey population. Less than half of participants described literally auditory experiences (ie, voices indistinguishable from voices or other sounds), and 14 (9%) individuals reported exclusively thought-like voices (ie, with no auditory qualities; table 3). We encouraged description of the differences in the characteristics of these experiences (panel 1) by using questions that directly invited participants to compare voices with their thoughts and actual voices in the room (appendix). 56 (37%) participants—coded as auditory–thought mixed—reported either a combination of auditory and thought-like voices or experiences that were somewhere between literally auditory and thought-like. Notably, most individuals who described their experiences as non-literally auditory still referred to them as voices. About a fifth (30 individuals) of the sample deemed voice an inadequate term for their experience, instead using terms such as “intuitive knowing” or “telepathic experience”, or descriptors such as “alters”, “parts”, or “fellow system members”.

eir experiences as non-literally auditory still referred to them as voices. About a fifth (30 individuals) of the sample deemed voice an inadequate term for their experience, instead using terms such as “intuitive knowing” or “telepathic experience”, or descriptors such as “alters”, “parts”, or “fellow system members”. 124 (81%) participants reported the presence of several voices, with only 10 (7%) individuals reporting a single voice. Most participants reported having had multiple voices, with a quarter (39 individuals) reporting undifferentiated or ambiguous collections of voices, such as crowds, gangs, or classroom groups. Voices with a physical location were equally likely to be external or internal. Most voices were described as being characterful in some way (table 4)—ie, people or person-like entities with distinct characteristics, such as gender, age, patterned emotional responses, or intentions. “I hear distinct voices. Each voice has their own personality. They often try to tell me what to do or try to interject their own thoughts or feelings about a certain subject or matter […] My voices range in age and maturity. Many of them have identified themselves and given themselves names.” “I hear a mixture of men and women, but no children. They usually tell me to do things, but not dangerous things. Like they'll tell me to take out the garbage or check the lock on the window or call someone. Sometimes they comment on what I'm doing and whether I'm doing a good job or what I could be doing better.”

“I hear distinct voices. Each voice has their own personality. They often try to tell me what to do or try to interject their own thoughts or feelings about a certain subject or matter […] My voices range in age and maturity. Many of them have identified themselves and given themselves names.” “I hear a mixture of men and women, but no children. They usually tell me to do things, but not dangerous things. Like they'll tell me to take out the garbage or check the lock on the window or call someone. Sometimes they comment on what I'm doing and whether I'm doing a good job or what I could be doing better.” Roughly a fifth (33 [22%] of 153) of participants described voices that were recognised as specific, existing individuals. 24 (16%) participants described voices that were understood to be supernatural or spiritual entities. Common characteristics of address were conversational voices (engaging the voice-hearer directly) or voices that commented on specific things. Few people reported only so-called simple voices—single words or brief phrases—or voices that did not address them directly. Only 8 (5%) participants reported voices which predominantly issued negative commands; overall experiences of abusive or violent voices were much more common. Although many voices were described as either positive or neutral in tone, negative emotions were often associated with them, especially fear, anxiety, depression, and stress.

Common characteristics of address were conversational voices (engaging the voice-hearer directly) or voices that commented on specific things. Few people reported only so-called simple voices—single words or brief phrases—or voices that did not address them directly. Only 8 (5%) participants reported voices which predominantly issued negative commands; overall experiences of abusive or violent voices were much more common. Although many voices were described as either positive or neutral in tone, negative emotions were often associated with them, especially fear, anxiety, depression, and stress. “Starting when I was about 20 years old, I heard the voices of demons screaming at me, telling me that I was damned, that God hated me, and that I was going to hell… The voices were so frightening and disruptive that much of the time I was unable to focus or concentrate on anything else.” “To a point, they generally are anything but kind to me. They can be brutally sarcastic and intrusive.” About two-thirds of participants (101 individuals) reported changes in bodily experience when they heard voices (table 4), which varied substantially. “My body and brain felt like they were on fire when I heard the voices; I had constant tingling sensations throughout my extremities and shock-like sensations in my solar plexus.” “Yes, my body felt more distant from me—the whole experience felt a bit dreamlike (like living a dream), surreal, other worldly.” “At the very beginning I experienced a heat and a strong irritation in the right frontal part of my brain.”

“My body and brain felt like they were on fire when I heard the voices; I had constant tingling sensations throughout my extremities and shock-like sensations in my solar plexus.” “Yes, my body felt more distant from me—the whole experience felt a bit dreamlike (like living a dream), surreal, other worldly.” “At the very beginning I experienced a heat and a strong irritation in the right frontal part of my brain.” 28 (18%) people had multisensory voices, suggesting that their voices were perceived simultaneously through more than one sensory modality. 43 (28%) participants reported distinct hallucinations in other senses, and some people also described voices that gave access to other minds, or information that would not otherwise be available. A few (10–20) participants reported experiences of tiredness, sleep disturbance, and mania.

more than one sensory modality. 43 (28%) participants reported distinct hallucinations in other senses, and some people also described voices that gave access to other minds, or information that would not otherwise be available. A few (10–20) participants reported experiences of tiredness, sleep disturbance, and mania. In cases where participants described their first voice experiences, the experiences often occurred in childhood (table 5). Many participants reported negative or explicitly traumatic circumstances, with few voices (17 [11%] of 153 individuals) arising in positive or neutral circumstances. More than a third (53 of 153 individuals) of participants described structural transformations in the number and presence of voices over time, with a few (19 [12%] individuals) also reporting changes in voice content, frequency, or valence (emotional reaction elicited). Only one respondent specifically stated that their voice had not changed over time. Although 34 (22%) participants stated that they were unable to influence their voices, 54 (35%) reported that they could influence their voices indirectly (through strategies of avoidance, medication, or environmental change), and 69 (45%) individuals reported influencing their voices by engaging directly with them or exploring their meaning. The effect of the voices on participants' relationships with others was largely negative: 48 (31%) participants cited direct negative effects (eg, voices interrupting conversation or making it difficult to understand what others were saying), and 61 participants (40%) referenced a general negative effect, including experiences of stigma, fear, and loneliness.

relationships with others was largely negative: 48 (31%) participants cited direct negative effects (eg, voices interrupting conversation or making it difficult to understand what others were saying), and 61 participants (40%) referenced a general negative effect, including experiences of stigma, fear, and loneliness. To investigate the distinction between auditory and mixed auditory and thought-like voices, we compared numbers of people reporting each type of voice for a selection of the codes identified during the qualitative analysis (table 6). Participants with mixed auditory and thought-like voices were more likely than those with purely auditory experiences to report voices that were internal (p=0·010), conversational (p=0·010), had changed over time (p=0·030), and gave access to other minds (p=0·026). Mixed voices trended non-significantly towards being associated with voices that gave access to information that was otherwise unknown by the participant (p=0·051). No other contrasts were significant (table 6). We compared participants with and without characterful voices (table 7). People who heard characterful voices were significantly more likely to be able to influence their voices (p=0·040) and, at the non-significant trend level, were more likely to experience voices that were abusive or violent (p=0·051) than were those who heard non-characterful voices (table 7).

acterful voices (table 7). People who heard characterful voices were significantly more likely to be able to influence their voices (p=0·040) and, at the non-significant trend level, were more likely to experience voices that were abusive or violent (p=0·051) than were those who heard non-characterful voices (table 7). We compared participants who specifically reported effects on the body with those who did not (table 8). Participants with bodily experiences were more likely to report voices that were abusive or violent (p=0·024) and to be able to anticipate their voices (p=0·025) than were those with no bodily effect. Reporting of bodily experiences seemed to be associated with reporting of traumatic circumstances when participants first heard voices, voices that were associated with shame, and few positive and useful voices (p=0·05–0·06; table 8). A unique characteristic of our sample was its cross-diagnostic nature, including some participants who specifically reported that they had never received a psychiatric diagnosis (26 [17%] of 153 individuals). Based on previous research with similar populations,16–18 we compared people who had received a clinical diagnosis with those who had not (table 9). Participants who had not been clinically diagnosed were significantly less likely to associate their voices with fear (p=0·010) or depression (p=0·015) than were those with a clinical diagnosis. We detected no differences for any other categories (table 9).

eceived a clinical diagnosis with those who had not (table 9). Participants who had not been clinically diagnosed were significantly less likely to associate their voices with fear (p=0·010) or depression (p=0·015) than were those with a clinical diagnosis. We detected no differences for any other categories (table 9). To help with comparison with previous studies, we also did an exploratory analysis to compare participants who reported schizophrenia-related diagnoses (schizophrenia or schizoaffective disorder, n=38) with all other participants for a selection of codes associated with the classic understanding of auditory hallucinations in schizophrenia as auditory, externally located, and commanding phenomena. We identified no significant differences, even if we used an uncorrected p value cutoff (codes used: auditory, auditory-thought mixed, internal location, external location, single voice, multiple voices, and commanding nature).

y hallucinations in schizophrenia as auditory, externally located, and commanding phenomena. We identified no significant differences, even if we used an uncorrected p value cutoff (codes used: auditory, auditory-thought mixed, internal location, external location, single voice, multiple voices, and commanding nature). The sample of respondents included a large proportion of female participants. To check for the effect of gender, we did χ2 analyses to compare men and women for group membership in the four subgroups analysed (auditory voices, characterful voices, bodily effect, and clinical diagnosis), and in association with all codes analysed (to avoid type II errors, we did not apply a false discovery rate correction). We detected no significant associations between gender and subgroup, and only three codes were significantly associated: paranoia was more likely in men (p=0·036), shame (p=0·022), and stress (p=0·006) women. However, the relative percentage of women did noticeably vary between diagnostic groups (appendix).

orrection). We detected no significant associations between gender and subgroup, and only three codes were significantly associated: paranoia was more likely in men (p=0·036), shame (p=0·022), and stress (p=0·006) women. However, the relative percentage of women did noticeably vary between diagnostic groups (appendix). Discussion We used an open-ended, internet-based survey to obtain detailed information about the phenomenology of auditory hallucination from a diverse array of individuals, including those without psychiatric diagnoses (panel 2). Several of our findings are consistent with other large-sample studies of auditory hallucinations5,6,18,19 and longstanding clinical observations—ie, the high prevalence of multiple voices, typically with distinct characteristics; variations in acoustic properties, linguistic complexity and location; and strong associations with negative emotion, especially for individuals with psychiatric diagnoses.5,6,20–22 However, unlike the published scientific literature, our findings also suggest novel and under-researched aspects of auditory hallucination phenomenology. Specifically, we focus on distinctions between thought-like, mixed, and strictly auditory voices; voices with somatic effects; and the experiential complexities of characterful voices.

published scientific literature, our findings also suggest novel and under-researched aspects of auditory hallucination phenomenology. Specifically, we focus on distinctions between thought-like, mixed, and strictly auditory voices; voices with somatic effects; and the experiential complexities of characterful voices. Although auditory hallucinations are usually understood as predominantly perceptual experiences, nearly half of our participants described their voices either as thought-like or as having both auditory and thought-like qualities. Such mixed voices were significantly more likely to be conversational, show change over time, and be experienced as giving access to other minds. So-called sound talk (mentions of loudness, timbre, pitch, resonance, accent, and rhythm) was very common throughout the sample, complicating clear distinctions between thoughts and perceptions (eg, “My thoughts are shouting” or “I experience a silent scream […] a presence, an emotional energy, or potential that I can feel but not hear”). These findings are similar to historical, cognitive, and phenomenological research on the qualities of imagined sound23,24 and raise the question of whether some voices might be better understood as passive or uncontrolled imagined perceptions, rather than perceptual hallucinations. The extent to which the message of auditory hallucinations can be understood without being heard is also worthy of further study.

lities of imagined sound23,24 and raise the question of whether some voices might be better understood as passive or uncontrolled imagined perceptions, rather than perceptual hallucinations. The extent to which the message of auditory hallucinations can be understood without being heard is also worthy of further study. Participants also frequently reported multisensory voices, concurrent somatic events, and hallucinations in other sensory modalities. Whether we classify these other-sensory or somatic features as adjunctive components of auditory hallucinations or instead as events distinct from specifically auditory hallucinations, the implications of our findings are potentially important to attempts to understand and assign subtypes to hallucinatory phenomena. The high prevalence of multisensory voices and somatic features is also important in view of the scarce attention to such features in existing clinical interventions, and could inform further development of theoretical models that link self-recognition to deficits in sensory-motor control at the level of body schema.25 Notably, voices with effects on the body were also significantly more likely be associated with an overall experience of voices that were abusive or violent, and voices that could be anticipated in some way. Although we did not detect a significant association between voices with somatic aspects and trauma, the strong associations between abusive voices and childhood adversity,26 especially sexual and physical trauma,27 suggest that this association might be promising for future study.

hat could be anticipated in some way. Although we did not detect a significant association between voices with somatic aspects and trauma, the strong associations between abusive voices and childhood adversity,26 especially sexual and physical trauma,27 suggest that this association might be promising for future study. Command hallucinations are widely regarded as distressing and indicative of high risk of harm to self and others,28 and yet their content, severity, and importance have tended to be assumed rather than fully investigated. Command hallucinations were reported by 84% of 100 participants in Nayani and David's study5 and “constant, commanding and commenting” auditory hallucinations were reported by 86% of 199 participants in McCarthy-Jones and colleagues' study.6 We coded voices that issued negative commands or instructions to do harmful things as commanding, distinct from voices that issued requests or instructions to do things that were benign or helpful. Thus defined, command hallucinations characterised the overall experience of voice-hearing for only 8 (5%) of 153 participants. This discrepancy between our study and other phenomenological surveys could be caused by differences in populations and settings between studies: command hallucinations might be the dominant experience for individuals with a schizophrenia diagnosis, or those who are reporting on their voices in a clinical context and engaging with health-care services. Alternatively, a substantial number of people who hear voices who receive advice or strong suggestions from their voices might have been mislabelled as experiencing commands that are presumed to be inherently violent or potentially harmful.

g on their voices in a clinical context and engaging with health-care services. Alternatively, a substantial number of people who hear voices who receive advice or strong suggestions from their voices might have been mislabelled as experiencing commands that are presumed to be inherently violent or potentially harmful. The characterful or person-like nature of voices has been widely documented,4,10 is directly addressed by existing psychological interventions for voices,29 and was one of the most common aspects of voice-hearing reported in our analysis. However, little investigation has been done on the different ways that voices might be experienced as personified. The descriptions in our data suggest a range of person-like qualities, from amorphous entitativity (an undefined disembodied personality), to stereotypical person-like presentations (an angry man, an old woman), spiritual entities with anthropomorphic traits, specifically recognisable individuals, and voices that are subjectively experienced as representing all or part of the person's own self. Characterful voices were also distinguishable from other voices in their susceptibility to influence by the voice-hearer: more characterful voices could be directly engaged with in a meaningful way. These findings raise important conceptual, philosophical, and clinical questions for future research, including how the characterological features of voices are shaped by individuals' explanatory beliefs and local cultures.30 The heterogeneity of characterful voices also underscores the importance of existing relational interventions29,31 to address variability in the types of voices and their person-like qualities.

h, including how the characterological features of voices are shaped by individuals' explanatory beliefs and local cultures.30 The heterogeneity of characterful voices also underscores the importance of existing relational interventions29,31 to address variability in the types of voices and their person-like qualities. One limitation of the present study was the coding of characteristics derived from free-text written responses; some participants might have had particular experiences (such as command hallucinations), but not independently volunteered this information in our questionnaire. Our results might therefore underestimate the prevalence of features we coded for. Conversely, characteristics that are routinely discussed in clinical settings (such as voice location) might have been over-represented compared with less studied aspects of auditory hallucination experience. Ultimately, phenomenological investigation provides “no means to check the ‘truth’ of the responses recorded”, as noted by Nayani and David,5 and the departure from psychometrically validated measures limits the extent to which comparisons can be drawn between this study and other studies of auditory hallucination phenomenology. However, adoption of an exploratory, rather than prescriptive, approach to what counts as a voice or voice-like experience yields new insights into what people who hear voices themselves regard as most important. These insights are potentially of great importance to existing research frameworks that depend on assumptions that our data call into question, such as a focus on auditory hallucination as a primarily perceptual event.

experience yields new insights into what people who hear voices themselves regard as most important. These insights are potentially of great importance to existing research frameworks that depend on assumptions that our data call into question, such as a focus on auditory hallucination as a primarily perceptual event. Second, the online questionnaire was accessible only to English-speakers with basic internet literacy and access. Although the online platform might be thought to limit participation, results from research have shown that people with severe mental illness have rates of smartphone access and usage similar to the general public.32 We mainly recruited to the study through existing research, clinical, and service-user networks. High-functioning users of social media who are already engaged in such networks or communities might be over-represented, while individuals who are currently in acute care settings are almost certainly under-represented. Moreover, although the capacity to participate anonymously might have encouraged frank responses from some participants, we were unable to verify participants' self-reports. Because these self-reports include self-reported diagnoses, we have restricted ourselves to clinical versus non-clinical diagnoses and schizophrenia-spectrum versus other comparisons, rather than more specific distinctions between clinical diagnoses. In-depth comparison of voice phenomenology in different diagnostic contexts—including dissociative identity disorder and post-traumatic stress disorder—is a crucial topic for future studies of this kind.

schizophrenia-spectrum versus other comparisons, rather than more specific distinctions between clinical diagnoses. In-depth comparison of voice phenomenology in different diagnostic contexts—including dissociative identity disorder and post-traumatic stress disorder—is a crucial topic for future studies of this kind. Third, our overall sample shows substantial bias in terms of gender and ethnicity, limiting the representativeness and generalisability of our findings. 2·5 times as many women as men completed the study, which might be indicative of wider trends in survey response rates33 and hallucination proneness,34 and the cross-diagnostic nature of our sample. Although people from black and minority ethnic origins are up to nine times more likely than people from other ethnic origins to present with symptoms of psychosis,35 they were under-represented in our study. When we analysed gender effects in our data, we detected differences for only three codes: paranoia (which was more likely in men), childhood onset, and structured longitudinal change (which were both more likely in women than in men). These results might be caused by differences worthy of future attention, but their exploratory nature makes these findings tentative at best.

differences for only three codes: paranoia (which was more likely in men), childhood onset, and structured longitudinal change (which were both more likely in women than in men). These results might be caused by differences worthy of future attention, but their exploratory nature makes these findings tentative at best. Despite these limitations, our methods allowed us to reach a demographically and diagnostically diverse sample, which included participants with little or no current contact with mental health services. The use of more prescriptive clinical tools, or confining of our sample to clinical settings, would possibly have limited the range of experiences reported. If full understanding of the phenomenology of auditory hallucination is important, across diagnoses and between clinical and non-clinical populations,2,21 then such methods are a necessary starting point. By engaging a sample of people who hear voices with varying diagnoses and clinical histories, we report both overlap with past qualitative investigations of auditory hallucination and potentially important new findings that depart from previous studies of the phenomenology of voices. These findings underscore the importance of future investigations of the association between acoustic perception and thought, the somatic and multisensorial features of auditory hallucination, and the link between auditory hallucination and characterological entities. This online publication has been corrected. The corrected version first appeared at thelancet.com/psychiatry on April 2, 2015

By engaging a sample of people who hear voices with varying diagnoses and clinical histories, we report both overlap with past qualitative investigations of auditory hallucination and potentially important new findings that depart from previous studies of the phenomenology of voices. These findings underscore the importance of future investigations of the association between acoustic perception and thought, the somatic and multisensorial features of auditory hallucination, and the link between auditory hallucination and characterological entities. This online publication has been corrected. The corrected version first appeared at thelancet.com/psychiatry on April 2, 2015 Supplementary Material Supplementary appendix Supplementary audio What does it mean to hear voices? Authors Angela Woods and Ben Alderson-Day discuss the results of their new study with Lancet Psychiatry Editor Niall Boyce. Acknowledgments This study was funded by a Wellcome Trust Strategic Award (WT098455MA). We thank Matthew Ratcliffe for contributing to the conceptualisation and design of the study, and Sam Wilkinson and David Smailes for contributing to the preliminary data analysis. Contributors AW, NJ, and CF conceived the study. All authors contributed to the study design. AW and NJ coded, analysed, and interpreted the data, in liaison with other authors. BA-D did the statistical analyses and produced the tables. AW drafted the initial manuscript, with extensive contributions from NJ and BA-D. All authors contributed to editing and finalising the report. Declaration of interests We declare no competing interests.

Contributors AW, NJ, and CF conceived the study. All authors contributed to the study design. AW and NJ coded, analysed, and interpreted the data, in liaison with other authors. BA-D did the statistical analyses and produced the tables. AW drafted the initial manuscript, with extensive contributions from NJ and BA-D. All authors contributed to editing and finalising the report. Declaration of interests We declare no competing interests. Table 1 Diagnostic information by gender Female (n=100) Male (n=40) Other*(n=13) Schizoaffective disorder 14 (9%) 9 (6%) 1 (1%) Bipolar disorder 16 (10%) 5 (3%) 0 Major depression 11 (7%) 2 (1%) 1 (1%) Schizophrenia 5 (3%) 9 (6%) 0 Post-traumatic stress disorder 9 (6%) 1 (1%) 1 (1%) Dissociative identity disorder 7 (5%) 0 4 (3%) Borderline personality disorder 5 (3%) 2 (1%) 1 (1%) Depression (mixed) 4 (3%) 2 (1%) 1 (1%) Generalised anxiety disorder 5 (3%) 0 1 (1%) Psychosis (NOS) 2 (1%) 1 (1%) 1 (1%) Obsessive compulsive disorder 1 (1%) 1 (1%) 1 (1%) Other diagnosis 3 (2%) 1 (1%) 1 (1%) No diagnosis 18 (12%) 7 (5%) 1 (1%) Not all patients gave all details, therefore percentages do not always sum to 100%. NOS=not otherwise specified. * Other includes androgyny, genderfluid, genderqueer, transgender, non-binary, and bigender. Table 2 Demographic information

Female (n=100) Male (n=40) Other*(n=13) Schizoaffective disorder 14 (9%) 9 (6%) 1 (1%) Bipolar disorder 16 (10%) 5 (3%) 0 Major depression 11 (7%) 2 (1%) 1 (1%) Schizophrenia 5 (3%) 9 (6%) 0 Post-traumatic stress disorder 9 (6%) 1 (1%) 1 (1%) Dissociative identity disorder 7 (5%) 0 4 (3%) Borderline personality disorder 5 (3%) 2 (1%) 1 (1%) Depression (mixed) 4 (3%) 2 (1%) 1 (1%) Generalised anxiety disorder 5 (3%) 0 1 (1%) Psychosis (NOS) 2 (1%) 1 (1%) 1 (1%) Obsessive compulsive disorder 1 (1%) 1 (1%) 1 (1%) Other diagnosis 3 (2%) 1 (1%) 1 (1%) No diagnosis 18 (12%) 7 (5%) 1 (1%) Not all patients gave all details, therefore percentages do not always sum to 100%. NOS=not otherwise specified. * Other includes androgyny, genderfluid, genderqueer, transgender, non-binary, and bigender. Table 2 Demographic information Number of participants (n=153) Country UK 48 (31%) USA 76 (50%) Australia 9 (6%) Canada 7 (5%) Other 13 (8%) Ethnic origin* White 106 (69%) Mixed-race 16 (10%) Country-defined 13 (8%) Black or ethnic minority 9 (6%) Other 3 (2%) Not specified 6 (4%) Sexuality* Heterosexual 89 (58%) Bisexual 19 (12%) Homosexual, gay, or lesbian 13 (8%) Queer or pansexual 10 (7%) Asexual 9 (6%) Other 2 (1%) Not specified 11 (7%) Religious beliefs* Christian 45 (29%) None or atheist 44 (29%) Spiritual or mixed 9 (6%) Pagan or pantheistic 8 (5%) Buddhist 4 (3%) Jewish 2 (1%) Other 7 (5%) Not specified 34 (22%) How did you hear about the study? Social media (Twitter, Tumblr, Facebook) 32 (21%) Hearing the Voice project 27 (18%) Referred by a friend 24 (16%) Other (unspecified) 21 (14%) Mental health forum or blog 18 (12%) Referred by a mental health professional 11 (7%) Lived Experience Research Network 10 (7%) Intervoice 7 (5%) Newspaper article 6 (4%) Other hearing voices groups 3 (2%) Not all patients gave all details, therefore percentages do not always sum to 100%.

) Other (unspecified) 21 (14%) Mental health forum or blog 18 (12%) Referred by a mental health professional 11 (7%) Lived Experience Research Network 10 (7%) Intervoice 7 (5%) Newspaper article 6 (4%) Other hearing voices groups 3 (2%) Not all patients gave all details, therefore percentages do not always sum to 100%. * Codes derived from free-text responses. Table 3 Nature and location of voices Number of participants (n=153) Auditory* 67 (44%) Thought-like* 14 (9%) Mixed auditory or thought-like* 56 (37%) External 69 (45%) Internal 67 (44%) Single* 10 (7%) Multiple* 124 (81%) Undifferentiated voices 39 (25%) Voice as inadequate description 30 (20%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%. * Mutually exclusive categorical codes. Table 4 Character, emotion, experiences associated with voices

Number of participants (n=153) Auditory* 67 (44%) Thought-like* 14 (9%) Mixed auditory or thought-like* 56 (37%) External 69 (45%) Internal 67 (44%) Single* 10 (7%) Multiple* 124 (81%) Undifferentiated voices 39 (25%) Voice as inadequate description 30 (20%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%. * Mutually exclusive categorical codes. Table 4 Character, emotion, experiences associated with voices Number of participants (n=153) Characteristics Characterful* 106 (69%) Not characterful* 22 (14%) Recognised individual 33 (22%) Supernatural entity 24 (16%) Simple address 16 (10%) No direct address 16 (10%) Commenting voices 18 (12%) Conversational voices 56 (37%) Commanding voices 8 (5%) Abusive and violent voices 54 (35%) Positive and helpful voices 46 (30%) Spiritual purpose 24 (16%) Emotions Fear 63 (41%) Positive 48 (31%) Neutral 49 (32%) Anxiety 47 (31%) Depression 44 (29%) Anger 32 (21%) Stress 26 (17%) Suicidal 26 (17%) Sadness 21 (14%) Shame 21 (14%) Loneliness 16 (10%) Other kinds of experiences Bodily effect* 101 (66%) No bodily effect* 41 (27%) Tiredness 10 (7%) Sleep disturbance 20 (13%) Mania 13 (8%) Paranoia 23 (15%) Musical 17 (11%) Non-verbal 21 (14%) Other hallucinations 43 (28%) Multisensory 28 (18%) Access to other minds 21 (14%) Access to other information 19 (12%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%. * Mutually exclusive categorical codes. Table 5 Causes and effects of voices

Number of participants (n=153) Characteristics Characterful* 106 (69%) Not characterful* 22 (14%) Recognised individual 33 (22%) Supernatural entity 24 (16%) Simple address 16 (10%) No direct address 16 (10%) Commenting voices 18 (12%) Conversational voices 56 (37%) Commanding voices 8 (5%) Abusive and violent voices 54 (35%) Positive and helpful voices 46 (30%) Spiritual purpose 24 (16%) Emotions Fear 63 (41%) Positive 48 (31%) Neutral 49 (32%) Anxiety 47 (31%) Depression 44 (29%) Anger 32 (21%) Stress 26 (17%) Suicidal 26 (17%) Sadness 21 (14%) Shame 21 (14%) Loneliness 16 (10%) Other kinds of experiences Bodily effect* 101 (66%) No bodily effect* 41 (27%) Tiredness 10 (7%) Sleep disturbance 20 (13%) Mania 13 (8%) Paranoia 23 (15%) Musical 17 (11%) Non-verbal 21 (14%) Other hallucinations 43 (28%) Multisensory 28 (18%) Access to other minds 21 (14%) Access to other information 19 (12%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%. * Mutually exclusive categorical codes. Table 5 Causes and effects of voices Number of participants (n=153) Voice onset Child* 52 (34%) Adolescent* 32 (21%) Adult* 29 (19%) Circumstances Positive 17 (11%) Negative 36 (24%) Traumatic 35 (23%) Substance use 10 (7%) Change, influence, and anticipation Structured change to voices 53 (35%) Change within a voice 19 (12%) Influence Can influence directly 69 (45%) Can influence indirectly 54 (35%) Cannot influence 34 (22%) Anticipation Can generally anticipate 32 (21%) Can specifically anticipate 35 (23%) Cannot anticipate 70 (46%) Continuous voices 22 (14%) Effect on personal relationships General negative effect 61 (40%) Direct negative effect 48 (31%) Positive effect 14 (9%) No effect 42 (27%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%.

cally anticipate 35 (23%) Cannot anticipate 70 (46%) Continuous voices 22 (14%) Effect on personal relationships General negative effect 61 (40%) Direct negative effect 48 (31%) Positive effect 14 (9%) No effect 42 (27%) Data are n (%). Not all patients gave all details, therefore percentages do not always sum to 100%. * Mutually exclusive categorical codes. Table 6 Characteristics of voice-hearing associated with type of nature of voices Auditory voices (n=67) Mixed voices (n=56) Internal location* 19 (28%) 33 (59%) External location 34 (51%) 28 (50%) Multisensory 8 (12%) 12 (21%) Conversational* 18 (27%) 31 (55%) Direct influence 25 (37%) 30 (54%) Structured longitudinal change* 19 (28%) 29 (52%) Access to other minds* 4 (6%) 13 (23%) Access to information 4 (6%) 11 (20%) Bodily effect 40 (60%) 41 (73%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%. * Significant associations (all p<0·05, corrected for false discovery rate). Table 7 Characteristics of voice-hearing associated with characterful voices Characterful (n=106) Not characterful (n=22) Direct influence* 60 (57%) 6 (27%) Bodily effect 74 (70%) 15 (68%) Abusive or violent 41 (39%) 3 (14%) Fear 48 (45%) 5 (23%) Anxiety 35 (33%) 6 (27%) Depression 32 (30%) 5 (23%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%. * Significant associations (all p<0·05, corrected for false discovery rate).

Characterful (n=106) Not characterful (n=22) Direct influence* 60 (57%) 6 (27%) Bodily effect 74 (70%) 15 (68%) Abusive or violent 41 (39%) 3 (14%) Fear 48 (45%) 5 (23%) Anxiety 35 (33%) 6 (27%) Depression 32 (30%) 5 (23%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%. * Significant associations (all p<0·05, corrected for false discovery rate). Table 8 Characteristics of voice-hearing associated with bodily effect Bodily effect (n=101) No bodily effect (n=41) Multisensory 21 (21%) 5 (12%) Positive or useful 25 (25%) 18 (44%) Abusive or violent* 43 (43%) 7 (17%) Traumatic circumstances 28 (28%) 4 (10%) Fear 47 (47%) 13 (32%) Anxiety 35 (35%) 8 (20%) Shame 17 (17%) 1 (2%) Anticipation* 48 (48%) 9 (22%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%. * Significant associations (all p<0·05, corrected for false discovery rate). Table 9 Characteristics of voice-hearing associated with diagnosis Clinical (n=127) Non-clinical (n=26) Auditory 52 (41%) 15 (58%) Positive and useful voices 34 (27%) 12 (46%) Abusive and violent voices 49 (39%) 5 (19%) Fear* 60 (47%) 3 (12%) Anxiety 41 (32%) 6 (23%) Depression* 43 (34%) 1 (4%) Bodily effect 87 (69%) 14 (54%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%.

sive and violent voices 49 (39%) 5 (19%) Fear* 60 (47%) 3 (12%) Anxiety 41 (32%) 6 (23%) Depression* 43 (34%) 1 (4%) Bodily effect 87 (69%) 14 (54%) Data are n (%). Percentages are for participants within a subgroup receiving that code. Not all patients gave all details, therefore percentages do not always sum to 100%. * Significant associations (all p<0·05, corrected for false discovery rate). Panel 1 Nature of experiences Auditory “[M]ost of the time I can hear it like it was just someone standing next to me. It's a different feeling than when you think words inside of your head, when you think inside your head your voice isn't distinct like it is when you speak out loud. You think words, not tone. But there is definite distinct tone and individuality that's unfamiliar with the voices.” Thought-like “I did not hear the voices aurally. They were much more intimate than that, and inescapable. It's hard to describe how I could ‘hear’ a voice that wasn't auditory; but the words the voices used and the emotions they contained (hatred and disgust) were completely clear, distinct, and unmistakable, maybe even more so than if I had heard them aurally.” Mixed “I have all kinds of voice-type experiences […] Some are voices that are clearly in my head but which feel ‘different’ from my own thoughts. Some are voices that seem to come from outside but which I know don't.” Panel 2 Research in context Systematic review

“I did not hear the voices aurally. They were much more intimate than that, and inescapable. It's hard to describe how I could ‘hear’ a voice that wasn't auditory; but the words the voices used and the emotions they contained (hatred and disgust) were completely clear, distinct, and unmistakable, maybe even more so than if I had heard them aurally.” Mixed “I have all kinds of voice-type experiences […] Some are voices that are clearly in my head but which feel ‘different’ from my own thoughts. Some are voices that seem to come from outside but which I know don't.” Panel 2 Research in context Systematic review Before constructing the survey, we did a systematic review of the published literature on hallucinations across diagnostic (and non-clinical) populations. We initially employed the search terms “phenomenology” and “hallucinations”— where possible also limiting the methods employed to “qualitative”—across the PsycINFO and PubMed databases. These searches returned 237 and 125 initial articles, respectively. Each article abstract was then reviewed individually; of those directly relevant to our project (ie, moderate to large-sample [n>50] phenomenological studies of auditory or verbal hallucinations), we searched cited references to identify any additional relevant articles, in addition to future articles that used the base article as a reference. We searched cited references until no additional articles of relevance could be identified. Although we were able to identify a sub-set of articles employing structured or semi-structured measures and comparing the phenomenology of hallucinations across specific diagnostic groups (eg, Parkinson's disease vs schizophrenia), we did not identify any published studies that simultaneously surveyed both clinical and non-clinical individuals; included individuals with any diagnosis (psychiatric, neurological, or medical); and used open-ended (unstructured) prompts.

ns across specific diagnostic groups (eg, Parkinson's disease vs schizophrenia), we did not identify any published studies that simultaneously surveyed both clinical and non-clinical individuals; included individuals with any diagnosis (psychiatric, neurological, or medical); and used open-ended (unstructured) prompts. Interpretation We report the findings of what is, to our knowledge, the largest open-ended survey of the phenomenology of voices and voice-like events in the published scientific literature. We departed from other large-sample qualitative studies of auditory hallucination by targeting a diverse, naturalistic sample of individuals with and without clinical histories and with a broad range of (self-reported) diagnoses. Potentially important new findings concern the association between acoustic perception and thought, somatic and multisensorial features of auditory hallucinations, and the link between auditory hallucinations and characterological entities. Awareness and further investigation of these characteristics has substantial implications for experimental and applied clinical research programmes, especially with respect to further development of interventions targeting the way voice-hearers relate to their voices.

Introduction Depression is the leading contributor to the worldwide burden of disease in young people aged 10–24 years.1 Rates of depression increase substantially during adolescence,2, 3 a period of transition characterised by social, emotional, and physiological changes. Understanding of specific risk factors during this important transition is therefore needed to inform prevention strategies. During this period, peers are the main sources for social comparison and appraisal, and self-consciousness is heightened.4 Studies have shown that peers report similar levels of depressive symptoms to each other,5 although evidence for peer contagion effects on depression and related phenotypes (eg, self-harm) is mixed.6, 7, 8, 9, 10 Identification of young people at heightened risk of depression and related phenotypes is a key area for future research.10 A strong and robust increased risk of self-harm and attempted suicide has been reported in young people identifying with contemporary goth subculture and related subcultures generally described as alternative youth.11 A goth is defined in the Oxford Dictionary as “a member of a subculture favouring black clothing, white and black make-up, and goth music”. Much diversity exists within the goth subculture, making definition of the average adolescent goth difficult;12 however, many social norms are associated with being a goth, including alternative clothing and music, and a dark, morbid mood and aesthetic. The goth subculture has been suggested to provide an important source of validation and community to individuals who do not conform with societal norms.12

scent goth difficult;12 however, many social norms are associated with being a goth, including alternative clothing and music, and a dark, morbid mood and aesthetic. The goth subculture has been suggested to provide an important source of validation and community to individuals who do not conform with societal norms.12 Why affiliation with goth subculture is associated with an increased risk of self-harm and whether it is also associated with increased depression is unclear. If the observed association represents underlying social transmission, such as emulation of subcultural icons or self-harming peers, we might postulate that the association would be specific to self-harm. A contrasting explanation is that the observed association represents shared exposure to stressors, or social selection mechanisms, whereby vulnerable young people are attracted to others with similar underlying risks. In this case, goth affiliation would be associated with related phenotypes (specifically depression) in addition to self-harm and suicidal ideation, and observed associations would be confounded by baseline characteristics of young people before identification as a goth, and concurrent risk factors that might be specific to this group (eg, peer victimisation). Notably, the study that identified an association between goth subculture identification and deliberate self-harm by Young and colleagues11 did not adjust for several salient early risk factors that could have confounded the association, including specific emotional and behavioural problems, peer victimisation, and maltreatment. With these issues in mind, we aimed to test whether self-identification with the goth subculture at 15 years of age was associated with self-harm and depression at 18 years of age in a large cohort of young people followed up prospectively.

cific emotional and behavioural problems, peer victimisation, and maltreatment. With these issues in mind, we aimed to test whether self-identification with the goth subculture at 15 years of age was associated with self-harm and depression at 18 years of age in a large cohort of young people followed up prospectively. Research in context Evidence before this study We searched PubMed to identify potential literature published before Nov 5, 2014, using the search terms “goth or emo or subculture” and “depress*, suic*, self-harm, or mental health”. We identified 98 articles, of which 2 examined an association between goth or “alternative youth” affiliation and depression or self-harm, only one of which was prospective in design. Added value of this study To our knowledge, our study is the largest so far to prospectively examine the association between self-identification with the goth subculture and later self-harm and depression. Our study was designed to address some of the limitations in the original study by adjusting for the potential confounding effects of specific emotional and behavioural problems, peer victimisation, and maltreatment. Implications of all the available evidence

To our knowledge, our study is the largest so far to prospectively examine the association between self-identification with the goth subculture and later self-harm and depression. Our study was designed to address some of the limitations in the original study by adjusting for the potential confounding effects of specific emotional and behavioural problems, peer victimisation, and maltreatment. Implications of all the available evidence A strong, dose–response association between identification as “alternative” or “goth” and self-harm has now been reported in samples from Scotland, Germany, and England. Our study also identified a strong association with self-identification as a goth and adult depression. Together, these findings suggest that youths who identify with the goth subculture might represent a vulnerable group, although establishing causal links from these observational studies is not possible.

and England. Our study also identified a strong association with self-identification as a goth and adult depression. Together, these findings suggest that youths who identify with the goth subculture might represent a vulnerable group, although establishing causal links from these observational studies is not possible. Methods Study design and participants The Avon Longitudinal Study of Parents and Children (ALSPAC) is a longitudinal cohort study that recruited pregnant women resident in the former county of Avon, UK, who had expected dates of delivery between April 1, 1991, and Dec 31, 1992 (the administrative county of Avon was abolished in 1996). 14 541 pregnant women initially enrolled and returned at least one questionnaire or attended a Children in Focus clinic. Of these 14 541 pregnancies, 68 have no known birth outcome; of the remaining 14 472 pregnancies, 195 were twins, three were triplets, and one was quadruplets, resulting in 14 676 known fetuses. The 14 541 pregnancies resulted in 14 062 live born children, of whom 13 988 were alive at 1 year of age (74 infants died). When the oldest children were about 7 years of age, an attempt was made to bolster the initial sample with eligible participants who had declined to join the study originally; a further 456 children were enrolled at this point. As a result, when taking into account the variables obtained from the age of 7 years onwards (and potentially abstracted from obstetric notes), data were available for more than the 14 541 pregnancies mentioned. The phases of enrolment are described in detail in the cohort profile report.13 A fully searchable data dictionary is available online .

o account the variables obtained from the age of 7 years onwards (and potentially abstracted from obstetric notes), data were available for more than the 14 541 pregnancies mentioned. The phases of enrolment are described in detail in the cohort profile report.13 A fully searchable data dictionary is available online . Children were invited to attend annual assessment clinics, including face-to-face interviews and psychological and physical tests from the age of 7 years onwards. In this analysis, we included children who were still participating in the study and who completed a computer-based survey asking them about their self-identification with eight different social groups during clinic visits at 15 years of age. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees. From the beginning of the study, ALSPAC has had its own Ethics and Law Committee. Initial ethical approval was obtained for gaining written, informed consent from pregnant mothers. At each follow-up clinic assessment, mothers provided informed written consent, and children provided assent after receiving a full explanation of the study. Written informed consent was obtained from participants for the clinic at 18 years.

oval was obtained for gaining written, informed consent from pregnant mothers. At each follow-up clinic assessment, mothers provided informed written consent, and children provided assent after receiving a full explanation of the study. Written informed consent was obtained from participants for the clinic at 18 years. Procedures We adapted the Peer Crowd Questionnaire14, 15 on the basis of interviews with focus groups of 14-year-old adolescents (seven girls, four boys) from a large comprehensive school in south Bristol, in the former county of Avon in January, 2006, to identify salient youth subcultures at that time and location. We identified eight different social groups: “sporty”, “populars”, “skaters”, “chavs”, “loners”, “keeners”, “bimbos”, and “goths”. At 15 years of age, study participants were invited to attend the research clinic for the annual study follow-up visit. During the clinic, participants completed a computer-based survey in which they were asked a series of questions about their self-identification with these social groups. For self-identification as a goth, participants were asked, “is there a group of teens in your school or neighbourhood with the reputation of…rebelling against the norm (in clothing or ideas, for example), or in attempting not to conform to social ideals (eg, the ‘goths’)? How much do you identify with...the goths?” Participants responded “not at all”, “not very much”, “somewhat”, “more than somewhat”, or “very much”.

or neighbourhood with the reputation of…rebelling against the norm (in clothing or ideas, for example), or in attempting not to conform to social ideals (eg, the ‘goths’)? How much do you identify with...the goths?” Participants responded “not at all”, “not very much”, “somewhat”, “more than somewhat”, or “very much”. Outcomes The outcomes of the study were depression and self-harm at 18 years of age. We measured the prospective associations between self-identification with the goth subculture at 15 years of age and depression and self-harm at 18 years of age. At the 15 year visit, we assessed depressive mood with the Development and Wellbeing Assessment (DAWBA).16 To account for the full range of symptoms of depression at the age of 15 years, we derived a total symptom count, summing all of the symptoms from the depression section of the DAWBA. At the 18 year visit, participants completed a self-administered computerised version of the Clinical Interview Schedule-Revised (CIS-R)17 to assess depression. CIS-R is designed for, and has been widely used within, community samples.18 We used a binary variable (depressed or not depressed) to record depression; cases were those meeting the criteria for a primary diagnosis of mild, moderate, or severe depression with the CIS-R, which generates these diagnoses with algorithms based on the International Classification of Diseases (ICD)-10 criteria.

s.18 We used a binary variable (depressed or not depressed) to record depression; cases were those meeting the criteria for a primary diagnosis of mild, moderate, or severe depression with the CIS-R, which generates these diagnoses with algorithms based on the International Classification of Diseases (ICD)-10 criteria. We assessed self-harm with self-report at the 15 year and 18 year research clinics. Self-harm was assessed at the 15 year research clinic with an item from the DAWBA: “over the whole of your lifetime have you ever tried to harm or hurt yourself?” At the 18 year clinic, we assessed self-harm with CIS-R; participants were classified as having a lifetime history of self-harm if they responded positively to the question “have you ever hurt yourself on purpose in any way (eg, by taking an overdose of pills or by cutting yourself)?” We did not distinguish between individuals who had harmed themselves with and without suicidal intent in the present study.

having a lifetime history of self-harm if they responded positively to the question “have you ever hurt yourself on purpose in any way (eg, by taking an overdose of pills or by cutting yourself)?” We did not distinguish between individuals who had harmed themselves with and without suicidal intent in the present study. We assessed parental occupational social class on the basis of the lower of the mother's or mother's partner's occupational social class,19 dichotomised into professional, managerial, or skilled professions versus partly or unskilled occupations. We coded highest maternal education as the percentage of mothers with a university degree versus those without a degree. Both of these variables were measured during pregnancy. Maternal depression (assessed during pregnancy, at 18 weeks' gestation) was measured using a postal questionnaire based on the Edinburgh Postnatal Depression Scale20 (original internal consistency, a=0·87; in the present study, a=0·85). Maternal history of severe depression (self-report of past history of severe depression, yes or no) was also assessed by postal questionnaire at 12 weeks' gestation.

ed using a postal questionnaire based on the Edinburgh Postnatal Depression Scale20 (original internal consistency, a=0·87; in the present study, a=0·85). Maternal history of severe depression (self-report of past history of severe depression, yes or no) was also assessed by postal questionnaire at 12 weeks' gestation. Peer victimisation was assessed when children were 8 and 10 years of age using a modified version of the Bullying and Friendship Interview Schedule.21 Children's exposure to life events at 3·5 years was assessed using maternal report. The items included in this questionnaire were taken from other studies.22, 23 Children's internalising and externalising problems were assessed using maternal reports from the Strengths and Difficulties Questionnaire (SDQ)24 when children were 11 years of age. Previous depression, defined as scores reaching clinical significance (scores of 11 or more on the Short Moods and Feelings Questionnaire, yes vs no), was assessed via self-report when children were 10, 13, and 16 years of age.25 This scale has been validated against the CIS-R.26 Any axis-1 disorders classified using the Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV (no vs yes) were assessed using maternal reports from the DAWBA27 when children were 10 years old. Diagnoses were derived with a computer algorithm, which has been shown to be similar in accuracy to a clinical rating method.16

disorders classified using the Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV (no vs yes) were assessed using maternal reports from the DAWBA27 when children were 10 years old. Diagnoses were derived with a computer algorithm, which has been shown to be similar in accuracy to a clinical rating method.16 Children's self-perception was self-reported using a shortened form of Harter's Self Perception Profile for Children28 during a clinic when children were on average aged 8·5 years. This measure consisted of 12 items relating to worldwide self-worth and scholastic competence. We gave the Emotionality, Activity, and Sociability Temperament Scale29 questionnaire to mothers when children were aged 6 years old and this scale had four subscales (emotionality, shyness, sociability, and activity). Statistical analysis We examined the univariable prospective association between identification with goth subculture and depression in logistic regression analyses, then adjusted for baseline depression at 15 years (using the continuous score derived from the DAWBA), and baseline self-harm at 15 years to minimise the possibility of reverse causality. We then additionally adjusted for a range of individual, family, and social confounders, including previous depression, internalising and externalising difficulties at 11 years, self-perception, victimisation by bullies, antenatal depression, maternal history of depression, maternal education, and temperament (emotionality and activity).

y adjusted for a range of individual, family, and social confounders, including previous depression, internalising and externalising difficulties at 11 years, self-perception, victimisation by bullies, antenatal depression, maternal history of depression, maternal education, and temperament (emotionality and activity). We selected covariates a priori from the scientific literature on the basis of their potential to confound the association between self-identification as a goth, depression, and self-harm.3, 12, 30 When data were subject to particularly high attrition, we used the earliest measure for the covariate, to maximise our sample size. We adjusted for depression at 15 years of age, and for many of the same confounding variables as the original study by Young and colleagues11 and a range of further confounding variables, including baseline self-harm, early emotional and behavioural difficulties, psychiatric disorder, peer problems, and child maltreatment.

djusted for depression at 15 years of age, and for many of the same confounding variables as the original study by Young and colleagues11 and a range of further confounding variables, including baseline self-harm, early emotional and behavioural difficulties, psychiatric disorder, peer problems, and child maltreatment. Similarly to most prospective studies, missing data because of attrition were a concern. We used a sample with complete data across all exposure, outcome, and confounding variables (n=2351) to investigate main and independent effects of self-identification with the goth subculture. To address the possibility of bias, we used collected data to predict and impute missing variables and did analyses with imputed datasets, allowing individuals with incomplete data to be included in the analyses. We used a fully conditional specification as implemented in the Multiple Imputation by Chained Equation algorithm in Stata version 12.31, 32

s, we used collected data to predict and impute missing variables and did analyses with imputed datasets, allowing individuals with incomplete data to be included in the analyses. We used a fully conditional specification as implemented in the Multiple Imputation by Chained Equation algorithm in Stata version 12.31, 32 The ALSPAC sample has substantial information on sociodemographic variables that predict missing data, allowing the construction of an imputation model using strong auxiliary information. Because missing data in the ALSPAC study have been previously found to be dependent on several variables (eg, gender, parental social, and maternal education), these variables were included in the imputation models, in addition to other measures that have been identified as closely associated with adolescent self-harm and with our outcome variables (eg, similar measures from the domains of mental health or substance use obtained earlier in the study) and all other variables included in analyses (appendix). Variable estimates were averaged from 100 imputed datasets using Rubin's rules.32 In longitudinal studies, earlier measures of child depression can be used to predict later depression,33 allowing us to impute up to a starting sample of 5342, those with at least one measure of adolescent depression and complete exposure data. The imputations were separated into these two stages to establish that extending the model to those without any earlier depression data produced similar results. Analyses were done using Stata version 12 (StataCorp, TX, USA). We deemed p values less than 0·05 as significant.

scent depression and complete exposure data. The imputations were separated into these two stages to establish that extending the model to those without any earlier depression data produced similar results. Analyses were done using Stata version 12 (StataCorp, TX, USA). We deemed p values less than 0·05 as significant. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. LBo, RC, RP, and JH had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results 10 962 adolescents in the ALSPAC study were invited to attend the 15 year clinic visit, of whom 5515 attended (average age of 15·5 years [SD 0·4]) and 5357 completed the survey about subculture identification. From the eight subcultures identified, the goth subculture was recognised by all adolescents interviewed. The most frequently endorsed subcultures were “sporty”, “populars”, “skaters”, “chavs”, and “goths”. At the 18 year visit, 3694 adolescents attended and provided outcome data (mean age 17·8 years [SD 0·5]). Overall, subculture identification and outcome data at 18 years were available for 3694 adolescents (figure 1).

. The most frequently endorsed subcultures were “sporty”, “populars”, “skaters”, “chavs”, and “goths”. At the 18 year visit, 3694 adolescents attended and provided outcome data (mean age 17·8 years [SD 0·5]). Overall, subculture identification and outcome data at 18 years were available for 3694 adolescents (figure 1). Participants lost to follow-up between the 15 and 18 year visits were no more likely to self-identify as a goth (χ2 0·558, p=0·455), to have self-harmed by the age of 15 years (χ2 2·001, p=0·157), or to have shown high levels of depressive symptoms at 15 years of age (χ2 9·713, p=0·205) than were those with data for all variables. Table 1 shows how individuals who identified as goths “more than somewhat” to “very much” differed in early individual and family characteristics (assessed between the age of 8 years and 15 years) compared with young people who reported they “did not” or “somewhat” self-identify with the goth subculture. Those who identified with the goth subculture were more likely to be girls. They were also more likely to have mothers with a history of depression, report being bullied at the age of 8 and 10 years, and have a history of emotional and behavioural difficulties, according to mother's reports from the SDQ, including symptoms of depression and anxiety, hyperactivity, and peer relationship difficulties. Young people who self-identified with the goth subculture also self-reported more symptoms of depression at 10, 12, and 13 years of age. However, generally, the magnitude of these differences was small and unlikely to be clinically significant for most participants; p values should be interpreted with caution in view of the large sample and number of comparisons made.

culture also self-reported more symptoms of depression at 10, 12, and 13 years of age. However, generally, the magnitude of these differences was small and unlikely to be clinically significant for most participants; p values should be interpreted with caution in view of the large sample and number of comparisons made. In our sample of young people with complete data for goth self-identification and depression at 18 years, 105 (6%) of the 1841 individuals who responded that they did not at all identify with the goth subculture had depression scores in the clinical range at 18 years compared with 47 (9%) of the 523 young people who identified with goth subculture “somewhat” and 28 (18%) of the 154 individuals who responded that they identified with the goth subculture “very much” (figure 2).

ey did not at all identify with the goth subculture had depression scores in the clinical range at 18 years compared with 47 (9%) of the 523 young people who identified with goth subculture “somewhat” and 28 (18%) of the 154 individuals who responded that they identified with the goth subculture “very much” (figure 2). Table 2 shows clear dose–response associations between the extent to which young people self-identified as a goth and depression scores in the clinical range at 18 years. For example, compared with young people who did not identify as a goth, those who somewhat identified as being a goth were 1·6 times more likely to have scores in the clinical range for depression at 18 years (unadjusted odds ratio [OR] 1·63, 95% CI 1·14–2·34, p<0·001) and were more than three times as likely to have scores in the clinical range for depression at 18 years (unadjusted OR 3·67, 2·33–4·79, p<0·001). For each unit increase in goth affiliation, the (unadjusted) odds of depression increased by 1·36 (1·23–1·49). Adjustment for potential confounders led to only a fairly small attenuation of this OR in the final model (OR 1·27, 1·11–1·47; table 3). As expected, we noted the largest attenuation after adjustment for baseline symptoms of depression (OR 1·29, 1·15–1·45).

(unadjusted) odds of depression increased by 1·36 (1·23–1·49). Adjustment for potential confounders led to only a fairly small attenuation of this OR in the final model (OR 1·27, 1·11–1·47; table 3). As expected, we noted the largest attenuation after adjustment for baseline symptoms of depression (OR 1·29, 1·15–1·45). We also identified evidence of a dose–response association between goth identification and self-harm at 18 years of age. For example, compared with young people who did not identify as a goth, those who indicated that they somewhat identified as being a goth were 2·33 times more likely to report having self-harmed, whereas those who very much identified themselves as being a goth were more than five times more likely to report self-harm. For each unit increase in affiliation with goth subculture, the (unadjusted) OR of self-harm increased by 1·52 (1·42–1·63) The OR decreased by 12·5% in the adjusted model (OR 1·33, 1·19–1·48; table 3).

s those who very much identified themselves as being a goth were more than five times more likely to report self-harm. For each unit increase in affiliation with goth subculture, the (unadjusted) OR of self-harm increased by 1·52 (1·42–1·63) The OR decreased by 12·5% in the adjusted model (OR 1·33, 1·19–1·48; table 3). Of all subcultures identified, 154 young people who identified with the goth subculture very much were most at risk of depression and self-harm, with 28 (18%) having scores in the clinical range for depression, and 57 (37%) reporting self-harm at 18 years of age. Of the 341 young people who identified as skaters very much, 37 (11%) had depression, and 85 (25%) reported self-harm at 18 years of age; and of the 47 adolescents who self-identified as loners, 4 (9%) had depression scores in the clinical range, and 12 (26%) reported self-harm. Those who self-identified as sporty were least likely to have depression and to have self-harmed at age 18 years (31 [4%] of 786 sporty individuals had depression and 47 [6%] had self-harmed). Discussion In this analysis of data from a longitudinal cohort study, we noted a dose–response association between the extent to which young people self-identified with the goth subculture at the age of 15 years and both depression and self-harm at 18 years of age. This association was independent of the potentially confounding characteristics of previous depressive symptoms and self-harm, personality, history of bullying, behavioural issues, maternal depression, and perception of body image.

subculture at the age of 15 years and both depression and self-harm at 18 years of age. This association was independent of the potentially confounding characteristics of previous depressive symptoms and self-harm, personality, history of bullying, behavioural issues, maternal depression, and perception of body image. Young people who self-identified as goths were more likely to be girls (contrasting with the findings from the original Young and colleagues' study sample in Glasgow in which they were more likely to be boys11), to have mothers with a history of depression, to have a history of emotional issues, including depression themselves, and to report issues with peers, including being bullied. Such vulnerability factors for depression suggest a degree of self-selection, with young people more susceptible to depression and self-harm being more likely to be attracted to the goth subculture. Yet, even after adjustment for these early risk factors, young people who self-identified as goths remained at an increased risk of depression and self-harm compared with those who did not identify with the subculture. Although some residual confounding is likely to remain, our findings support earlier evidence11 that goths represent a vulnerable group.

e early risk factors, young people who self-identified as goths remained at an increased risk of depression and self-harm compared with those who did not identify with the subculture. Although some residual confounding is likely to remain, our findings support earlier evidence11 that goths represent a vulnerable group. Our study has several strengths, including the large sample size, the prospective design from before birth to 18 years (a time when rates of depression peak), and the detailed information on a range of potential confounding variables. Individuals who are susceptible to depression might be more drawn to subcultures, such as the goth subculture, which are known to embrace marginalised individuals from all backgrounds, including those with previous mental health difficulties. Thus, as we originally postulated, the reported association between goth affiliation and depression could be due to social selection factors not addressed adequately in the previous study.11 To investigate this, we adjusted the analysis for concurrent depressive symptoms and controlled for the effects of previous victimisation, emotional and behavioural issues, and a range of other potential confounders to test whether noted associations remained. Although we adjusted for many potential confounders, we cannot exclude the possibility of residual confounding. A second limitation is the loss to follow-up from the original ALSPAC sample. Young adults who attended the clinic at 18 years of age came from families with higher levels of education and social class, which might have reduced statistical power in detection of an association between goth affiliation at 15 years and depression and self-harm at 18 years. The wealth of data about participants who have not been followed up in ALSPAC makes the assumptions behind our handling of missing data much more reasonable, and the results of our imputation analyses were consistent with our complete case findings. We therefore think that such a strong association is unlikely to be explained by attrition. The questions used to assess self-harm at the age of 15 and 18 years were worded differently and could have led to different responses by participants. Our definition of self-harm included individuals who had self-harmed with and without suicidal intent.

ch a strong association is unlikely to be explained by attrition. The questions used to assess self-harm at the age of 15 and 18 years were worded differently and could have led to different responses by participants. Our definition of self-harm included individuals who had self-harmed with and without suicidal intent. The extent to which suicidal and non-suicidal self-harm represent distinct concepts or more or less extreme versions of the same behaviour is a source of debate.34, 35, 36, 37, 38 Further research should examine whether associations with goth subculture and self-harm differ according to self-reported suicidal intent. Finally, our findings support those of Scottish11 and German39 young people, reporting a link between goth affiliation and self-harm; however, whether our findings generalise beyond these populations is unclear. The meaning and cultural identity of goths are likely to vary within and across cultures and time. Only eight subcultures were investigated and therefore the role of young people who identify with rarer youth subcultures is not known and is an important direction of future research.

beyond these populations is unclear. The meaning and cultural identity of goths are likely to vary within and across cultures and time. Only eight subcultures were investigated and therefore the role of young people who identify with rarer youth subcultures is not known and is an important direction of future research. Peer contagion might represent one mechanism by which young people who affiliate with other at-risk goths might be at increased risk of depression and self-harm. Evidence of peer contagion effects for both depression and self-harm have been previously reported.40 Although corumination, excessive reassurance seeking, and negative feedback seeking might represent mechanisms through which peer contagion increases risk of depression,41 why peer contagion might operate to increase risk of self-harm is unclear. Such effects might arise if young people actively discuss self-harming as an effective emotion regulation strategy. In a study by Young and colleagues,39 young people who identified with “alternative” subcultures (including goths) were more likely to report autonomic reasons for self-harming (ie, to reduce negative emotions), including to “stop bad feelings” and to “relieve feeling numb or empty” compared with non-alternative young people. Alternative young people were also more likely to endorse social reasons for self-harming (eg, to “feel more part of a group”). Emulation of peer behaviour has also been suggested,40 and in the context of the goth subculture, this might have some validity. Information about exposure to self-harm in others was not available in this study but is an important area for future research.

reasons for self-harming (eg, to “feel more part of a group”). Emulation of peer behaviour has also been suggested,40 and in the context of the goth subculture, this might have some validity. Information about exposure to self-harm in others was not available in this study but is an important area for future research. Well validated experimental manipulations that use sad music to induce symptoms of depression exist,36, 37 although effects in the laboratory are short lasting. Listening to repeated music from the goth genre might lower mood and exacerbate symptoms of depression. An alternative explanation for our findings could be an affiliation or attraction model in which the extent to which young people self-identify with the goth subculture might represent the extent to which at-risk young people feel isolated, ostracised, or stigmatised by society. Such young people might be attracted to other alternative goth young people who do not adhere to societal norms. The registering of hate crimes committed against goths and other subcultures by Manchester police after the murder of 20-year-old goth Sophie Lancaster in 2007 suggests that goths might be the target of social stigma and aggression.

e attracted to other alternative goth young people who do not adhere to societal norms. The registering of hate crimes committed against goths and other subcultures by Manchester police after the murder of 20-year-old goth Sophie Lancaster in 2007 suggests that goths might be the target of social stigma and aggression. Although our findings suggest that youths who identify with the goth subculture might represent a vulnerable group, our observational findings cannot be used to claim that becoming a goth causes an increased risk of self-harm and depression. Although peer contagion might operate within the goth youth community, other factors such as stigma and social ostracising might represent the underlying mechanisms of increased risk. Working with youths in the goth community to identify those at risk of depression and self-harm and provide support might be effective. Public campaigns to reduce stigma and aggression targeted to individuals from diverse subcultures might also be important. Supplementary Material Supplementary appendix Supplementary audio Lucy Bowes and Niall Boyce discuss the association between goth subculture identification, depression, and self-harm

Although our findings suggest that youths who identify with the goth subculture might represent a vulnerable group, our observational findings cannot be used to claim that becoming a goth causes an increased risk of self-harm and depression. Although peer contagion might operate within the goth youth community, other factors such as stigma and social ostracising might represent the underlying mechanisms of increased risk. Working with youths in the goth community to identify those at risk of depression and self-harm and provide support might be effective. Public campaigns to reduce stigma and aggression targeted to individuals from diverse subcultures might also be important. Supplementary Material Supplementary appendix Supplementary audio Lucy Bowes and Niall Boyce discuss the association between goth subculture identification, depression, and self-harm Acknowledgments We thank all the families who took part in this study, the midwives for help in recruiting them, and the whole ALSPAC team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists, and nurses. The UK Medical Research Council and the Wellcome Trust (grant number 102215/2/13/2) and the University of Bristol provided core support for ALSPAC. Our research was specifically funded by a Wellcome Trust grant held by GL (grant number 084268/Z/07/Z) and a Medical Research Council Programme grant held by BMau (grant number G0500953). RP and BMar are both supported by Elizabeth Blackwell Institute for Health Research Institutional Wellcome Strategic Awards. LBo is supported by a Leverhulme Early Career Research Fellowship. The ALSPAC data resource is publicly available online.

al Research Council Programme grant held by BMau (grant number G0500953). RP and BMar are both supported by Elizabeth Blackwell Institute for Health Research Institutional Wellcome Strategic Awards. LBo is supported by a Leverhulme Early Career Research Fellowship. The ALSPAC data resource is publicly available online. Contributors BMau designed the goth self-identification questionnaire, GL developed the CIS-R, JH conceptualised the manuscript. RC, RP, JH, and LBo wrote the statistical analysis plan and all authors act as guarantors for the manuscript. RC and RP formatted and analysed the data, LBo drafted the manuscript, and all authors read, drafted, and revised the whole report. Declaration of interests We declare no competing interests. Figure 1 Flow chart of ALSPAC study participants ALSPAC=The Avon Longitudinal Study of Parents and Children. CIS-R=Clinical Interview Schedule-Revised. Figure 2 Percentage of sample depressed according to affiliation with goth subculture based on the trend in table 2 Table 1 Baseline characteristics of study participants according to goth self-identification

Figure 1 Flow chart of ALSPAC study participants ALSPAC=The Avon Longitudinal Study of Parents and Children. CIS-R=Clinical Interview Schedule-Revised. Figure 2 Percentage of sample depressed according to affiliation with goth subculture based on the trend in table 2 Table 1 Baseline characteristics of study participants according to goth self-identification Low goth self-identification (n=4709) High goth self-identification (n=648) p value Gender Female 2453 (52%) 374 (58%) 0·01 Male 2256 (48%) 274 (42%) 0·01 Parental social class High (professional, managerial, or skilled occupations) 2072 (44%) 286 (44%) 0·94 Low (partly skilled or unskilled occupations) 2637 (56%) 362 (56%) 0·95 Mother with degree 871 (18%) 119 (18%) 0·94 Maternal depression score during pregnancy 6·5 (4·8) 7·0 (5·1) 0·02 Maternal history of severe depression 283 (6%) 64 (10%) <0·001 Bullying Bullied at age of 8 years 1822 (39%) 307 (47%) <0·001 Bullied at age of 13 years 542 (12%) 106 (16%) 0·02 Life events score at age 3·5 years 2·7 (2·5) 3·1 (2·8) <0·001 Strengths and difficulties (age 12 years) Prosocial score 8·4 (1·6) 8·3 (1·7) 0·13 Hyperactivity score 2·6 (2·1) 2·8 (2·2) 0·04 Emotional symptoms score 1·4 (1·7) 1·6 (1·7) 0·03 Conduct problems score 1·1 (1·3) 1·1 (1·3) 0·24 Peer problems score 1·0 (1·4) 1·3 (1·6) <0·001 Total difficulties score 5·8 (4·4) 6·5 (4·7) <0·001 MFQ score >11 MFQ>11 at age of 10 years 198 (4%) 27 (4%) 0·97 MFQ>11 at age 12·5 years 231 (5%) 51 (8%) <0·001 MFQ>11 at 13·5 years 364 (8%) 115 (18%) <0·001 DAWBA depression score at age of 15 years 3·2 (4·5) 5·7 (6·7) <0·001 Self-harm at age of 15 years 457 (10%) 166 (26%) <0·001 Self-perception at age of 11 years Kind 3899 (83%) 509 (79%) <0·001 Friendly 4309 (92%) 584 (90%) 0·02 Confident 2731 (58%) 360 (56%) <0·001 Sporty 2816 (60%) 262 (40%) <0·001 Good looking 1681 (36%) 190 (29%) <0·001 Easily bored 1333 (28%) 226 (35%) 0·01 Different from others 739 (16%) 222 (34%) <0·001 Worries a lot 782 (17%) 154 (24%) <0·001 Messes about a lot 1267 (27%) 211 (33%) 0·02 Temperament at age of 6 years (child in top quartile) Emotionality 749 (16%) 157 (24%) <0·001 Activity 739 (16%) 80 (12%) 0·20 Shyness 871 (18%) 120 (19%) 0·94 Sociability 876 (19%) 123 (19%) 0·17 Data are n (%) or mean (SD). Self-identification by binary variable of low self-identification (not at all to somewhat) or high self-identification (more than somewhat to very much).

onality 749 (16%) 157 (24%) <0·001 Activity 739 (16%) 80 (12%) 0·20 Shyness 871 (18%) 120 (19%) 0·94 Sociability 876 (19%) 123 (19%) 0·17 Data are n (%) or mean (SD). Self-identification by binary variable of low self-identification (not at all to somewhat) or high self-identification (more than somewhat to very much). Table 2 Odds ratio for depression and self-harm at 18 years for each category of goth identification Number of participants (n=5357) Number of participants with data for goth self-identification at 15 years and outcomes at 18 years (n=3694) Depression Self-harm OR (95% CI) p value OR (95% CI) p value Linear effect* 5357 3694 1·36 (1·23–1·49) p<0·001 1·52 (1·42–1·63) p<0·001 Not at all 2759 1841 Reference .. Reference .. Not very much 1234 884 1·16 (0·83–1·62) .. 1·52 (1·20–1·93) .. Somewhat 716 523 1·63 (1·14–2·34) .. 2·33 (1·80–3·02) .. More than somewhat 410 292 2·33 (1·56–3·47) .. 3·65 (2·72–4·89) .. Very much 238 154 3·67 (2·33–4·79) .. 5·14 (3·58–7·36) .. LR test 5357 3694 χ2 38·99 p<0·001 χ2 132·65 p<0·001 OR=odds ratio. LR=likelihood ratio. * Goth identification was treated as a continuous variable; therefore, the linear effect refers to OR of depression for a one point increase in goth identification. The LR test provides a test for the dose–response trend. Table 3 Odds ratio for depression and self-harm given for each category increase in goth identification

* Goth identification was treated as a continuous variable; therefore, the linear effect refers to OR of depression for a one point increase in goth identification. The LR test provides a test for the dose–response trend. Table 3 Odds ratio for depression and self-harm given for each category increase in goth identification Depression Self-harm OR (95% CI) p value OR (95% CI) p value Unadjusted (n=3694) 1·36 (1·23–1·49) p<0·001 1·52 (1·42–1·63) p<0·001 Complete cases (n=2351) 1·39 (1·22–1·58) p<0·001 1·51 (1·37–1·66) p<0·001 Adjusted for baseline depression score (n=2351) 1·33 (1·16–1·52) p<0·001 1·46 (1·32–1·61) p<0·001 Adjusted for ever self-harmed by 15 years (n=2351) 1·30 (1·14–1·49) p<0·001 1·33 (1·19–1·48) p<0·001 Additionally adjusted for all covariates* (n=2351) 1·28 (1·11–1·47) p<0·001 1·33 (1·19–1·48) p<0·001 Imputed confounders (n=3694) 1·20 (1·08–1·33) p<0·001 FMI 0·155 1·28 (1·18–1·39) p<0·001 FMI† 0·155 Imputed outcome (n=5342) 1·18 (1·06–1·31) p=0·002 FMI 0·472 1·25 (1·16–1·36) p<0·001 FMI 0·469 OR=odds ratio. FMI=fraction of missing information. sMFQ=Short Mood and Feelings Questionnaire. * Gender, previous depression (an sMFQ score of >11 at ages 10, 13, or 16 years), internalising and externalising difficulties at 11 years, self-perception, victimisation by bullies, antenatal depression, maternal history of depression, maternal education, and temperament (emotionality and activity).

Depression Self-harm OR (95% CI) p value OR (95% CI) p value Unadjusted (n=3694) 1·36 (1·23–1·49) p<0·001 1·52 (1·42–1·63) p<0·001 Complete cases (n=2351) 1·39 (1·22–1·58) p<0·001 1·51 (1·37–1·66) p<0·001 Adjusted for baseline depression score (n=2351) 1·33 (1·16–1·52) p<0·001 1·46 (1·32–1·61) p<0·001 Adjusted for ever self-harmed by 15 years (n=2351) 1·30 (1·14–1·49) p<0·001 1·33 (1·19–1·48) p<0·001 Additionally adjusted for all covariates* (n=2351) 1·28 (1·11–1·47) p<0·001 1·33 (1·19–1·48) p<0·001 Imputed confounders (n=3694) 1·20 (1·08–1·33) p<0·001 FMI 0·155 1·28 (1·18–1·39) p<0·001 FMI† 0·155 Imputed outcome (n=5342) 1·18 (1·06–1·31) p=0·002 FMI 0·472 1·25 (1·16–1·36) p<0·001 FMI 0·469 OR=odds ratio. FMI=fraction of missing information. sMFQ=Short Mood and Feelings Questionnaire. * Gender, previous depression (an sMFQ score of >11 at ages 10, 13, or 16 years), internalising and externalising difficulties at 11 years, self-perception, victimisation by bullies, antenatal depression, maternal history of depression, maternal education, and temperament (emotionality and activity). † The FMI can be used to judge whether the number of imputations is sufficient for analysis.32 A rule of thumb is that the number of imputations (M) should be more than or equal to 100 × FMI. In our case, the number of imputations (M=100) was sufficient.

Introduction Sleep problems are pervasive in people with schizophrenia. In a study1 of patients with persecutory delusions, 54% had clinical insomnia, 30% had subthreshold insomnia, and only 16% were sleeping well. In outpatients with clinically stable schizophrenia, poor sleep is common2, 3 and associated with an increased severity of positive symptoms.3 Relatives of patients with schizophrenia notice sleep problems more than any other sign preceding relapse,4 and a meta-analysis5 concluded that “sleep disorders are an intrinsic feature of schizophrenia”. These findings extend into the general population. Findings from two national epidemiological studies6, 7 have shown that insomnia is strongly associated with paranoia; additionally, sleep problems are associated with psychotic-like experiences in children.8 Yet, the treatment (or even routine assessment) of sleep problems in people with schizophrenia has received scant attention. Clinical trials in populations without a diagnosis of schizophrenia have shown that cognitive behavioural therapy (CBT) is highly effective for treatment of insomnia,9, 10, 11, 12 and CBT is considered by many as the recommended treatment for this disorder.13 However, this therapy is yet to be tested in a randomised controlled trial in patients with schizophrenia. Research in context Evidence before this study

Introduction Sleep problems are pervasive in people with schizophrenia. In a study1 of patients with persecutory delusions, 54% had clinical insomnia, 30% had subthreshold insomnia, and only 16% were sleeping well. In outpatients with clinically stable schizophrenia, poor sleep is common2, 3 and associated with an increased severity of positive symptoms.3 Relatives of patients with schizophrenia notice sleep problems more than any other sign preceding relapse,4 and a meta-analysis5 concluded that “sleep disorders are an intrinsic feature of schizophrenia”. These findings extend into the general population. Findings from two national epidemiological studies6, 7 have shown that insomnia is strongly associated with paranoia; additionally, sleep problems are associated with psychotic-like experiences in children.8 Yet, the treatment (or even routine assessment) of sleep problems in people with schizophrenia has received scant attention. Clinical trials in populations without a diagnosis of schizophrenia have shown that cognitive behavioural therapy (CBT) is highly effective for treatment of insomnia,9, 10, 11, 12 and CBT is considered by many as the recommended treatment for this disorder.13 However, this therapy is yet to be tested in a randomised controlled trial in patients with schizophrenia. Research in context Evidence before this study We searched PubMed and the ISRCTN trial registry up to April 10, 2015, with the terms “insomnia”, “therapy”, and “schizophrenia”, without date restrictions, for English-language publications of randomised controlled trials investigating the treatment of insomnia with patients with schizophrenia. We did not find any randomised trials of treatment of insomnia in schizophrenia with a psychological therapy. Apart from our own case report series, we did not find any descriptions of cognitive behavioural therapy (CBT) for insomnia with patients with schizophrenia. Three further searches with the terms “insomnia”, “randomized”, “schizophrenia”, and “hypnotics/benzodiazepine/melatonin”, and reading of review papers, identified only three small randomised controlled trials testing the short-term effects of melatonin and eszopiclone. We found no randomised controlled trials assessing treatment of insomnia in patients selected for having current delusions and hallucinations.

cs/benzodiazepine/melatonin”, and reading of review papers, identified only three small randomised controlled trials testing the short-term effects of melatonin and eszopiclone. We found no randomised controlled trials assessing treatment of insomnia in patients selected for having current delusions and hallucinations. Added value of this study Our trial is the first randomised controlled trial to treat insomnia in patients with current psychotic experiences, use a psychological treatment for insomnia with patients with schizophrenia, and examine effects of an insomnia treatment for patients with schizophrenia for up to 6 months. Our findings show that CBT for insomnia is likely to be beneficial for reducing insomnia in patients with schizophrenia. Implications of all the available evidence Trial data are insufficient for the treatment of sleep problems in patients with schizophrenia. The present study, in combination with studies of CBT for insomnia in other disorders, suggests that CBT could be offered as psychological treatment for patients with schizophrenia. However a larger, suitably powered phase 3 study is indicated.

onth. Exclusion criteria were a primary diagnosis of sleep apnoea, alcohol or substance dependency, organic syndrome or learning disability, a command of spoken English inadequate for engagement in therapy, and current engagement in individual CBT. Patient enrolment was done by one full-time graduate psychologist (RL). The trial received ethics approval from the NHS Research Ethics Committee South Central–Oxford C (reference 12/SC/0138) and the protocol has been published elsewhere.27 All patients provided written informed consent. Randomisation and masking We randomly assigned patients (1:1), via a web-based randomisation system, to receive either CBT plus standard care or standard care alone. The randomisation system was designed to balance three variables with a non-deterministic minimisation algorithm: sex (male vs female), severity of sleep problem (moderate [ISI score 15–21] vs high [22–28]), and psychotic experiences (hallucination only vs delusion only vs hallucinations and delusions).

ficient for the treatment of sleep problems in patients with schizophrenia. The present study, in combination with studies of CBT for insomnia in other disorders, suggests that CBT could be offered as psychological treatment for patients with schizophrenia. However a larger, suitably powered phase 3 study is indicated. Treatment of sleep problems in patients with psychosis might have another important benefit: reductions in delusions and hallucinations. Sleep disturbance is increasingly recognised as a potential contributory factor to the occurrence of a wide range of mental health problems.14, 15 Sleep disturbance might also have a role in the occurrence of psychotic experiences such as delusions and hallucinations. Findings from longitudinal studies have shown that insomnia predicts new inceptions of paranoia16 and its persistence.17 Insomnia increases negative affect, anomalous perceptions, and reasoning errors, which are all factors implicated in the development of persecutory ideation. Studies have also linked insomnia and hallucinatory experiences.18, 19 In adolescents at an ultra-high risk of psychosis, sleep difficulties are a predictor of positive psychotic experiences.20 Importantly, research in twins has shown overlap in the genetic and environmental causes of insomnia and psychotic experiences.18 If a causal link does exist, the clinical implication is that treatment of insomnia in patients with schizophrenia could lessen psychotic experiences, which would provide a new treatment route for these patients.

in twins has shown overlap in the genetic and environmental causes of insomnia and psychotic experiences.18 If a causal link does exist, the clinical implication is that treatment of insomnia in patients with schizophrenia could lessen psychotic experiences, which would provide a new treatment route for these patients. In a case series,21 we used CBT for insomnia in 15 patients with persistent persecutory delusions in the context of non-affective psychosis. After the brief CBT intervention, we recorded large reductions in levels of insomnia and paranoia. We also recorded significant reductions in levels of hallucinations, anxiety, and depression. A methodologically rigorous assessment is now needed. We planned the Better Sleep Trial (BEST) as a randomised controlled pilot trial assessing treatment of insomnia in patients with persistent delusions or hallucinations. We followed the recommendation of Lee and colleagues22 that “pilot studies are more about learning than confirming: they are not designed to formally assess evidence of benefit”. As such, we aimed to establish recruitment and follow-up rates, indicate levels of compliance with the treatment, examine the use of sleep assessments in this population, and provide an estimation of the potential efficacy of the intervention.

n confirming: they are not designed to formally assess evidence of benefit”. As such, we aimed to establish recruitment and follow-up rates, indicate levels of compliance with the treatment, examine the use of sleep assessments in this population, and provide an estimation of the potential efficacy of the intervention. The study is part of our process of improving treatments for patients with delusions and hallucinations, translating advances in understanding into intervention. One putative causal factor at a time is targeted and the effect on the psychotic experiences examined23—a so-called interventionist-causal model approach.24 Our intention was to use clinical techniques focused on sleep and then examine the subsequent effects. In this mechanistic approach, the central need is to establish change in the putative causal factor (ie, sleep) in one group compared with a group in which sleep patterns remain relatively stable. Therefore, the appropriate design was to compare the targeting of sleep with standard care. Identification of the active ingredients of intervention was not the question that the trial was designed to address.

causal factor (ie, sleep) in one group compared with a group in which sleep patterns remain relatively stable. Therefore, the appropriate design was to compare the targeting of sleep with standard care. Identification of the active ingredients of intervention was not the question that the trial was designed to address. We had two primary outcome hypotheses: that CBT for insomnia added to standard care would improve sleep in patients with psychosis compared with standard care alone, and that CBT for insomnia added to standard care would reduce delusions and hallucinations compared with standard care alone. Our secondary hypotheses were that improvements in sleep and psychotic symptoms would be maintained over at least 3 months, that improvement in sleep would be associated with improvements in psychotic symptoms, and that CBT for insomnia would lead to improvements in other outcomes, such as patient wellbeing and feelings of fatigue.

hypotheses were that improvements in sleep and psychotic symptoms would be maintained over at least 3 months, that improvement in sleep would be associated with improvements in psychotic symptoms, and that CBT for insomnia would lead to improvements in other outcomes, such as patient wellbeing and feelings of fatigue. Methods Study design and patients We did this prospective, assessor-blind, randomised controlled pilot study at two centres in the UK: Oxford Health National Health Service (NHS) Foundation Trust—a large mental health service covering a population of roughly 1·2 million people—and, in the final 4 months of recruitment, at an additional site at Northamptonshire Healthcare NHS Foundation Trust. We sought to include patients who had persistent, distressing delusions or hallucinations in the context of non-affective psychosis and insomnia. Inclusion criteria were current delusion or hallucination; a score of at least 2 on the distress items of the Psychotic Symptoms Rating Scale (PSYRATS)25 for either a delusion or hallucination; delusion or hallucination that had persisted for at least 3 months; a clinical diagnosis of schizophrenia, schizoaffective disorder, or delusional disorder (ie, a diagnosis of non-affective psychosis); sleep difficulties lasting 1 month or longer and a score of 15 or more on the Insomnia Severity Index26 (ISI; ie, clinical insomnia); an age of 18–65 years; and a medication dosage that had been stable for at least the past month. Exclusion criteria were a primary diagnosis of sleep apnoea, alcohol or substance dependency, organic syndrome or learning disability, a command of spoken English inadequate for engagement in therapy, and current engagement in individual CBT. Patient enrolment was done by one full-time graduate psychologist (RL).

ard care alone. The randomisation system was designed to balance three variables with a non-deterministic minimisation algorithm: sex (male vs female), severity of sleep problem (moderate [ISI score 15–21] vs high [22–28]), and psychotic experiences (hallucination only vs delusion only vs hallucinations and delusions). Research assessors (RL for most of the trial, with the last follow-up assessments completed by a replacement graduate psychologist [JM]) were masked to group allocation; the trial therapists informed patients of the randomisation outcome to maintain allocation concealment. Precautionary strategies to prevent unmasking included the therapist and assessor considering room use and booking arrangements; patients being reminded by the assessor not to talk about treatment allocation; and, after the initial assessment, the assessor not looking at the patient's clinical notes. In the case of an allocation being revealed, we remasked by using another assessor (which happened six times in total, three times at each follow-up point). All assessments were therefore done blind to allocation.

t allocation; and, after the initial assessment, the assessor not looking at the patient's clinical notes. In the case of an allocation being revealed, we remasked by using another assessor (which happened six times in total, three times at each follow-up point). All assessments were therefore done blind to allocation. Procedures The CBT intervention was provided one to one by clinical psychologists (EM and FW), either in NHS clinics or at the patient's home. DF and HS did weekly clinical supervision. The aim was to provide the insomnia intervention in about eight sessions over 12 weeks, with four sessions defined as a minimum therapeutic dose, with flexibility in length and number of sessions as appropriate in this clinical group. We also included telephone calls and texts between sessions to maintain treatment momentum. The main techniques standard for CBT sleep interventions were taken from four main sources.28, 29, 30, 31 The intervention was written in a manual to guide the work, which was shared with the patient. Initially the sessions focused on psycho-education about sleep difficulties, assessment of the triggering and maintenance of sleep difficulties, and goal setting. We used a checklist of factors likely to cause sleep difficulties, which was generated by the team. On the basis of the assessment, the active therapeutic techniques used could have included stimulus control therapy (eg, setting of appropriate and regular sleep times, ensuring the bed or bedroom were used only for sleeping, not staying in bed if unable to sleep for longer than 20–30 min, reducing sleep in the daytime), establishment of appropriate daytime activity and circadian rhythms (eg, obtaining natural light in the morning, regular mealtimes, gradually shifting sleep and wake times for sleep-phase problems), sleep hygiene, relaxation, and cognitive techniques addressing unhelpful beliefs and attitudes about sleep. The intervention was deliberately simplified, with the principal therapeutic techniques being stimulus control (ie, learning to associate bed with sleep) and improvement of daytime activity levels.

ase problems), sleep hygiene, relaxation, and cognitive techniques addressing unhelpful beliefs and attitudes about sleep. The intervention was deliberately simplified, with the principal therapeutic techniques being stimulus control (ie, learning to associate bed with sleep) and improvement of daytime activity levels. A detailed description of our approach to treatment of sleep problems in this group is available elsewhere,32 and includes noting of adaptations needed for the particular problems of delusions and hallucinations interfering with sleep, attempts to sleep being overused by patients as an escape from voices, extensive disruption of circadian rhythms, insufficient daytime activity, and fear of the bed based on past adverse experiences. Sessions were taped with patient agreement. To assess treatment fidelity, six tapes, chosen at random, were rated on the Cognitive Therapy Scale–Revised33 by an independent clinical psychologist experienced in CBT for psychosis. All tapes were rated as providing at least satisfactory cognitive therapy (ie, an average score of at least three on scale items). Standard care was delivered according to national and local service protocols and guidelines and mainly consisted of antipsychotic medication and contact with the local clinical team. We recorded medication and hospital admissions from clinical notes and other service provision using a modified version of the Client Service Receipt Inventory.34 In the baseline assessment, the trial patients were also assessed for insomnia using the Duke Structured Interview Schedule for Sleep Disorder Diagnoses.35

. We recorded medication and hospital admissions from clinical notes and other service provision using a modified version of the Client Service Receipt Inventory.34 In the baseline assessment, the trial patients were also assessed for insomnia using the Duke Structured Interview Schedule for Sleep Disorder Diagnoses.35 Outcomes As a pilot study, our main outcomes concerned the number of participants who were recruited, complied with treatment, and followed up. The prespecified primary outcome measures were levels of insomnia, assessed with the ISI,26 and levels of delusions and hallucinations, assessed with the PSYRATS.25 Higher scores on these measures indicate greater severity.

ncerned the number of participants who were recruited, complied with treatment, and followed up. The prespecified primary outcome measures were levels of insomnia, assessed with the ISI,26 and levels of delusions and hallucinations, assessed with the PSYRATS.25 Higher scores on these measures indicate greater severity. Secondary outcome measures assessed sleep using different methods: a second self-report questionnaire measure of sleep (Pittsburgh Sleep Quality Index [PSQI]),36 a sleep diary, and actigraphy data obtained by participants wearing an actiwatch (CamNtech MotionWatch 8, CamNtech, Cambridge, UK) for at least 7 days. Additionally, we assessed secondary outcome measures for psychiatric symptoms: a self-report measure of paranoid thinking (Paranoid Thoughts Scale);37 a standard psychiatric interviewer-rated assessment (the Positive and Negative Syndromes Scale; PANSS);38 and an adapted patient reported outcome measure (CHoice of Outcome In Cbt for psychosEs),39 which assessed, for example, self-confidence, peace of mind, and a sense of being in control. Other secondary outcomes were fatigue (Multidimensional Fatigue Inventory),40 quality of life (Euroqol 5 Dimensions 5 Levels; responses were converted into a single summary measure using UK population tariffs),41 and psychological wellbeing (Warwick-Edinburgh Mental Well-being Scale).42 We included several measures, such as the Beck Depression Inventory (BDI),43 for the purposes of potential mediation analysis. The outcome measures were completed at weeks 0 (baseline), 12 (post-intervention), and 24 (follow-up).

ion tariffs),41 and psychological wellbeing (Warwick-Edinburgh Mental Well-being Scale).42 We included several measures, such as the Beck Depression Inventory (BDI),43 for the purposes of potential mediation analysis. The outcome measures were completed at weeks 0 (baseline), 12 (post-intervention), and 24 (follow-up). During the trial, we recorded any adverse event that came to our attention. Medical notes were also checked at the end of the trial for the following events prespecified as adverse: all deaths, suicide attempts, serious violent incidents, admissions to secure units, formal complaints about therapy.

ion tariffs),41 and psychological wellbeing (Warwick-Edinburgh Mental Well-being Scale).42 We included several measures, such as the Beck Depression Inventory (BDI),43 for the purposes of potential mediation analysis. The outcome measures were completed at weeks 0 (baseline), 12 (post-intervention), and 24 (follow-up). During the trial, we recorded any adverse event that came to our attention. Medical notes were also checked at the end of the trial for the following events prespecified as adverse: all deaths, suicide attempts, serious violent incidents, admissions to secure units, formal complaints about therapy. Statistical analysis The trial statisticians (LC and L-MY) prepared a fully detailed statistical analysis plan and the chief investigator (DF) approved the plan before any analysis. No formal sample size calculation was done for this pilot study. However, the target sample size was based on recruitment for 15 months at an estimated rate of four patients per month with one research worker (ie, 60 patients), which was considered adequate to obtain reasonably reliable sample size estimates.44 Outcomes were assessed separately for assessment points at weeks 12 and 24. We used ANCOVA to obtain estimates and 95% CIs of continuous outcomes, with adjustment for baseline variables. The analysis plan did not include reporting of p values, in accordance with recommendations that “The analysis of a pilot study should be mainly descriptive or should focus on confidence interval estimation.”45 As a sensitivity analysis, we controlled for initial overall symptom severity as assessed by the PANSS and use of antipsychotic medication. In a post-hoc analysis for the main outcomes, we constructed a mixed model to incorporate the repeated measures at the two assessments (weeks 12 and 24) for each patient. We calculated effect sizes with Cohen's d by taking the estimated coefficient of treatment allocation from the ANCOVA divided by the pooled baseline SD. An effect size of 0·3 is considered a small effect, 0·5 a medium effect, and 0·8 or higher a large effect. We used correlation coefficients to examine potential associations between changes in sleep and changes in psychotic experiences (changes in scores for each measure were calculated as week 12 score minus baseline score). All main analyses were done at the end of the last follow-up assessments (ie, there were no interim analyses) and were based on the intention-to-treat population. Analyses were done with SAS (version 9.3)46 and were repeated by an independent statistician. This study is registered with ISRCTN, number 33695128.

score). All main analyses were done at the end of the last follow-up assessments (ie, there were no interim analyses) and were based on the intention-to-treat population. Analyses were done with SAS (version 9.3)46 and were repeated by an independent statistician. This study is registered with ISRCTN, number 33695128. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. DF, L-MY, and LC had full access to all the data in the study and had final responsibility for the decision to submit for publication.

score). All main analyses were done at the end of the last follow-up assessments (ie, there were no interim analyses) and were based on the intention-to-treat population. Analyses were done with SAS (version 9.3)46 and were repeated by an independent statistician. This study is registered with ISRCTN, number 33695128. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. DF, L-MY, and LC had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The figure shows the trial profile. Between Dec 14, 2012, and May 22, 2013, and Nov 7, 2013, and Aug 26, 2014, we randomly assigned 50 patients to receive CBT plus standard care (n=24) or standard care alone (n=26), with the break in enrolment due to employment of a new trial therapist. 47 (94%) patients were from Oxford Health and three (6%) were from Northamptonshire Healthcare. The last assessments were completed on Feb 10, 2015. All 50 patients completed the baseline assessment. 48 (96%) patients provided follow-up data for the primary efficacy measures (figure). In the CBT group, the mean number of sessions received was 7·3 (SD 1·9). On the basis of at least four CBT sessions constituting a minimum therapeutic dose, 23 (96%) of 24 patients had a dose of the intervention. The actual number of treatment sessions attended in the 12 week period was three (n=1), four (n=1), five (n=1), six (n=4), seven (n=5), eight (n=7), nine (n=2), ten (n=2), or 11 (n=1) sessions. The appendix shows descriptive comments about the intervention received from five of the first seven patients who had CBT.

on. The actual number of treatment sessions attended in the 12 week period was three (n=1), four (n=1), five (n=1), six (n=4), seven (n=5), eight (n=7), nine (n=2), ten (n=2), or 11 (n=1) sessions. The appendix shows descriptive comments about the intervention received from five of the first seven patients who had CBT. Baseline and demographic characteristics were similar between groups (table 1). In line with other studies of persistent psychotic experiences, both groups included slightly more men, the average age was about 40 years, most participants were unemployed, and the main diagnosis was schizophrenia (table 1). There were high levels of depression in both groups (table 1), although BDI scores were not significantly correlated with level of insomnia at baseline (r=0·07, p=0·623). All but one patient (in the standard care alone group) met the Duke Structured Interview Schedule for Sleep Disorder Diagnoses criteria for insomnia disorder. Provision of standard care was similar between groups and was fairly stable during the trial (table 2). All but four participants (n=3 given CBT, n=1 given standard care alone) were taking antipsychotic medication (table 2). Most participants were also prescribed a hypnotic, anxiolytic, or antidepressant medication (19 [79%] in the CBT group, 21 [81%] in the standard care alone group; appendix). All participants were outpatients.

our participants (n=3 given CBT, n=1 given standard care alone) were taking antipsychotic medication (table 2). Most participants were also prescribed a hypnotic, anxiolytic, or antidepressant medication (19 [79%] in the CBT group, 21 [81%] in the standard care alone group; appendix). All participants were outpatients. Compared with standard care alone, CBT had a treatment benefit on insomnia in the large effect size range at 12 weeks (table 3). Benefits were still apparent at 24 weeks' follow-up (table 2). At baseline, no patients' ISI scores were lower than the cutoff indicating normal sleep (a score of 0–7). By week 12, nine (41%) of 22 patients in the CBT group and one (4%) of 25 patients in the standard care alone group scored less than this cutoff. The wide confidence intervals for the effects of CBT on delusions and hallucinations cover a range from decreasing to increasing psychotic experiences (table 3). Changes in the other sleep assessments were relatively consistent with those for the primary outcome measure of insomnia (table 3). We recorded moderate to large effect sizes in sleep quality as assessed by the PSQI (table 3). Data collection was less complete for the sleep diaries and actigraphy than for the other secondary outcome measures, and the ensuing effect sizes were small to moderate (table 3). Compared with standard care alone, patients reported reduced fatigue at week 12, and improved quality of life and psychological wellbeing at week 24, with small to medium effect sizes overall for both these categories (table 3). The confidence intervals for paranoia and overall psychiatric symptomatology again span CBT potentially reducing or increasing these problems (table 3). Correlations between changes in insomnia and changes in hallucinations and paranoia (although not PSYRATS delusions) were mainly small and positive, but the confidence intervals are wide and include negative correlations (appendix).

ptomatology again span CBT potentially reducing or increasing these problems (table 3). Correlations between changes in insomnia and changes in hallucinations and paranoia (although not PSYRATS delusions) were mainly small and positive, but the confidence intervals are wide and include negative correlations (appendix). All the notes of the trial patients were checked. Four patients were admitted to hospital during the trial (n=2 in each group; appendix). There were no deaths or complaints about therapy. Three patients, all in the CBT group, had a total of five adverse events: two suicide attempts, two serious violent incidents, and one admission to a secure unit (following one of the violent incidents). No adverse events were considered to be related to study treatment. The appendix shows results of the sensitivity analysis. Results of the post-hoc analysis using a mixed model for repeated measures analysis were similar to those obtained using ANCOVA (appendix).

All the notes of the trial patients were checked. Four patients were admitted to hospital during the trial (n=2 in each group; appendix). There were no deaths or complaints about therapy. Three patients, all in the CBT group, had a total of five adverse events: two suicide attempts, two serious violent incidents, and one admission to a secure unit (following one of the violent incidents). No adverse events were considered to be related to study treatment. The appendix shows results of the sensitivity analysis. Results of the post-hoc analysis using a mixed model for repeated measures analysis were similar to those obtained using ANCOVA (appendix). Discussion This is the first randomised controlled trial testing the effects of CBT for insomnia in patients with current psychotic experiences, which comprised very long-standing insomnia, delusions, and hallucinations. Overall our study shows the feasibility of testing the psychological treatment adapted for this population: it was possible to recruit patients to the trial, to randomise them, to keep assessors masked to allocation, to retain patients in the trial, to assess a wide range of measures, and to implement the treatment. Indeed, there was a very low dropout rate for the trial assessments and a very high uptake of CBT. Our findings also show that CBT for insomnia, a relatively brief intervention, is likely to have substantial benefits in improving sleep for patients with current delusions and hallucinations. Benefits in sleep were sustained up to the final follow-up assessment.

for the trial assessments and a very high uptake of CBT. Our findings also show that CBT for insomnia, a relatively brief intervention, is likely to have substantial benefits in improving sleep for patients with current delusions and hallucinations. Benefits in sleep were sustained up to the final follow-up assessment. Findings from a qualitative study47 of a subsample of the trial participants are consistent with those of the present quantitative analysis. Our results compare well with the only randomised controlled trial that we are aware of that targeted insomnia in patients with schizophrenia with one of the commonly prescribed sedative hypnotics (so-called Z-drugs). In a randomised double-blind trial48 done over 8 weeks in 39 clinically stable outpatients with schizophrenia who had insomnia but were not recruited for current psychotic experiences, eszoplicone was compared with placebo. The medication led to a decrease in ISI scores of 3·78 (95% CI 0·2–7·5) versus placebo. The reduction in ISI score in the present study is very similar to that reported in a meta-analysis of CBT for comorbid insomnia for patients with physical or psychiatric disorders (ISI mean change 6·4 [SE 1·27]).11 Our trial did not prove informative about the potential effect of the sleep treatment on the psychotic experiences. The wide confidence intervals include the possibility that the treatment can either reduce or increase delusions and hallucinations. A larger trial will provide a better estimate of the effects on psychotic experiences.

Contributors DF, L-MY, HS, and EM designed the study. DF was the principal investigator and had the main responsibility for drafting the study protocol and report. FW and EM were the trial therapists. DF, HS, and AGH provided the training and supervision for the trial therapists. RL and JM were the research assessors. L-MY and LC were responsible for the trial outcome analyses. DF, L-MY, and LC take responsibility for the integrity of the data and the accuracy of the data analysis. All authors contributed to, read, and approved the final report. Declaration of interests We declare no competing interests. Figure Trial profile ISI=Insomnia Severity Index. PSYRATS=Psychotic Symptoms Rating Scale. CBT=cognitive behavioural therapy. Table 1 Baseline demographic and clinical characteristics CBT group (n=24) Standard care alone group (n=26) Age (years) 39·6 (11·6) 42·2 (13·5) Sex Male 16 (67%) 18 (69%) Female 8 (33%) 8 (31%) Ethnic origin White 22 (92%) 25 (96%) Black 1 (4%) 1 (4%) Chinese 1 (4%) 0 Employment status Unemployed 21 (88%) 23 (88%) Part-time employed 1 (4%) 0 Full-time employed 0 1 (4%) Volunteer 1 (4%) 1 (4%) Retired 1 (4%) 0 Student 0 1 (4%) Clinical diagnosis Schizophrenia 16 (67%) 17 (65%) Schizoaffective disorder 5 (21%) 5 (19%) Delusional disorder 0 0 Psychosis not otherwise specified 3 (13%) 4 (15%) Severity of insomnia (ISI) Moderate (15–21) 19 (79%) 19 (73%) High (22–28) 5 (21%) 7 (27%) Presence of psychotic experiences Both delusions and hallucinations 18 (75%) 19 (73%) Delusion only 4 (17%) 3 (12%) Hallucination only 2 (8%) 4 (15%) Depression (BDI) None (0–13) 3 (13%) 3 (12%) Mild (14–19) 5 (21%) 3 (12%) Moderate (20–28) 2 (8%) 6 (23%) Severe (29–63) 14 (58%) 14 (54%) Data are mean (SD) or n (%), unless otherwise indicated. CBT=cognitive behavioural therapy. ISI=Insomnia Severity Index. BDI=Beck Depression Inventory.

ot prove informative about the potential effect of the sleep treatment on the psychotic experiences. The wide confidence intervals include the possibility that the treatment can either reduce or increase delusions and hallucinations. A larger trial will provide a better estimate of the effects on psychotic experiences. The clear caution when discussing potential clinical effects is that this study was a pilot study. Our trial is insufficiently powered to detect anything but the largest effect sizes. There was no formal power calculation because we did not primarily intend to detect treatment effects. We did not achieve the target sample size of 60 patients, but this information will form part of the recruitment strategy plan for the future phase 3 trial. Our main purpose was to assess the feasibility of the assessment and to refine procedures. A key observation was the wide variety of sleep disturbance within the patient group, including patients going to bed early in the evening and simply lying awake for hours, patients not having a bed, patients having an association with their bed as a place of trauma, patients awake and pacing all night, patients' sleep being disturbed through the use of hypnotics, patients being drowsy and oversleeping at various times of the day, and patients only going to bed in the early hours (delayed-phase sleep), whereas others had obviously irregular sleep–wake patterns. Our treatment manual will become more specific to each of these types of presentation and it would be of interest to establish whether they are associated with different outcomes. Such presentations are consistent with evidence of varied circadian rhythm dysfunction in patients with schizophrenia49 and, indeed, the complexity of sleep disturbance beyond insomnia in other disorders such as bipolar disorder.50 Notably, our trial did not include polysomnography. Obvious practical difficulties exist in implementation of such a measurement, and subjective reports of problems must be the key clinical priority. Instead of polysomnography we used actigraphy to assess sleep–wake timing and fragmentation. This form of assessment (wearing a watch-like device) was not agreed to by all participants. Furthermore, we have concerns that the movement recordings are not able to differentiate between actual sleep and drowsy but awake inactivity, which is common in this patient group.

o assess sleep–wake timing and fragmentation. This form of assessment (wearing a watch-like device) was not agreed to by all participants. Furthermore, we have concerns that the movement recordings are not able to differentiate between actual sleep and drowsy but awake inactivity, which is common in this patient group. Such recordings might be better used to assess changes in overall activity levels, which is an important outcome in itself. Sleep improvement might also lead to physical health benefits.51

o assess sleep–wake timing and fragmentation. This form of assessment (wearing a watch-like device) was not agreed to by all participants. Furthermore, we have concerns that the movement recordings are not able to differentiate between actual sleep and drowsy but awake inactivity, which is common in this patient group. Such recordings might be better used to assess changes in overall activity levels, which is an important outcome in itself. Sleep improvement might also lead to physical health benefits.51 The key efficacy question addressed in the design of the present study was whether CBT can improve sleep in patients with current psychotic experiences and insomnia. We assessed this question in the context of our translational research that aims to improve treatments for people with psychosis.23 The causal framework that we use leads to the initial aim of showing change in the targeted mechanism (eg, sleep disruption), hence the comparison of a sleep treatment (whereby the expectation is that sleep will improve) with continued standard care (whereby sleep patterns are likely to remain stable). The chosen design also addresses the important clinical question of the potential overall benefit provided when the intervention is added to standard care. From some perspectives, this question is key: “The primary question facing clinicians and policy makers is whether adding a particular form of treatment will significantly improve symptoms or functioning compared to the often-constrained services as usual.”52 However, this question would only establish whether the extra treatment has benefits, not which elements of the package are actually necessary or whether there are better forms of treatment. These latter questions require different choices of control group in order to answer them. Nonetheless, our clinical impression is that, for example, the extra contact time alone provided by the CBT is an insufficient explanation for such large improvements in sleep in a patient group that typically has long-standing and complex problems. Indeed, CBT has already been shown to have benefit compared with attention-control conditions in the treatment of primary insomnia.53

extra contact time alone provided by the CBT is an insufficient explanation for such large improvements in sleep in a patient group that typically has long-standing and complex problems. Indeed, CBT has already been shown to have benefit compared with attention-control conditions in the treatment of primary insomnia.53 Sleep difficulties are a problem that patients with psychosis find distressing and very much wish to receive effective treatment for from mental health services. A large, multicentre trial is now needed that can establish the effects of CBT for insomnia, delivered by front-line mental health staff, both on sleep and psychotic experiences. Supplementary Material Supplementary appendix Acknowledgments The trial was funded by a grant from the National Health Service (NHS) National Institute for Health Research (NIHR) Research for Patient Benefit Programme (reference PB-PG-0211-10007). DF is supported by a Medical Research Council senior clinical fellowship. Research support to RGF, DF, and RL is provided by a Wellcome Trust Strategic Award (098461/Z/12/Z) for the Oxford Sleep and Circadian Neurosciences Institute. JG is an NHS NIHR senior investigator. We thank Annabel Ivins for facilitating recruitment in Northamptonshire. The views and opinions expressed herein are those of the authors and do not necessarily reflect those of the Research for Patient Benefit Programme, NIHR, NHS, or the Department of Health.

sciences Institute. JG is an NHS NIHR senior investigator. We thank Annabel Ivins for facilitating recruitment in Northamptonshire. The views and opinions expressed herein are those of the authors and do not necessarily reflect those of the Research for Patient Benefit Programme, NIHR, NHS, or the Department of Health. Contributors DF, L-MY, HS, and EM designed the study. DF was the principal investigator and had the main responsibility for drafting the study protocol and report. FW and EM were the trial therapists. DF, HS, and AGH provided the training and supervision for the trial therapists. RL and JM were the research assessors. L-MY and LC were responsible for the trial outcome analyses. DF, L-MY, and LC take responsibility for the integrity of the data and the accuracy of the data analysis. All authors contributed to, read, and approved the final report. Declaration of interests We declare no competing interests. Figure Trial profile ISI=Insomnia Severity Index. PSYRATS=Psychotic Symptoms Rating Scale. CBT=cognitive behavioural therapy. Table 1 Baseline demographic and clinical characteristics

ucination only 2 (8%) 4 (15%) Depression (BDI) None (0–13) 3 (13%) 3 (12%) Mild (14–19) 5 (21%) 3 (12%) Moderate (20–28) 2 (8%) 6 (23%) Severe (29–63) 14 (58%) 14 (54%) Data are mean (SD) or n (%), unless otherwise indicated. CBT=cognitive behavioural therapy. ISI=Insomnia Severity Index. BDI=Beck Depression Inventory. Table 2 Provision of standard care CBT group Standard care alone group n Mean (SD) n Mean (SD) Antipsychotic medication (chlorpromazine equivalent dose, mg/day) Baseline 24 363·7 (266·5) 26 495·8 (358·1) 12 weeks 23 359·8 (289·3) 25 530·7 (366·0) 24 weeks 23 372·5 (302·9) 25 531·3 (373·2) 6 months before the trial (n) Psychiatric hospital admission 24 0·1 (0·3) 26 0·1 (0·3) Psychiatrist meetings 23 1·2 (1·4) 26 2·5 (4·9) Meetings with community psychiatric nurse 23 11·7 (13·6) 26 15·4 (16·2) Visits to day-care centre 23 6·0 (14·8) 26 2·0 (6·8) Meetings with general practitioner 22 2·8 (3·6) 25 3·0 (5·3) 6 months during the trial (n) Psychiatric hospital admission 24 0·1 (0·3) 26 0·1 (0·3) Psychiatrist meetings 19 1·1 (0·9) 24 2·3 (3·9) Meetings with community psychiatric nurse 19 7·6 (8·2) 24 13·5 (13·9) Visits to day-care centre 19 6·1 (18·6) 24 7·9 (24·9) Meetings with general practitioner 18 3·5 (3·9) 24 3·2 (3·2) Table 3 Scores for the primary and secondary outcome measures

24 0·1 (0·3) 26 0·1 (0·3) Psychiatrist meetings 19 1·1 (0·9) 24 2·3 (3·9) Meetings with community psychiatric nurse 19 7·6 (8·2) 24 13·5 (13·9) Visits to day-care centre 19 6·1 (18·6) 24 7·9 (24·9) Meetings with general practitioner 18 3·5 (3·9) 24 3·2 (3·2) Table 3 Scores for the primary and secondary outcome measures CBT group Standard care alone group Adjusted mean difference (95% CI) Effect size (d) n Mean (SD) n Mean (SD) Primary outcome measure Insomnia (ISI) 0 weeks 24 18·6 (3·2) 26 18·8 (3·3) .. .. 12 weeks 22 9·3 (5·5) 25 15·4 (5·4) 6·1 (3·0 to 9·22) 1·9 24 weeks 23 11·0 (5·6) 25 15·0 (5·7) 3·9 (0·9 to 6·8) 1·2 Delusions (PSYRATS) 0 weeks 24 16·1 (3·2) 26 15·3 (4·9) .. .. 12 weeks 22 13·9 (4·8) 25 13·8 (4·1) 0·3 (−2·0 to 2·6) 0·1 24 weeks 23 14·0 (4·7) 25 12·7 (5·7) −0·8 (−3·6 to 2·1) −0·2 Hallucinations (PSYRATS) 0 weeks 24 25·1 (12·1) 26 26·7 (9·2) .. .. 12 weeks 22 27·5 (9·2) 25 25·9 (8·1) −1·9 (−6·5 to 2·7) −0·2 24 weeks 23 24·6 (11·6) 25 22·0 (10·2) −3·4 (−9·2 to 2·3) −0·3 Sleep secondary outcome measures Sleep quality (PSQI) 0 weeks 23 11·9 (3·2) 23 11·6 (2·9) .. .. 12 weeks 21 8·7 (3·8) 23 10·5 (4·8) 1·9 (−0·4 to 4·1) 0·6 24 weeks 23 6·9 (4·4) 23 9·6 (4·8) 2·8 (0·2 to 5·4) 0·9 Time to sleep onset (min)* 0 weeks 21 61·5 (51·4) 25 61·1 (32·7) .. .. 12 weeks 22 34·6 (40·3) 19 46·6 (37·0) 18·3 (−1·5 to 38·2) 0·4 24 weeks 20 33·3 (29·6) 21 57·9 (41·7) 26·4 (2·5 to 50·3) 0·6 Total sleep time (min)* 0 weeks 18 393·7 (127·8) 21 403·4 (146·3) .. .. 12 weeks 17 456·4 (107·7) 18 437·5 (118·1) 33·0 (−17·9 to 83·8) 0·2 24 weeks 18 465·9 (117·1) 20 412·4 (128·2) 46·6 (−15·1 to 108·3) 0·3 Waking in night (min)* 0 weeks 19 43·5 (40·2) 22 59·1 (63·3) .. .. 12 weeks 18 19·9 (22·2) 18 47·6 (52·1) 16·9 (−4·7 to 38·5) 0·3 24 weeks 18 32·7 (35·7) 20 42·7 (46·6) 14·6 (−11·1 to 40·3) 0·3 Total sleep time (min)† 0 weeks 20 396·3 (141·7) 23 470·0 (128·7) .. .. 12 weeks 18 406·2 (100·0) 19 471·0 (124·4) −15·5 (−78·9 to 47·9) −0·1 24 weeks 18 445·1 (86·7) 21 449·8 (141·1) 33·1 (−27·0 to 93·3) 0·2 Psychiatric secondary measures Paranoia (GPTS) 0 weeks 24 90·8 (28·7) 26 90·5 (29·8) .. .. 12 weeks 22 89·6 (36·8) 24 96·2 (37·3) 6·7 (−10·2 to 23·5) 0·2 24 weeks 20 78·3 (34·8) 25 88·1 (35·0) 9·8 (−7·5 to 27·2) 0·3 Total symptoms (PANSS) 0 weeks 24 83·6 (16·2) 26 79·7 (14·1) .. .. 12 weeks 22 77·5 (12·1) 24 79·3 (14·6) 3·4 (−2·3 to 9·0) 0·2 24 weeks 21 74·8 (14·7) 24 75·8 (11·8) 3·3 (−2·9 to 9·5) 0·2 Fatigue (MFI) 0 weeks 23 43·8 (16·4) 26 47·6 (15·3) .. ..

o 23·5) 0·2 24 weeks 20 78·3 (34·8) 25 88·1 (35·0) 9·8 (−7·5 to 27·2) 0·3 Total symptoms (PANSS) 0 weeks 24 83·6 (16·2) 26 79·7 (14·1) .. .. 12 weeks 22 77·5 (12·1) 24 79·3 (14·6) 3·4 (−2·3 to 9·0) 0·2 24 weeks 21 74·8 (14·7) 24 75·8 (11·8) 3·3 (−2·9 to 9·5) 0·2 Fatigue (MFI) 0 weeks 23 43·8 (16·4) 26 47·6 (15·3) .. .. 12 weeks 22 29·1 (19·0) 24 45·4 (19·6) 10·5 (2·1 to 18·9) 0·7 24 weeks 21 25·9 (21·4) 25 38·4 (18·1) 9·0 (−1·1 to 19·1) 0·6 Patient outcomes (CHOICE) 0 weeks 23 52·2 (21·1) 26 55·0 (14·4) .. .. 12 weeks 22 58·0 (22·7) 24 49·9 (18·3) 8·5 (−1·4 to 18·5) 0·5 24 weeks 21 60·0 (22·8) 23 57·5 (21·8) 3·8 (−7·3 to 14·9) 0·2 Quality of life (EQ-5D-5L) 0 weeks 24 0·55 (0·23) 26 0·60 (0·22) .. .. 12 weeks 22 0·63 (0·25) 24 0·55 (0·22) 0·11 (0·02 to 0·23) 0·5 24 weeks 22 0·63 (0·27) 25 0·58 (0·20) 0·08 (−0·05 to 0·21) 0·4 Wellbeing (WEMWBS) 0 weeks 24 35·3 (9·3) 26 37·0 (7·8) .. .. 12 weeks 22 36·1 (10·7) 24 34·0 (8·9) 2·6 (−2·1 to 7·4) 0·3 24 weeks 21 39·4 (9·9) 25 34·7 (7·9) 4·8 (−0·5 to 9·3) 0·6 CBT=cognitive behavioural therapy. ISI=Insomnia Severity Index. PSYRATS=Psychotic Symptoms Rating Scale. PSQI=Pittsburgh Sleep Quality Index. GPTS=Green et al Paranoid Thought Scales. PANSS=Positive and Negative Syndromes Scale. MFI=Multidimensional Fatigue Inventory. CHOICE=CHoice of Outcome In Cbt for psychoses. EQ–5D=Euroqol 5 Dimensions 5 Levels. WEMWBS=Warwick-Edinburgh Mental Well-being Scale. * Sleep diary. † Actigraphy data.

Introduction More than 10 million people are currently in prison worldwide,1 and substantially larger numbers of ex-prisoners are living in society.2 Despite reported decreases in violence in many countries,3 repeat offending remains high across many high-income and middle-income countries.4 In the USA and UK, more than a third of released prisoners are reconvicted for a new crime within 2 years, and more than half within 5 years.5,6 Furthermore, about 70% of those convicted in the USA are repeat offenders.7 In England and Wales, this figure is estimated at 90%,8 and the proportion of individuals convicted who have had 15 or more previous offences has been increasing since 2008.9

for a new crime within 2 years, and more than half within 5 years.5,6 Furthermore, about 70% of those convicted in the USA are repeat offenders.7 In England and Wales, this figure is estimated at 90%,8 and the proportion of individuals convicted who have had 15 or more previous offences has been increasing since 2008.9 Much research has focused on identification of individuals at high risk of reoffending. Although a substantial amount is known about demographic risk factors for reoffending,10–12 uncertainty remains about its mental health determinants.13 Research specifically related to reoffending is different from that in the general population because in the general population, several psychiatric disorders have been shown to be associated with an increased risk of committing and conviction for violence and violent crime,14–16 whereas in offenders, this association is not consistent. Because psychiatric disorders are prevalent and mostly treatable, with some studies suggesting that one in seven prisoners has a psychotic illness or major depression, and about one in five people enter prison with clinically significant substance use disorders,17 tackling them has the potential to substantially reduce adverse outcomes in released prisoners. In the USA, for example, estimates suggest that 15% of prisoners have a severe mental illness,18 and the number of individuals with mental illness in prisons and jails is ten times that in public psychiatric hospitals.19

,17 tackling them has the potential to substantially reduce adverse outcomes in released prisoners. In the USA, for example, estimates suggest that 15% of prisoners have a severe mental illness,18 and the number of individuals with mental illness in prisons and jails is ten times that in public psychiatric hospitals.19 The little research into psychiatric disorders and reoffending that has been done has led to divergent findings. Authors of systematic reviews with heterogeneous samples10,20 have concluded that psychosis is inversely related to reoffending. By contrast, authors of a focused review13 reported that psychosis increased risk of reoffending, although it was only based on four studies that used control groups without psychiatric disorder. However, even in these investigations, causality has not been shown, and several potential confounders have not been fully examined.13 First, whether this association is attributable to sociodemographic and criminological factors is uncertain.21 Second, findings from some studies suggest that the association is mainly due to substance misuse,22–24 and whether other common psych iatric disorders are independently related to risk of reoffending needs further examination. Third, although both criminal activity25 and most psychiatric disorders26 have long been known to run in families, the contribution of familial (genetic and early environmental) factors to the association has not been investigated. Finally, few studies have been done on female prisoners, who have higher prevalences of psychiatric disorders than do men in prison.17

ost psychiatric disorders26 have long been known to run in families, the contribution of familial (genetic and early environmental) factors to the association has not been investigated. Finally, few studies have been done on female prisoners, who have higher prevalences of psychiatric disorders than do men in prison.17 In this population-based longitudinal study of released prisoners, we aimed to investigate the association between psychiatric disorders and violent reoffending and to address three questions. First, whether being diagnosed with any psychiatric disorder is independently associated with violent reoffending. Second, whether this association differs by psychiatric diagnosis. Finally, whether this association is explained or moderated by comorbid substance use disorder. We did the analyses by controlling for sociodemographic and criminological factors, but also comparing sibling prisoners with and without psychiatric disorder, a powerful approach to control for familial confounding.

gnosis. Finally, whether this association is explained or moderated by comorbid substance use disorder. We did the analyses by controlling for sociodemographic and criminological factors, but also comparing sibling prisoners with and without psychiatric disorder, a powerful approach to control for familial confounding. Methods Study setting We linked the following population-based registers in Sweden: the National Crime Register, which includes detailed information about all criminal convictions since 1973; the National Patient Register, which provides diagnoses for all inpatient psychiatric hospital admissions since 1973 and outpatient care since 2001; the Migration Register, which supplies information about dates of migration into or out of Sweden; the Cause of Death Register, which contains information about dates and causes of all deaths since 1958; the Multi-Generation Register, which contains information about biological relationships for all individuals living in Sweden since 1933; and the Longitudinal Integration Database for Health Insurance and Labour Market studies, which contains yearly assessments of income, marital and employment status, and education for all individuals aged 16 years or older since 1990.

ation about biological relationships for all individuals living in Sweden since 1933; and the Longitudinal Integration Database for Health Insurance and Labour Market studies, which contains yearly assessments of income, marital and employment status, and education for all individuals aged 16 years or older since 1990. In Sweden, all residents (including immigrants) have a unique personal identifier used in all national registers, thus enabling data linkage.27 We selected a cohort of all convicted prisoners who have been imprisoned since Jan 1, 2000, and released before Dec 31, 2009. All individuals were followed up from the day of release until first reoffence of violent crime, death, emigration, or end of the study (Dec 31, 2009). We identified prisoners with full siblings using the Multi-Generation Register. This study was approved by the Regional Ethics Committee at the Karolinska Institutet (Stockholm, Sweden).

re followed up from the day of release until first reoffence of violent crime, death, emigration, or end of the study (Dec 31, 2009). We identified prisoners with full siblings using the Multi-Generation Register. This study was approved by the Regional Ethics Committee at the Karolinska Institutet (Stockholm, Sweden). Measures We linked prisoners within the study cohort to the National Patient Register to obtain information about diagnosed psychiatric disorders. We identified those with any lifetime psychiatric diagnoses (based on the ICD Eighth [ICD-8; code 290–315], Ninth [ICD-9; code 290–319], and Tenth [ICD-10; code F00–F99] Revisions) before release from prison. To explore the difference between individual disorders and the effect of comorbidity, we investigated the following specific psychiatric disorders: alcohol use disorder (ICD-8: 291 and 303; ICD-9: 291, 303, and 305A; ICD-10: F10), drug use disorder (ICD-8: 304; ICD-9: 292, 304, and 305 [except .A]; ICD-10: F11–F19), personality disorder (ICD-8: 301 [except .1]; ICD-9: 301 [except .B]; ICD-10: F60–F61), attention-deficit hyperactivity disorder (ICD-8: not applicable; ICD-9: 314; ICD-10: F90), and other developmental or childhood disorders (ICD-8: 308; ICD-9: 299A, 312, 313, and 315; ICD-10: F80–F98 [except F90]).

ICD-10: F11–F19), personality disorder (ICD-8: 301 [except .1]; ICD-9: 301 [except .B]; ICD-10: F60–F61), attention-deficit hyperactivity disorder (ICD-8: not applicable; ICD-9: 314; ICD-10: F90), and other developmental or childhood disorders (ICD-8: 308; ICD-9: 299A, 312, 313, and 315; ICD-10: F80–F98 [except F90]). We assigned a hierarchical approach to differentiate between schizophrenia spectrum disorders, bipolar disorder, depression, and anxiety disorder.28 We included any individual with one of the schizophrenia spectrum disorder diagnoses, including schizoaffective and delusional disorders (ICD-8: 295, 297, 298·1–9, and 299; ICD-9: 295, 297, 298 [except .A], and 299; ICD-10: F20–F29); one of the bipolar diagnoses (ICD-8: 296.1, 296.3, 296.8, 296A, 296C–296E, and 296W; ICD-10: F30–F31), but not schizophrenia spectrum disorders; one of the depression diagnoses (ICD-8: 296.2, 296.9, 298.0, and 300.4; ICD-9: 296B, 296X, 298A, 300E, and 311; ICD-10: F32–F39), but without schizophrenia spectrum or bipolar disorder; and one of the anxiety diagnoses (ICD-8: 300 [except .4], 305, and 307; ICD-9: 300 [except .E], 306, 308, and 309; ICD-10: F40–F48), but without schizophrenia spectrum or bipolar disorder or depression. Use of Swedish national registers for psychiatric research is well established, and the patient registry data have good to excellent validity for a range of psychiatric disorders.29–33 Overall, the positive predictive value has been reported to be 85–95% for most diagnoses.34

spectrum or bipolar disorder or depression. Use of Swedish national registers for psychiatric research is well established, and the patient registry data have good to excellent validity for a range of psychiatric disorders.29–33 Overall, the positive predictive value has been reported to be 85–95% for most diagnoses.34 The main outcome was any conviction of violent crime after release. In keeping with previous work, we defined violent crime as homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), illegal threats, or intimidation.33,35 If no date of the crime was recorded, we used the date of conviction. Measured covariates were sex, age, immigration status (defined as being born outside Sweden), criminological factors (length of incarceration [categorised into four levels], violent index offence, and any previous violent crime), and sociodemographic factors (civil status [categorised into four levels], employment, highest level of completed education [categorised into three levels], disposable income, and neighbourhood deprivation) at the year of release. For all analyses, we investigated the index offence, which is the most serious offence that led to the prison sentence. We did not replace missing data by imputation or other methods because this imputation needs some assumptions to be made and the number of individuals with missing values was quite small, but in a sensitivity analysis, we recalculated the results with missing values imputed.

ous offence that led to the prison sentence. We did not replace missing data by imputation or other methods because this imputation needs some assumptions to be made and the number of individuals with missing values was quite small, but in a sensitivity analysis, we recalculated the results with missing values imputed. Statistical analysis To explore the association between psychiatric disorders and risk of violent reoffending, we compared prisoners with and without a psychiatric disorder. We used Kaplan-Meier survival curves to show the timing of violent reoffending after release from prison. To quantify the association, we used the Cox proportional hazards model, and estimated hazard ratios (HRs) in three models. In the first model, we adjusted for age and immigration status. In the second, we also adjusted for socio demographic and criminological factors. In the third, we also used sibling comparison to adjust for possible familial confounding.36 We did this familial adjustment by fitting a fixed-effect model37 (stratified Cox regression) to the subsample of same-sex full sibling prisoners. This model adjusts for all unmeasured genetic and environ mental factors that are shared by siblings, and also included the measured covariates adjusted for in models 1 and 2. We stratified all analyses by sex. We verified the proportional hazards assumption by visually checking the Kaplan-Meier curves and tested it using Schoenfeld residuals.38

unmeasured genetic and environ mental factors that are shared by siblings, and also included the measured covariates adjusted for in models 1 and 2. We stratified all analyses by sex. We verified the proportional hazards assumption by visually checking the Kaplan-Meier curves and tested it using Schoenfeld residuals.38 To explore the association between each individual psychiatric disorder and risk of violent reoffending, we constructed Cox regression models for each of the diagnoses investigated. We calculated HRs in three models, with progressive adjustment for age and immigration status, sociodemographic and criminological covariates, and alcohol and drug use disorders. We further examined whether the association between psychiatric disorder and violent reoffending was moderated by substance use disorder (defined as diagnoses of alcohol or drug use disorders). We used a likelihood ratio test to examine the interaction between psychiatric disorder and substance use disorder (with p<0·05 indicating a significant interaction). Additionally, we analysed the moderating effect of substance use disorder on schizophrenia spectrum disorders and bipolar disorder. These diagnoses had the best diagnostic validity in our sample. Because comorbidity is common, we further examined the association between the number of diagnosed psychiatric disorders and violent reoffending.

ysed the moderating effect of substance use disorder on schizophrenia spectrum disorders and bipolar disorder. These diagnoses had the best diagnostic validity in our sample. Because comorbidity is common, we further examined the association between the number of diagnosed psychiatric disorders and violent reoffending. To assess the population effect of psychiatric disorder on violent reoffending, we used the population attributable fraction (PAF). The PAF measures the proportion of violent reoffending in the population that can be attributed to psychiatric disorder, assuming that a causal relation exists. In the presence of confounding, PAF can be calculated as Pr(X=1∣Y=1)(1 – HRα−1),39 where Pr(X=1∣Y=1) is the probability of exposure given outcome and HRα−1 is the adjusted HR. We calculated confidence intervals for PAFs using the Bonferroni inequality method.40

chiatric disorder, assuming that a causal relation exists. In the presence of confounding, PAF can be calculated as Pr(X=1∣Y=1)(1 – HRα−1),39 where Pr(X=1∣Y=1) is the probability of exposure given outcome and HRα−1 is the adjusted HR. We calculated confidence intervals for PAFs using the Bonferroni inequality method.40 To test whether the association between psychiatric disorders and violent reoffending was different depending on type of crime, we did sensitivity analyses using different outcomes. First, we restricted the outcome to specific crimes for which interpersonal violence is known to have occurred, including homicide and attempted homicide, all forms of assault (including aggravated, and assault of an officer), rape, sexual coercion, and child molestation. Second, we examined the association with other violent crime: arson, indecent exposure, sexual harassment, illegal threats, and intimidation. This breakdown also provides a proxy for testing of associations by severity of violent crime. We also examined the association between individual psychiatric disorders and these two subgroups of violent reoffending in male prisoners. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ZC had full access to all the data in the study and, with SF, had final responsibility for the decision to submit for publication.

To test whether the association between psychiatric disorders and violent reoffending was different depending on type of crime, we did sensitivity analyses using different outcomes. First, we restricted the outcome to specific crimes for which interpersonal violence is known to have occurred, including homicide and attempted homicide, all forms of assault (including aggravated, and assault of an officer), rape, sexual coercion, and child molestation. Second, we examined the association with other violent crime: arson, indecent exposure, sexual harassment, illegal threats, and intimidation. This breakdown also provides a proxy for testing of associations by severity of violent crime. We also examined the association between individual psychiatric disorders and these two subgroups of violent reoffending in male prisoners. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ZC had full access to all the data in the study and, with SF, had final responsibility for the decision to submit for publication. Results We identified 47 326 prisoners during the study period (43 840 male and 3486 female prisoners), who we followed up for 10 years after release from prison. Baseline sociodemographic and criminological information and psychiatric diagnoses, and follow-up data in male and female prisoners are presented in table 1, and their associations with violent reoffending are presented in the appendix (pp 3–6). In male prisoners, 18 563 (42%) of 43 840 had been diagnosed with at least one psychiatric disorder before release, and 10 884 (25%) reoffended for violent crimes during follow-up. In female prisoners, a higher proportion (2233 [64%] of 3486) had been diagnosed with psychiatric disorder than had male prisoners, and fewer (379 [11%]) reoffended for violent crimes than did male prisoners. 11 804 (57%) of prisoners with psychiatric disorders had both inpatient and outpatient diagnoses (10 669 [57%] men and 1135 [51%] women). Types of violent reoffending are presented in the appendix (p 2); the most common category was assault (7171 [64%] of 11 263 individuals reoffending for a violent crime), followed by threats and intimidation, robbery, sexual offences, and homicide.

tient and outpatient diagnoses (10 669 [57%] men and 1135 [51%] women). Types of violent reoffending are presented in the appendix (p 2); the most common category was assault (7171 [64%] of 11 263 individuals reoffending for a violent crime), followed by threats and intimidation, robbery, sexual offences, and homicide. The overall Kaplan-Meier curve for violent reoffending in released prisoners is presented in the appendix (p 1). Prisoners with any psychiatric disorder had a higher rate of violent reoffending than did those without a disorder (figure 1). In male prisoners, the median time to first violent reoffending was 2·4 months shorter for those with psychiatric disorder (median 14·2 [IQR 5·1–31·8]) than with those without (16·6 [6·2–35·2]). In female prisoners, time to violent reoffending was 4·8 months shorter for those with psychiatric disorder (18·4 [6·0–38·3]) than with those without (23·2 [10·3–41·5]). Prisoners with psychiatric disorder had a high probability of violent reoffending (figure 2): over 5 years, the probability was 0·41 (95% CI 0·40–0·42) for male prisoners with psychiatric disorder and 0·25 (0·25–0·26) for those without. In female prisoners, violent reoffending probabilities were 0·20 (0·17–0·22) for those with psychiatric disorder and 0·08 (0·06–0·10) for those without.

reoffending (figure 2): over 5 years, the probability was 0·41 (95% CI 0·40–0·42) for male prisoners with psychiatric disorder and 0·25 (0·25–0·26) for those without. In female prisoners, violent reoffending probabilities were 0·20 (0·17–0·22) for those with psychiatric disorder and 0·08 (0·06–0·10) for those without. Cox regression analysis showed that, in male prisoners, psychiatric disorder was associated with an increased hazard of violent reoffending (model 1: HR 2·10 [95% CI 2·02–2·19]; table 2). The association was attenuated but remained substantial after adjustment for sociodemographic and criminological factors (model 2: 1·63 [1·57–1·70]). We further compared prisoners who were full siblings, and psychiatric disorder was still associated with an increased hazard of violent reoffending (model 3: 2·01 [1·66–2·43]). In female prisoners, psychiatric disorder was also associated with a higher hazard of violent reoffending (model 1: 2·76 [2·15–3·55]), and after adjustment (model 2: 2·02 [1·54–2·63]). However, the association was non-significant in the sibling model, with wide confidence intervals. We recorded similar results when analysing all siblings of prisoners (including non-prisoner siblings, appendix p 7). We also found similar results in young and adult men (appendix p 8). Even in the most adjusted model, our data provide sufficient events per variable (EPV; men: 602 EPV [10 844 events per 18 variables]; women 21 EPV [379 events per 18 variables]; 20 is deemed a sufficient number of EPV).41,42

r siblings, appendix p 7). We also found similar results in young and adult men (appendix p 8). Even in the most adjusted model, our data provide sufficient events per variable (EPV; men: 602 EPV [10 844 events per 18 variables]; women 21 EPV [379 events per 18 variables]; 20 is deemed a sufficient number of EPV).41,42 10 884 incidents of violent reoffending occurred in male prisoners after release. Of these, 2187 were potentially attributable to psychiatric disorder. This corresponds to a PAF of 20% (95% CI 19–22). In female prisoners, 152 of 379 incidents of violent reoffending were potentially attributable to psychiatric disorder, with a corresponding PAF of 40% (27–52). When we explored individual psychiatric disorders, all diagnoses were associated with an increased hazard of violent reoffending, even after adjustment for possible confounders (except for schizophrenia spectrum disorders in women in model 3), but the magnitude of associations varied and some hazards were not significantly increased in women (table 3). We found the strongest associations for alcohol and drug use disorders, personality disorder, attention-deficit hyperactivity disorder, other developmental or childhood disorders, schizophrenia spectrum disorders, and bipolar disorder.

iations varied and some hazards were not significantly increased in women (table 3). We found the strongest associations for alcohol and drug use disorders, personality disorder, attention-deficit hyperactivity disorder, other developmental or childhood disorders, schizophrenia spectrum disorders, and bipolar disorder. Because of the small sample size of female prisoners, we did the following analyses in male prisoners only. The proportion of male prisoners who violently reoffended and had any psychiatric disorder along with substance use disorder comorbidity was higher than in those without this comorbidity, and the adjusted HR was also higher (table 4). A test of interaction between any psychiatric disorder and substance use disorder was not significant. We noted similar results for schizophrenia spectrum and bipolar disorder. The hazard of violent reoffending increased in a stepwise way according to the number of psychiatric disorders (figure 3). Individuals with four or more psychiatric disorders had a substantially increased hazard of reoffending compared with those without psychiatric disorder (adjusted HR 2·74 [95% CI 2·45–3·06]). In sensitivity analyses (appendix pp 9–11), our findings did not differ when we restricted outcomes to interpersonal violent crimes or other violent crimes, or used imputed samples.

ad a substantially increased hazard of reoffending compared with those without psychiatric disorder (adjusted HR 2·74 [95% CI 2·45–3·06]). In sensitivity analyses (appendix pp 9–11), our findings did not differ when we restricted outcomes to interpersonal violent crimes or other violent crimes, or used imputed samples. Discussion In this longitudinal study, we have shown that psychiatric disorders were associated with a substantially increased hazard of violent reoffending. The association was independent of a number of measured sociodemographic, criminological, and familial factors, except that the finding in female prisoners was non-significant when taking familial factors into account. To our knowledge, we are the first to use a sibling design to study reoffending in an unselected prison population. Additionally, with important caveats, we have estimated the population impact of psychiatric disorders on violent reoffending.

male prisoners was non-significant when taking familial factors into account. To our knowledge, we are the first to use a sibling design to study reoffending in an unselected prison population. Additionally, with important caveats, we have estimated the population impact of psychiatric disorders on violent reoffending. Our study has three main findings. First, any diagnosed psychiatric disorder was associated with a substantially increased hazard of violent reoffending; however, the hazard ratio decreased after adjustment for sociodemographic and criminological factors, suggesting that about 40% of the excess violent reoffending was due to these factors. More importantly, this result suggests that psychiatric disorders (which included both inpatient and outpatient diagnoses) were associated with an increased hazard of violent reoffending independently of these factors. This finding is by contrast with the findings of systematic reviews,10,20 some expert opinion,21 and scores assigned to mental disorders in widely used risk assessment instruments in criminal justice.43 But it is in keeping with a few cohort investigations, although these studies have used small numbers of prisoners44 or selected samples of high-risk prisoners45 or community offenders.46 Furthermore, our findings are consistent with those from a large retrospective study in the USA,47 which showed that psychiatric disorders are associated with increased hazard of previous incarcerations. However, because of the retrospective design, investigators of this study were unable to show a temporal sequence between exposure and outcome, and exclude the possibility of reverse causation.

e study in the USA,47 which showed that psychiatric disorders are associated with increased hazard of previous incarcerations. However, because of the retrospective design, investigators of this study were unable to show a temporal sequence between exposure and outcome, and exclude the possibility of reverse causation. Additionally, we noted no evidence of familial confounding on the association between psychiatric disorder and violent reoffending in men, but in women, adjustment for familial confounding made the finding non-significant. The temporality between measures of psychiatric disorders and violent reoffending, and the gradient effect of number of diagnoses on reoffending provide additional corroboration for a causal hypothesis.48 However, for causality to be clearly shown, these findings will need validation in other released prisoner cohorts and treatment trials will need to be done.

atric disorders and violent reoffending, and the gradient effect of number of diagnoses on reoffending provide additional corroboration for a causal hypothesis.48 However, for causality to be clearly shown, these findings will need validation in other released prisoner cohorts and treatment trials will need to be done. To our knowledge, we calculated PAFs to estimate the population impact of diagnosed psychiatric disorders on violent reoffending for the first time. PAFs assume causality, so our estimates should be interpreted with much caution. Additionally, because we did not have reliable information about all possible covariates and thus could not include them in our models, the reported PAFs are likely to be overestimates. Diagnostic comorbidities (such as personality disorder) and social factors that co-occur with psychiatric disorders (including victimisation and homelessness) will probably reduce our PAF estimates. Generalisation to other countries with different criminal cultures should not be made without further research. Nevertheless, the PAF that we report shows a substantial contribution of psychiatric disorder to the high risk of reoffending. In some countries, this contribution to reoffending will also be important from a public health perspective in terms of absolute numbers of crimes. For example, in the USA, former prisoners account for an estimated 15–20% of all adult arrests,7 so even a small PAF would lead to substantial decreases in violent crimes from the 1·1 million committed in the USA in 2013.49 National violence prevention strategies, which have not included prison health in their targets, strategies, or surveillance,3 need review on the basis of our findings.

% of all adult arrests,7 so even a small PAF would lead to substantial decreases in violent crimes from the 1·1 million committed in the USA in 2013.49 National violence prevention strategies, which have not included prison health in their targets, strategies, or surveillance,3 need review on the basis of our findings. In line with previous research,17,50 we noted that a higher proportion of female prisoners had psychiatric disorders than did male prisoners. The hazard ratio for violent reoffending seemed to be higher in women prisoners than in men released from prison, although the absolute rate of violent reoffences were lower in women than in men. These findings are consistent with other research that shows that women with schizophrenia and related disorders have a higher relative risk of violence than do men with these disorders,15 and might be attributable to women who offend being more severely psychiatrically ill than are men who offend.51

women than in men. These findings are consistent with other research that shows that women with schizophrenia and related disorders have a higher relative risk of violence than do men with these disorders,15 and might be attributable to women who offend being more severely psychiatrically ill than are men who offend.51 The second main finding was that each individual psychiatric disorder was associated with a modest increased hazard of violent reoffending. This result was unexpected, particularly in men, for whom we found similar HRs for alcohol and drug use disorders, personality disorder, attention-deficit hyperactivity dis order, other developmental or childhood disorders, schizophrenia spectrum disorders, and bipolar disorder. This finding contrasts with studies in the general population showing substance use disorder to be associated with a higher risk of violent crime than are other psychiatric disorders (particularly if they are not comorbid with substance use disorder).52,53 A theoretical explanation for the nonspecificity that we report could be that psychiatric disorders share core psychopathological features,54 such as emotional dysregulation, which increase the risk of violence.

an are other psychiatric disorders (particularly if they are not comorbid with substance use disorder).52,53 A theoretical explanation for the nonspecificity that we report could be that psychiatric disorders share core psychopathological features,54 such as emotional dysregulation, which increase the risk of violence. The magnitude of the associations varied. Bipolar disorder was associated with a higher risk of violent reoffending in the familial adjusted model (model 3) than were other psychiatric disorders, apart from alcohol or drug use disorders, similar to a register-based US study.47 Prison health services have not focused on screening or treatment of bipolar disorder specifically, and replication of this finding and possible associations with severity and psychotic symptoms of the illness need further investigation. In women, we noted some evidence of heterogeneity by individual disorder, and the effect of alcohol use disorder seemed to be stronger than that of other psychiatric disorders. Additionally, the effect of alcohol use disorder seemed stronger in women than in men. Possible differences between various offender categories and types of violent reoffending should be considered (appendix pp 10–11). Although the overall effect of any psychiatric disorder was not materially different when different types of violent reoffending were investigated, whether affective disorders (eg, depression, bipolar disorder, and anxiety disorder) are associated with higher hazards of reoffending for less severe violent crimes than for more severe violent crimes needs further research.

er was not materially different when different types of violent reoffending were investigated, whether affective disorders (eg, depression, bipolar disorder, and anxiety disorder) are associated with higher hazards of reoffending for less severe violent crimes than for more severe violent crimes needs further research. The third main finding was that the association between psychiatric disorders and violent reoffending was not fully attributable to substance use disorder. In line with previous studies,22,23,33 we found that prisoners with severe mental illness (eg, schizophrenia spectrum disorders and bipolar disorder) and comorbid substance use disorder had a higher risk of violent reoffending than did those without comorbidity. However, we also showed that severe mental illness increased the hazard of violent reoffending, even without substance use disorder comorbidity (although this finding was not significant for bipolar disorder because of small numbers of patients with the disorder). Additionally, we found that the hazard of violent reoffending increased in a stepwise way according to the number of psychiatric disorders, and prisoners with multimorbidity of psychiatric disorders had a substantially increased risk of violent reoffending. These findings suggest that management of prisoners with psychiatric disorders should not merely focus on treatment of one disorder, but consider comorbidity and multimorbidity. The roles of antipsychotics,55 mood stabilisers,55 attention-deficit hyperactivity disorder medications,56 and psychological treatments in reduction of risks of repeat offending need investigation.

h psychiatric disorders should not merely focus on treatment of one disorder, but consider comorbidity and multimorbidity. The roles of antipsychotics,55 mood stabilisers,55 attention-deficit hyperactivity disorder medications,56 and psychological treatments in reduction of risks of repeat offending need investigation. A limitation of our study is reliance on data from patient registers for ascertainment of psychiatric diagnoses. Although these data have good diagnostic validity and the advantage of not relying on patient recall and self-report, the prevalence of some psychiatric disorders was underestimated. The prevalence of severe mental illness was similar to the pooled prevalence in a systematic review17 of more than 100 studies (eg, pooled prevalence of psychosis of 3·6% in male prisoners and 3·9% in female prisoners in the systematic review vs 3% in male prisoners and 4% in female prisoners in this study); however, the prevalence of attention-deficit hyper-activity disorder and other developmental or childhood disorders seems likely to have been under estimated.57,58 This underestimation might especially be the case for individuals released in the early 2000s who had shorter coverage of outpatient data than did individuals released later.

lence of attention-deficit hyper-activity disorder and other developmental or childhood disorders seems likely to have been under estimated.57,58 This underestimation might especially be the case for individuals released in the early 2000s who had shorter coverage of outpatient data than did individuals released later. Another important limitation is that personality disorder was probably underestimated in this study. The proportion of patients with the disorder in this study contrasts with findings from investigations that use structured instruments, which, despite very high heterogeneity between primary studies, report prevalences of more than 50% in male and about 40% in female prisoners.59 However, our estimates are similar to those of three large carefully done studies in both remand and sentenced populations in England and Wales of between 7% and 11%.60–62 This finding underscores a wider issue in personality disorder research of investigators using structured instruments reporting much higher prevalences than do those of clinically based investigations; these clinically based studies might more closely identify individuals with treatment needs than would structured instruments. The first implication of this underestimation is that we might have overstated the effect of other psychiatric disorders and substance use disorder on violent reoffending because these disorders are moderated by personality disorder. We think that this overestimation is unlikely because research in the general population has shown that comorbid personality disorder does not explain the associations between other psychiatric disorders and violent crime.63 Additionally, our sibling models partly adjust for personality disorder. A second implication is that our PAFs are overestimates because they do not fully include all possible risk factors for violent reoffending. At the same time, PAFs provide an indication of the possible effect of treatment of a risk factor on population estimates of violent reoffending, and the evidence base of effective treatment for personality disorders is weak, at least in the prison setting.64,65

nclude all possible risk factors for violent reoffending. At the same time, PAFs provide an indication of the possible effect of treatment of a risk factor on population estimates of violent reoffending, and the evidence base of effective treatment for personality disorders is weak, at least in the prison setting.64,65 A further limitation is that the study was done in one country. Although the prison population is small in Sweden,1 some key characteristics of prisoners in Sweden are not very different from those in other high-income countries (eg, prevalence of psychiatric disorders, proportion of prisoners reoffending, length of incarceration; appendix p 12). Nevertheless, the extent to which our findings can be generalised to other countries needs further research. Because Sweden has a well-developed public health system (similar to that of the UK, but more accessible than that of the USA), our findings are likely to be on the conservative side in terms of estimation of the effects of psychiatric disorders in the international context, and the association between psychiatric disorders and violent reoffending might be even stronger in countries with less resourced prison health services. Finally, because we have used registers, we have information about a restricted set of covariates. A complex set of risk factors is likely to be implicated in reoffending, with different factors acting at different points, some of which will be proximal and unaccounted for in registers.

esourced prison health services. Finally, because we have used registers, we have information about a restricted set of covariates. A complex set of risk factors is likely to be implicated in reoffending, with different factors acting at different points, some of which will be proximal and unaccounted for in registers. Many individuals with psychiatric disorders revolve between admission to hospital, homelessness, and the criminal justice system. Our findings underscore the need for improved detection, treatment, and management of prisoners with mental health disorders, and linkage of these prisoners to community-based mental care services on release.66 They also emphasise the need for further research into the role of psychiatric diagnoses in risk assessment for future offences and the effectiveness of diversion from criminal justice. Because the worldwide number of prisoners with psychiatric disorders is large, improvements to their treatment and management in custody and on release have the potential to improve their quality of life and counteract the cycle of reoffending.67 Supplementary Material Supplementary Appendix Acknowledgments This study was funded by grants from the Wellcome Trust (095806), Swedish Research Council, and Swedish Research Council for Health, Working Life and Welfare. Declaration of interests

Many individuals with psychiatric disorders revolve between admission to hospital, homelessness, and the criminal justice system. Our findings underscore the need for improved detection, treatment, and management of prisoners with mental health disorders, and linkage of these prisoners to community-based mental care services on release.66 They also emphasise the need for further research into the role of psychiatric diagnoses in risk assessment for future offences and the effectiveness of diversion from criminal justice. Because the worldwide number of prisoners with psychiatric disorders is large, improvements to their treatment and management in custody and on release have the potential to improve their quality of life and counteract the cycle of reoffending.67 Supplementary Material Supplementary Appendix Acknowledgments This study was funded by grants from the Wellcome Trust (095806), Swedish Research Council, and Swedish Research Council for Health, Working Life and Welfare. Declaration of interests SF reports travel expenses for a conference from Janssen and has provided expert testimony for deaths in custody in England and Northern Ireland. HL reports grants from Shire for eating disorders and speaking fees from Eli Lilly. All other authors declare no competing interests. See Online for appendix

Declaration of interests SF reports travel expenses for a conference from Janssen and has provided expert testimony for deaths in custody in England and Northern Ireland. HL reports grants from Shire for eating disorders and speaking fees from Eli Lilly. All other authors declare no competing interests. See Online for appendix Research in context Evidence before this study We searched PubMed from Jan 1, 2004, to Dec 31, 2014, using the search terms “psychiatric disorder*”, “mental disorder*”, “mental illness*”, “recidi*”, “reoffend*”, and “repeat offend*”, with no language restrictions. We identified one systematic review on the risk of repeat offending in individuals with psychotic disorders and a further 13 studies since publication of that review (appendix pp 16–18). Although authors of one review found some evidence in support of psychotic disorders increasing reoffending risk, only one included study looked at prisoners and none studied violent reoffending. In these 13 newer studies, findings were inconsistent, with those from eight studies showing no independent association between psychiatric disorders and reoffending. However, seven of these studies were small or in selected samples, and the remaining one did not account for important sociodemographic confounders. Furthermore, we identified no studies that considered the effect of familial factors on the link between psychiatric disorders and reoffending. We identified one systematic review on personality disorders and repeat offending and one review on risk of recidivism for offenders with mental disorders. Authors of the first noted that personality disorders were consistently associated with an increased risk of reoffending, with low heterogeneity between these primary reports and little difference between all personality disorders and antisocial personality disorder. Authors of the second, which included heterogeneous samples, concluded that antisocial personality disorder predicts recidivism.

ciated with an increased risk of reoffending, with low heterogeneity between these primary reports and little difference between all personality disorders and antisocial personality disorder. Authors of the second, which included heterogeneous samples, concluded that antisocial personality disorder predicts recidivism. Added value of this study We are the first, to our knowledge, to investigate the association between psychiatric disorders and violent reoffending while taking into account both measured (sociodemographic and criminological) and unmeasured (familial) confounding factors. Use of a total population cohort of released prisoners enabled us to provide precise effect sizes and estimate the possible population impact of psychiatric disorders on violent reoffending. We found that some psychiatric disorders were associated with a substantially increased hazard of violent reoffending. We reported some evidence for potentially important heterogeneity between individual diagnoses and risk of violent reoffending. Implications of all the available evidence Our findings underscore the need for improved detection, treatment, and management of prisoners with psychiatric and substance use disorders, and linkage of these prisoners to community-based mental health services on release. Further research into the role of psychiatric diagnoses in violent risk assessment and the effectiveness of diversion from criminal justice is needed. Figure 1 Kaplan-Meier curve (unadjusted model) for violent reoffending in released prisoners by sex and psychiatric disorder status HR=hazard ratio.

Implications of all the available evidence Our findings underscore the need for improved detection, treatment, and management of prisoners with psychiatric and substance use disorders, and linkage of these prisoners to community-based mental health services on release. Further research into the role of psychiatric diagnoses in violent risk assessment and the effectiveness of diversion from criminal justice is needed. Figure 1 Kaplan-Meier curve (unadjusted model) for violent reoffending in released prisoners by sex and psychiatric disorder status HR=hazard ratio. Figure 2 Probability of violent reoffending by sex, time after release, and psychiatric disorder status Error bars are 95% CIs. Figure 3 Association between number of psychiatric disorders and violent reoffending in male released prisoners Error bars are 95% CIs.

Figure 1 Kaplan-Meier curve (unadjusted model) for violent reoffending in released prisoners by sex and psychiatric disorder status HR=hazard ratio. Figure 2 Probability of violent reoffending by sex, time after release, and psychiatric disorder status Error bars are 95% CIs. Figure 3 Association between number of psychiatric disorders and violent reoffending in male released prisoners Error bars are 95% CIs. Table 1 Baseline sociodemographic and criminological information, and follow-up data for released prisoners in Sweden Men Women Number of individuals 43 840 3486 Number of person-years at risk 139 260 11 243 Incidents of violent reoffending during follow-up 10 884 (25%) 379 (11%) Age group (years) 16–24 8466 (19%) 361 (10%) 25–39 17 291 (39%) 1409 (40%) ≥40 18 083 (41%) 1716 (49%) Civil status Unmarried 26 910 (65%) 1614 (49%) Married 5066 (12%) 537 (16%) Divorced 9105 (22%) 1094 (33%) Widowed 222 (<1%) 67 (2%) Highest length of education (years) <9 19 546 (47%) 1765 (53%) 9–11 19 174 (46%) 1322 (40%) ≥12 2583 (6%) 225 (7%) Employed 8045 (20%) 355 (11%) Immigrant 13 710 (31%) 806 (23%) Disposable income (×100 Swedish Krona) 775 (473 to 1100) 749 (444 to 1082) Neighbourhood deprivation* 0·38 (−0·17 to 1·48) 0·35 (−0·15 to 1·46) Length of incarceration (months) <6 30 155 (69%) 2608 (75%) 6–11 7270 (17%) 506 (15%) 12–23 4408 (10%) 283 (8%) ≥24 2007 (5%) 89 (3%) Violent index offence 17 294 (39%) 643 (18%) Previous violent crime 23 960 (55%) 1112 (32%) Previous psychiatric disorder Any psychiatric disorder 18 563 (42%) 2233 (64%) Alcohol use disorder 9276 (21%) 968 (28%) Drug use disorder 9597 (22%) 1438 (41%) Personality disorder 2320 (5%) 353 (10%) Attention-deficit hyperactivity disorder 546 (1%) 51 (1%) Other developmental or childhood disorder 979 (2%) 139 (4%) Schizophrenia spectrum disorders 1237 (3%) 130 (4%) Bipolar disorder 216 (<1%) 35 (1%) Depression 2553 (6%) 418 (12%) Anxiety disorder 3247 (7%) 534 (15%) Data are n, n (%), or median (IQR). 2573 men and 174 women had missing values for civil status, highest length of education, employment, disposable income, and neighbourhood deprivation.

m disorders 1237 (3%) 130 (4%) Bipolar disorder 216 (<1%) 35 (1%) Depression 2553 (6%) 418 (12%) Anxiety disorder 3247 (7%) 534 (15%) Data are n, n (%), or median (IQR). 2573 men and 174 women had missing values for civil status, highest length of education, employment, disposable income, and neighbourhood deprivation. * Standardised score of the overall degree of socioeconomic deprivation in an individual’s residential area. Table 2 Association between any psychiatric disorder and violent crime reoffending Number of person-years at risk Number of violent reoffences Model 1* Model 2† Model 3‡ Men With psychiatric disorder 50 904 5658 2·10 (2·02–2·19) 1·63 (1·57–1·70) 2·01 (1·66–2·43)§ Without psychiatric disorder 88 356 5226 1 1 1 Women With psychiatric disorder 6595 301 2·76 (2·15–3·55) 2·02 (1·54–2·63) 0·52 (0·08–3·15)¶ Without psychiatric disorder 4648 78 1 1 1 Data are n or hazard ratio (95% CI). * Adjusted for age and immigration status. † Adjusted for age, immigration status, and sociodemographic and criminological covariates. ‡ Fixed-effect sibling model, adjusted for all factors shared by siblings and measured covariates adjusted for in models 1 and 2. § Based on 1417 pairs of male prisoners who were full siblings. ¶ Based on 41 pairs of female prisoners who were full siblings. Table 3 Association between individual psychiatric disorders and violent crime reoffending Model 1* Model 2† Model 3‡ Men

‡ Fixed-effect sibling model, adjusted for all factors shared by siblings and measured covariates adjusted for in models 1 and 2. § Based on 1417 pairs of male prisoners who were full siblings. ¶ Based on 41 pairs of female prisoners who were full siblings. Table 3 Association between individual psychiatric disorders and violent crime reoffending Model 1* Model 2† Model 3‡ Men Alcohol use disorder 2·14 (2·05–2·24) 1·63 (1·56–1·71) 1·45 (1·38–1·53) Drug use disorder 2·13 (2·05–2·22) 1·65 (1·58–1·72) 1·52 (1·45–1·59) Personality disorder 2·29 (2·14–2·45) 1·64 (1·53–1·76) 1·30 (1·21–1·40) Attention-deficit hyperactivity disorder 2·22 (1·89–2·61) 1·56 (1·31–1·85) 1·31 (1·10–1·55) Other developmental or childhood disorder 1·82 (1·65–2·01) 1·46 (1·32–1·61) 1·33 (1·20–1·47) Schizophrenia spectrum disorders 2·06 (1·87–2·26) 1·51 (1·37–1·67) 1·20 (1·09–1·33) Bipolar disorder 1·96 (1·50–2·58) 1·75 (1·32–2·32) 1·50 (1·13–1·99) Depression 1·41 (1·30–1·54) 1·28 (1·18–1·40) 1·09 (1·00–1·18) Anxiety disorder 1·41 (1·32–1·51) 1·23 (1·14–1·32) 1·09 (1·01–1·17) Women

Alcohol use disorder 2·14 (2·05–2·24) 1·63 (1·56–1·71) 1·45 (1·38–1·53) Drug use disorder 2·13 (2·05–2·22) 1·65 (1·58–1·72) 1·52 (1·45–1·59) Personality disorder 2·29 (2·14–2·45) 1·64 (1·53–1·76) 1·30 (1·21–1·40) Attention-deficit hyperactivity disorder 2·22 (1·89–2·61) 1·56 (1·31–1·85) 1·31 (1·10–1·55) Other developmental or childhood disorder 1·82 (1·65–2·01) 1·46 (1·32–1·61) 1·33 (1·20–1·47) Schizophrenia spectrum disorders 2·06 (1·87–2·26) 1·51 (1·37–1·67) 1·20 (1·09–1·33) Bipolar disorder 1·96 (1·50–2·58) 1·75 (1·32–2·32) 1·50 (1·13–1·99) Depression 1·41 (1·30–1·54) 1·28 (1·18–1·40) 1·09 (1·00–1·18) Anxiety disorder 1·41 (1·32–1·51) 1·23 (1·14–1·32) 1·09 (1·01–1·17) Women Alcohol use disorder 2·65 (2·15–3·26) 2·08 (1·66–2·60) 1·84 (1·46–2·32) Drug use disorder 2·59 (2·10–3·20) 1·84 (1·46–2·30) 1·58 (1·26–2·00) Personality disorder 2·57 (1·99–3·33) 1·66 (1·27–2·18) 1·27 (0·96–1·68) Attention-deficit hyperactivity disorder 2·01 (0·95–4·25) 1·53 (0·72–3·27) 1·20 (0·56–2·57) Other developmental or childhood disorder 1·84 (1·29–2·64) 1·20 (0·82–1·76) 1·04 (0·70–1·53) Schizophrenia spectrum disorders 1·75 (1·11–2·74) 1·04 (0·64–1·69) 0·74 (0·45–1·20) Bipolar disorder 2·84 (1·06–7·65) 1·81 (0·67–4·91) 1·35 (0·49–3·68) Depression 1·49 (1·11–2·00) 1·36 (1·00–1·86) 1·16 (0·85–1·59) Anxiety disorder 1·40 (1·07–1·83) 1·21 (0·92–1·60) 1·07 (0·81–1·41) Data are hazard ratio (95% CI). * Adjusted for age and immigration status. † Adjusted for age, immigration status, and sociodemographic and criminological covariates.

Alcohol use disorder 2·65 (2·15–3·26) 2·08 (1·66–2·60) 1·84 (1·46–2·32) Drug use disorder 2·59 (2·10–3·20) 1·84 (1·46–2·30) 1·58 (1·26–2·00) Personality disorder 2·57 (1·99–3·33) 1·66 (1·27–2·18) 1·27 (0·96–1·68) Attention-deficit hyperactivity disorder 2·01 (0·95–4·25) 1·53 (0·72–3·27) 1·20 (0·56–2·57) Other developmental or childhood disorder 1·84 (1·29–2·64) 1·20 (0·82–1·76) 1·04 (0·70–1·53) Schizophrenia spectrum disorders 1·75 (1·11–2·74) 1·04 (0·64–1·69) 0·74 (0·45–1·20) Bipolar disorder 2·84 (1·06–7·65) 1·81 (0·67–4·91) 1·35 (0·49–3·68) Depression 1·49 (1·11–2·00) 1·36 (1·00–1·86) 1·16 (0·85–1·59) Anxiety disorder 1·40 (1·07–1·83) 1·21 (0·92–1·60) 1·07 (0·81–1·41) Data are hazard ratio (95% CI). * Adjusted for age and immigration status. † Adjusted for age, immigration status, and sociodemographic and criminological covariates. ‡ Adjusted for age, immigration status, sociodemographic and criminological covariates, and alcohol and drug use disorders. Table 4 Violent reoffending in male prisoners with psychiatric disorder with and without substance use disorder comorbidity Incidents of violent reoffending Adjusted hazard ratio* Without substance use disorder With substance use disorder Without substance use disorder With substance use disorder p value for interaction† Any psychiatric disorder‡ 764/3426 (22%) 1912/5504 (35%) 1·39 (1·29–1·50) 2·43 (2·30–2·57) 0·85 Schizophrenia spectrum disorder 66/303 (22%) 390/934 (42%) 1·29 (1·00–1·67) 2·68 (2·41–2·98) 0·44 Bipolar disorder 11/72 (15%) 41/144 (28%) 1·45 (0·75–2·79) 3·22 (2·35–4·39) 0·67 Data are n/N (%) or hazard ratio (95% CI).

n† Any psychiatric disorder‡ 764/3426 (22%) 1912/5504 (35%) 1·39 (1·29–1·50) 2·43 (2·30–2·57) 0·85 Schizophrenia spectrum disorder 66/303 (22%) 390/934 (42%) 1·29 (1·00–1·67) 2·68 (2·41–2·98) 0·44 Bipolar disorder 11/72 (15%) 41/144 (28%) 1·45 (0·75–2·79) 3·22 (2·35–4·39) 0·67 Data are n/N (%) or hazard ratio (95% CI). * Compared with prisoners without any psychiatric disorder, adjusted for age, immigration status, and sociodemographic and criminological covariates. † Between any psychiatric disorder and substance use disorder. ‡ Excluding substance use disorder.

Introduction Obsessive compulsive disorder is considered the fourth most common mental disorder in high-income countries and ranks as the tenth leading cause of disability worldwide.1, 2 It is associated with increased mortality3 and can have a substantial impact on quality of life for both patients and family members or carers.2 Clomipramine and the SSRIs are currently recommended for pharmacological management of the disease.4 Psychotherapies and especially behavioural or cognitive behavioural interventions have been developed5, 6 and are also recommended.7

substantial impact on quality of life for both patients and family members or carers.2 Clomipramine and the SSRIs are currently recommended for pharmacological management of the disease.4 Psychotherapies and especially behavioural or cognitive behavioural interventions have been developed5, 6 and are also recommended.7 Previous systematic reviews and meta-analyses have generally compared the efficacy of pharmacological interventions with placebo, not with each other.8, 9, 10 Psychotherapeutic interventions have typically been compared with a waiting list or other inactive therapy.7, 11 Only a few studies have directly compared psychotherapeutic with pharmacological interventions or combinations of them, and their results are inconclusive.7 In the absence of available head-to-head comparisons, indirect evidence can be used to enhance the existing evidence base. Indirect comparisons between different medications have been done in the past, but statistical methods appropriate for such comparisons were poorly developed at that time.10 Network meta-analysis is a method of synthesising information from a network of trials addressing the same question, but involving different interventions. It aims to combine direct and indirect evidence into a single effect size and rank all available treatments in terms of efficacy, providing estimates for interventions even if they have not been directly compared. This approach has been applied successfully to schizophrenia, bipolar disorder, depression, and certain anxiety disorders (social phobia and generalised anxiety disorder), but not obsessive-compulsive disorder. We therefore did a systematic review and network meta-analysis with the aim to simultaneously compare all available treatments using both direct and indirect data.12 A more detailed report than this one will be published, and data collected for children and adolescents will also be separately published.

lsive disorder. We therefore did a systematic review and network meta-analysis with the aim to simultaneously compare all available treatments using both direct and indirect data.12 A more detailed report than this one will be published, and data collected for children and adolescents will also be separately published. Research in context Evidence before this study

lsive disorder. We therefore did a systematic review and network meta-analysis with the aim to simultaneously compare all available treatments using both direct and indirect data.12 A more detailed report than this one will be published, and data collected for children and adolescents will also be separately published. Research in context Evidence before this study During the protocol stage of our project (May 1 to June 30, 2013), we did a scoping search of the literature. We used the two specialised registers of controlled trials maintained and administered by the Cochrane Collaboration Common Mental Disorders Group. We searched the registers using the generic term “condition = obsess* OR compulsi*”, with no language or date restrictions. We found that the latest comprehensive review had been published in 2006 and specific meta-analyses had been published in 2008. Since then, several new trials have been done. Previous systematic reviews and meta-analyses have generally focused on the comparison between antidepressant medications and placebo or psychotherapeutic interventions and a waiting list. Few studies have directly compared the relative efficacy of serotonergic antidepressants versus each other, behavioural-type psychotherapies versus each other, or medications versus psychotherapies. Clinicians are often interested in pragmatic comparisons (Are all SSRIs equally effective? Is clomipramine more effective than are SSRIs? Is cognitive behavioural psychotherapy more effective than are medications?), but these questions have been examined in few studies in the past using statistical methods that have not always taken into account the complexity of such comparisons. We therefore did a network meta-analysis with the aim to simultaneously compare in a single analysis and rank in terms of efficacy all available interventions for management of obsessive-compulsive disorder in adults.

using statistical methods that have not always taken into account the complexity of such comparisons. We therefore did a network meta-analysis with the aim to simultaneously compare in a single analysis and rank in terms of efficacy all available interventions for management of obsessive-compulsive disorder in adults. Added value of this study We found small differences in efficacy between medications, and the hypothesis of clomipramine being better than SSRIs was not confirmed. Although certain psychotherapies were associated with larger effects than were medications, we underline an important limitation that, in most psychotherapeutic trials, patients who were taking stable doses of antidepressants were not excluded and therefore these therapies cannot be considered as pure monotherapies. Implications of all the available evidence Taking all evidence into account, the combination of psychotherapies with medications is possibly the most effective intervention and clinicians should consider this option more often than at present for patients with severe obsessive-compulsive disorder. Psychotherapy is effective in symptomatic patients taking antidepressant medications, and its effect as monotherapy is not known. Future research should try to differentiate more clearly the effect of medications versus psychotherapy and monotherapy versus combined therapy, avoiding the limitations that we have underlined in this study.

ffective in symptomatic patients taking antidepressant medications, and its effect as monotherapy is not known. Future research should try to differentiate more clearly the effect of medications versus psychotherapy and monotherapy versus combined therapy, avoiding the limitations that we have underlined in this study. Methods Search strategy and selection criteria In this systematic review and network meta-analysis, we searched the two controlled trials registers maintained by the Cochrane Collaboration Common Mental Disorders group for trials published up to Feb 16, 2016, by experienced staff of the Cochrane Common Mental Disorders group using their standard methodology. Reports of trials for inclusion in the Group's registers are collated from routine (weekly), generic searches of MEDLINE, Embase, and PsycINFO; quarterly searches of the Cochrane Central Register of Controlled Trials; and review-specific searches of additional databases. We searched the registers using the generic term “condition = obsess* OR compulsi*”, with no language restrictions. We included studies in the review if they were randomised controlled trials of adult patients with a diagnosis of obsessive-compulsive disorder. We allowed all comorbidities except for schizophrenia or bipolar disorder. We excluded studies that focused exclusively on treatment-resistant patient populations defined within the same study.

s in the review if they were randomised controlled trials of adult patients with a diagnosis of obsessive-compulsive disorder. We allowed all comorbidities except for schizophrenia or bipolar disorder. We excluded studies that focused exclusively on treatment-resistant patient populations defined within the same study. Eligible experimental interventions were all antidepressants7 and psychotherapeutic interventions7 recommended by current guidelines—ie, behavioural therapy, including exposure and response prevention but not explicit cognitive techniques (such as cognitive restructuring); cognitive therapy, including cognitive restructuring but not explicit behavioural techniques; and cognitive behavioural therapy (CBT). In psychotherapy trials that used both an individual and group format, we extracted data only for groups with the individual format. Eligible control interventions were drug placebo, psychological placebo (any credible psychological intervention that includes only non-specific components of therapy, such as general stress management or relaxation), and any other non-specific psychotherapeutic relationship. Inclusion and exclusion criteria were independently assessed by two reviewers (HB and PSk) and validated by one reviewer (PSk). For studies that were excluded, we noted the main reason for exclusion.

of therapy, such as general stress management or relaxation), and any other non-specific psychotherapeutic relationship. Inclusion and exclusion criteria were independently assessed by two reviewers (HB and PSk) and validated by one reviewer (PSk). For studies that were excluded, we noted the main reason for exclusion. Data analysis Data extraction was done independently by two reviewers (HB and PSk) and validated by one reviewer (PSk). We used standardised data extraction Word forms and structured Excel spreadsheets to extract data from published reports. In cases of duplicate data, we selected the manuscript with the largest sample. We also considered preliminary congress abstracts duplicate and did not select them if a full article had been published after the congress. We extracted data for inclusion and exclusion criteria (study design, experimental intervention, control intervention, age range, primary diagnosis, comorbid diagnoses, and use of diagnostic criteria), general details of the study (country, treatment setting, and length of follow-up), details of continuous outcome assessment (number of patients eligible for randomisation, randomised, dropped out, and remaining at the end of study, and baseline, end of treatment, and change from baseline Yale-Brown Obsessive Compulsive Scale [YBOCS] scores, with SDs), and details of the risk of bias assessment (intention-to-treat analysis, use of methods for handling missing data, and dropouts).

isation, randomised, dropped out, and remaining at the end of study, and baseline, end of treatment, and change from baseline Yale-Brown Obsessive Compulsive Scale [YBOCS] scores, with SDs), and details of the risk of bias assessment (intention-to-treat analysis, use of methods for handling missing data, and dropouts). For the quantitative synthesis, the primary outcome measure was continuous and it was symptom severity as measured by YBOCS.13 Our preferred measure was mean change from baseline score. For studies in which this measure was not reported, we used mean YBOCS scores at the end of study after checking that YBOCS at baseline was balanced across groups. We report mean differences with 95% credible intervals compared with placebo. We assessed risk of bias using the criteria suggested by the Cochrane Collaboration Handbook.14 We included studies with a high risk of bias in the main analysis but did sensitivity analyses to examine the effect of excluding them.

ss groups. We report mean differences with 95% credible intervals compared with placebo. We assessed risk of bias using the criteria suggested by the Cochrane Collaboration Handbook.14 We included studies with a high risk of bias in the main analysis but did sensitivity analyses to examine the effect of excluding them. We did pairwise and network meta-analyses for efficacy. We excluded studies that did not use YBOCS. This post-hoc decision was made for two reasons: YBOCS is the only available clinician-rated scale that has been extensively validated in controlled trials worldwide13 and use of a single scale allowed us to use the mean difference instead of the standardised mean difference, avoiding the methodological and interpretational difficulties associated with use of standardised mean difference.14 Where possible, we derived missing SDs from reported statistics following guidance in the Cochrane Collaboration Handbook.14 Where possible, we analysed the intention-to-treat population; otherwise, we used reported results for participants who completed the study.

associated with use of standardised mean difference.14 Where possible, we derived missing SDs from reported statistics following guidance in the Cochrane Collaboration Handbook.14 Where possible, we analysed the intention-to-treat population; otherwise, we used reported results for participants who completed the study. We did all analyses in a Bayesian framework using OpenBUGS version 3.2.3. We used the random-effect models described by Dias and colleagues,15 modified to incorporate an additional class hierarchy,16 such that all SSRIs were assumed to be similar, with a common class mean effect and between-SSRI variability about this class mean. We used flat priors for all parameters. We assessed heterogeneity by examining the posterior median of the between-study heterogeneity parameter from the random-effects model. To assess variability within studies, we used what was reported by trial authors. For continuous measures SDs were reported and for ratio measures typically SEs. However, where these statistics were not reported, we used methods recommended by the Cochrane Collaboration Handbook14 (eg, estimation of SEs from CIs). We measured goodness of fit with the posterior mean of the residual deviance. To assess inconsistency between direct and indirect evidence, we compared the fit of a model assuming consistency with that of one that relaxes this assumption (unrelated mean-effects model).17 We also compared the results of the pairwise meta-analysis with those of the network meta-analysis. All OpenBUGS code is available in the appendix.

cy between direct and indirect evidence, we compared the fit of a model assuming consistency with that of one that relaxes this assumption (unrelated mean-effects model).17 We also compared the results of the pairwise meta-analysis with those of the network meta-analysis. All OpenBUGS code is available in the appendix. Preplanned sensitivity analyses excluded studies at high risk of bias for the following domains: masking of the outcome assessor, incomplete outcome data, and high overall attrition or evidence of differential attrition between groups. We present the results both before (ie, the full dataset) and after excluding waiting list controlled trials. These trials are non-masked and evidence exists that they lead to biased results in favour of the active psychotherapeutic interventions.18, 19, 20 We did separate meta-regressions assuming a common interaction term for the following study-level characteristics: length of trial, publication date, industry sponsorship, and inclusion of patients with current comorbid depression. This study is registered with PROSPERO, number CRD42012002441. Role of the funding source The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Preplanned sensitivity analyses excluded studies at high risk of bias for the following domains: masking of the outcome assessor, incomplete outcome data, and high overall attrition or evidence of differential attrition between groups. We present the results both before (ie, the full dataset) and after excluding waiting list controlled trials. These trials are non-masked and evidence exists that they lead to biased results in favour of the active psychotherapeutic interventions.18, 19, 20 We did separate meta-regressions assuming a common interaction term for the following study-level characteristics: length of trial, publication date, industry sponsorship, and inclusion of patients with current comorbid depression. This study is registered with PROSPERO, number CRD42012002441. Role of the funding source The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results We identified 1480 articles in our search and assessed 158 (11%) full-text articles for eligibility (figure 1). We excluded 95 (60%) articles and included 64 trials reported in 63 (40%) articles21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83 in the qualitative review. A detailed list of the excluded studies is in the appendix. From the 63 articles eligible for inclusion in the network meta-analysis, we excluded ten (16%): nine (14%)22, 32, 35, 36, 54, 62, 73, 77, 80 did not use YBOCS and one (2%)47 was not connected to the network (details of these studies in appendix), leaving 54 trials reported in 53 (34%) articles21, 23, 24, 25, 26, 27, 28, 29, 30, 31, 33, 34, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 74, 75, 76, 78, 79, 81, 82, 83 included in the network meta-analysis (quantitative review). 7302 patients were randomly allocated in the qualitative review; however, 7014 (96%) were randomly allocated in the network meta-anlysis, with 288 (4%) excluded. Only 6652 (91%) contributed to the network meta-analysis since some trials did not report outcomes for all participants.

meta-analysis (quantitative review). 7302 patients were randomly allocated in the qualitative review; however, 7014 (96%) were randomly allocated in the network meta-anlysis, with 288 (4%) excluded. Only 6652 (91%) contributed to the network meta-analysis since some trials did not report outcomes for all participants. The 64 trials included in the qualitative review were published over a period of 33 years (1980–2012; table 1; detailed characteristics in appendix). In most psychotherapeutic trials, patients were not excluded if they were taking a stable dose of antidepressants for at least 3 months before inclusion (13 [72%] of all 18 psychotherapeutic trials and 12 [80%] of the 15 psychotherapeutic trials included in the network meta-analysis explicitly allowed antidepressants). In these trials, the proportion of patients on antidepressant medication varied, ranging from 13% to 100% and, in more than two-thirds of studies with the information available, was 45% or higher (detailed description in appendix). Patients were not allowed to make dose adjustments during trials, but no specific information was provided on how this criterion had been monitored by authors. Participants had long-standing and severe obsessive-compulsive disorder. Demographic and clinical characteristics of participants were similar across comparisons.

tients were not allowed to make dose adjustments during trials, but no specific information was provided on how this criterion had been monitored by authors. Participants had long-standing and severe obsessive-compulsive disorder. Demographic and clinical characteristics of participants were similar across comparisons. The 54 trials included in the network meta-analysis (quantitative review) involved 17 different treatments grouped into 12 classes (all six SSRIs were grouped into the same class; figure 2). Overall, of the 136 unique pairwise comparisons that could be made between the 17 treatment conditions, only 37 (27%) were studied head to head in the included studies. A detailed table of the data used in the analysis is in the appendix. Six (11%) trials used a waiting list control group: five (9%) CBT studies23, 31, 40, 50, 71 including 157 patients, 80 (51%) of whom had been randomly allocated to CBT, and one (2%) behavioural therapy study56 including 40 patients, 20 (50%) of whom had been randomly allocated to behavioural therapy. The behavioural therapy trial that used the waiting list as a control group56 was clearly an outlier in terms of efficacy (mean YBOCS difference from waiting list at the end of study −30·87). The network meta-analysis model gave an adequate fit to the data and we identified no evidence of inconsistency (posterior mean of the residual deviance was 104·6 in the network meta-analysis assuming consistency and 105·8 assuming inconsistency compared with 107 data points). Furthermore, the deviance information criterion was similar for the models with (480·8) and without (479·1) the consistency assumption. The posterior median SD for the consistency model was 3·10 (95% credible interval 2·46–3·95), whereas for the inconsistency model, this value was reduced to 1·75 (1·18–2·53).

points). Furthermore, the deviance information criterion was similar for the models with (480·8) and without (479·1) the consistency assumption. The posterior median SD for the consistency model was 3·10 (95% credible interval 2·46–3·95), whereas for the inconsistency model, this value was reduced to 1·75 (1·18–2·53). Most active interventions showed a significant reduction in mean YBOCS compared with drug placebo, regardless of inclusion or exclusion of trials using waiting list controls (table 2). Venlafaxine and psychological placebo both showed reductions in mean YBOCS, but they were not significant. The waiting list was the only so-called intervention that was associated with an increase in mean YBOCS compared with drug placebo. The effects of individual SSRIs were similar in magnitude. Clomipramine had a larger effect compared with placebo than did SSRIs, but the difference was not significant (table 3). All three psychotherapeutic interventions (behavioural therapy, cognitive therapy, and CBT) showed greater efficacy than did drug placebo. However, in the full analysis, CBT was less efficacious than were the other two and was not different from psychological placebo (appendix). Exclusion of studies that had used waiting list control groups led to a larger effect for CBT, which was significantly different from psychological placebo and similar to the other two psychotherapies.

nalysis, CBT was less efficacious than were the other two and was not different from psychological placebo (appendix). Exclusion of studies that had used waiting list control groups led to a larger effect for CBT, which was significantly different from psychological placebo and similar to the other two psychotherapies. In the full network, both behavioural and cognitive therapy had a larger reduction in mean YBOCS than did SSRIs as a class (table 3). CBT also had a lower mean YBOCS than SSRIs as a class, but only after excluding waiting list controlled trials. We observed similar results when comparing different types of psychotherapies with clomipramine as the reference (detailed results for all possible comparisons are shown in the appendix). However, in psychotherapeutic trials, most patients were taking stable doses of antidepressant medications for the whole duration of the trial. The same applies to the comparison between combinations of medications and psychotherapy versus psychotherapy alone as patients in these network comparisons were not in strict monotherapy (table 2, table 3). In all of these comparisons, differences were small. Excluding waiting list controlled trials, the combination of behavioural therapy with clomipramine was associated with the largest effect, but this combination has been used in only a single trial.38

work comparisons were not in strict monotherapy (table 2, table 3). In all of these comparisons, differences were small. Excluding waiting list controlled trials, the combination of behavioural therapy with clomipramine was associated with the largest effect, but this combination has been used in only a single trial.38 For all 64 trials included in the qualitative review, results of the risk of bias assessment for trials with at least one drug arm (46 [72%] of 64) and those with psychotherapy arms only (18 [28%] of 64) are presented in the appendix. Sequence generation (13 [20%] of 64) and random allocation concealment (eight [13%] of 64) were specifically described (ie, low risk of bias) in few studies. In trials with psychotherapy arms, masking of participants or those delivering the intervention was not possible (seven [39%] of these 18 trials used outcome assessors who were masked to treatment allocation). In the drug only trials, specification of the double-blind method (eg, identical capsules) was described in 15 (39%) of 38 trials. Handling of incomplete outcome data with an acceptable method was reported in 28 (61%) of the 46 trials with at least one drug arm and six (33%) of the 18 trials with psychotherapy arms only. A high proportion of the trials with drug arms were sponsored by pharmaceutical companies (table 1).

d in 15 (39%) of 38 trials. Handling of incomplete outcome data with an acceptable method was reported in 28 (61%) of the 46 trials with at least one drug arm and six (33%) of the 18 trials with psychotherapy arms only. A high proportion of the trials with drug arms were sponsored by pharmaceutical companies (table 1). For the sensitivity analyses, we used the full network (detailed results given in the appendix). In the first analysis, we included the 33 (61%) trials with low overall (<25%) and differential (<15%) attrition. This analysis led to a larger effect for CBT than in the full analysis, which was then very similar to the other two psychotherapies. In the second analysis, we included 34 (63%) trials that met the criterion of low risk of bias in the domain of incomplete outcome assessment, and the main finding was that clomipramine had a smaller effect than in the full analysis that was not different from that of SSRIs. In this analysis, we excluded all cognitive therapy trials as they had reported completers analyses. In the third analysis, we included the 17 (31%) trials that used a masked outcome assessor. Overall, results were similar to those of the full analysis, but the power was compromised because of the small sample size. We carried out separate meta-regressions to test the effect of length of trial, publication date, industry sponsorship, and inclusion of patients with current comorbid depression. The effects of these variables were small, and none were significant (appendix).

the power was compromised because of the small sample size. We carried out separate meta-regressions to test the effect of length of trial, publication date, industry sponsorship, and inclusion of patients with current comorbid depression. The effects of these variables were small, and none were significant (appendix). Discussion In this network meta-analysis, we found that several pharmacological and psychotherapeutic interventions can be considered more efficacious than is drug placebo. We found that SSRIs are generally equally efficacious, with no evidence to suggest that one drug is better than the others are. Their effect compared with placebo is statistically significant, but the estimated mean difference is generally moderate. In the full analysis, clomipramine showed a trend for a larger effect than with SSRIs that was not statistically significant. This finding contrasts with previous direct analyses, which postulated that clomipramine might be more efficacious than are SSRIs.10 This comparison was sensitive to studies with incomplete outcome assessment: some old clomipramine trials reported completers analyses only, and exclusion of these trials led to a lower effect for clomipramine than that of not excluding them, which was indistinguishable from that of SSRIs.

efficacious than are SSRIs.10 This comparison was sensitive to studies with incomplete outcome assessment: some old clomipramine trials reported completers analyses only, and exclusion of these trials led to a lower effect for clomipramine than that of not excluding them, which was indistinguishable from that of SSRIs. An unexpected finding was that in our main analysis, CBT had a smaller effect than that of behavioural or cognitive therapy. However, after exclusion of waiting list controlled trials, all differences between psychotherapies were not significant. The waiting list was the only so-called intervention that led to an increase in mean YBOCS score compared with drug placebo, and psychological placebo was very similar to drug placebo after exclusion of waiting list controlled trials. Research has also shown that trials using control groups with no or minimal contact with therapists usually lead to grossly overestimated effect sizes for active psychotherapeutic interventions.18, 84, 85 We obtained similar findings in the sensitivity analysis after exclusion of trials with high overall attrition to those from the main analysis after exclusion of waiting list controlled trials—ie, no difference between psychotherapies. The evidence for cognitive therapy mostly comes from trials that had compared it with behavioural therapy, with most of them not reporting intention-to-treat analyses, and these trials might have overestimated the effect of cognitive (and behavioural) therapy. The behavioural therapy trial that used the waiting list as a control group56 was clearly an outlier in terms of efficacy, and excluding it from the analysis reduced the effects for both behavioural and cognitive therapy, but not significantly. CBT has more links with other interventions and a more extensive network of trials than do cognitive and behavioural therapy and has been compared directly with several drugs in the same trial.26, 71, 74, 75 Taking all of this evidence into account, our analysis does not support the view that the three types of psychotherapy have different effects in obsessive-compulsive disorder.

e network of trials than do cognitive and behavioural therapy and has been compared directly with several drugs in the same trial.26, 71, 74, 75 Taking all of this evidence into account, our analysis does not support the view that the three types of psychotherapy have different effects in obsessive-compulsive disorder. Our analysis shows that all psychotherapies, either in the full dataset (for behavioural and cognitive therapy) or after exclusion of the waiting list controlled trials (for CBT), were more likely to lead to a larger effect than were medications. Some previous meta-analyses have reported similar results in favour of psychotherapy. For example, Cuijpers and colleagues86 examined the differential effect of pharmacotherapy and psychotherapy in major depression, dysthymia, panic disorder, social anxiety disorder, and obsessive-compulsive disorder, and reported a positive effect for psychotherapy compared with medications only for obsessive-compulsive disorder. One important limitation exists that, to our knowledge, has not been recognised before: most patients included in trials that used exclusively psychotherapeutic interventions were allowed to continue taking their antidepressant medications. Combination trials that had both psychotherapeutic and drug arms, or arms with both psychotherapy and drugs, explicitly excluded patients on antidepressant medications by design (and half of these trials were of CBT and half were of behavioural therapy). Therefore, psychotherapy trials have essentially compared different psychotherapeutic interventions in patients taking stable doses of antidepressant medications. Some evidence exists from other trials that focused exclusively on treatment-refractory patients that addition of CBT for patients with SSRI-refractory obsessive-compulsive disorder is more efficacious than is either psychological placebo87 or risperidone.88 In our analysis, although patients were symptomatic at study recruitment, what the effect would be if patients had been tapered off their antidepressant medication before randomisation is unknown because such studies have not been done. This issue has also been reported in meta-analyses of bipolar depression in which randomly allocated patients are allowed to continue using their mood stabilisers or anxiolytic medications.89 In any case, generalisation of these results for psychotherapeutic interventions in patients not taking concurrent antidepressant medications is difficult.

ported in meta-analyses of bipolar depression in which randomly allocated patients are allowed to continue using their mood stabilisers or anxiolytic medications.89 In any case, generalisation of these results for psychotherapeutic interventions in patients not taking concurrent antidepressant medications is difficult. Therefore, the question of what is better as monotherapy in obsessive-compulsive disorder—medications or psychotherapy—cannot be answered given the current evidence. Our analysis has several limitations. Most studies were of short-term duration. As most of the studies that tested the efficacy of psychotherapeutic interventions included patients who were taking stable doses of antidepressant medications, generalisation of these results to patients not on medications is not possible. We were unable to test different doses of the same drug to investigate potential dose-response associations.90 Because of the scarce data, we could not treat alternative dosing schemes in pharmacological trials as different nodes in the network. Several old studies only reported completers analyses, including all cognitive therapy studies, limiting the usefulness of the sensitivity analysis in this domain. We did not consider the relative efficacy of the various interventions in different symptom dimensions of obsessive-compulsive disorder, and generalisation of the results in subgroups of patients with specific symptoms, such as hoarding, should be made with caution.

sefulness of the sensitivity analysis in this domain. We did not consider the relative efficacy of the various interventions in different symptom dimensions of obsessive-compulsive disorder, and generalisation of the results in subgroups of patients with specific symptoms, such as hoarding, should be made with caution. The results of our analysis generally support current National Institute for Health and Care Excellence guidelines.7 For pharmacological management, the recommendation to use SSRIs rather than clomipramine as the first-line agents is supported by our findings since SSRIs have better tolerability than does clomipramine and we identified no convincing evidence for clomipramine being more efficacious than are SSRIs. For non-pharmacological management, all three types of psychotherapy are probably more efficacious than is non-specific therapy, but evidence is limited to patients taking stable doses of antidepressant medication before initiating psychotherapy. The combined initiation of both medication and psychotherapy (either behavioural therapy or CBT) seemed an efficacious treatment. In our analysis excluding waiting list controlled trials, this combined treatment was best, but with considerable uncertainty. Given that most psychotherapeutic trials can also be considered variants of combination trials (since most patients were taking stable doses of antidepressant medications), the combination of SSRIs or clomipramine with psychotherapy is likely to offer more benefit to patients with severe illness than is monotherapy, but more research is needed than at present to support this hypothesis, including cost-effectiveness analyses.

atients were taking stable doses of antidepressant medications), the combination of SSRIs or clomipramine with psychotherapy is likely to offer more benefit to patients with severe illness than is monotherapy, but more research is needed than at present to support this hypothesis, including cost-effectiveness analyses. Further research should try to differentiate more clearly than at present the effect of medications versus psychotherapy and monotherapy versus combined therapy. Trials that investigate the effect of psychotherapy should monitor use of antidepressants in included patients or recruit patients who are willing to taper off their antidepressant medication before entering randomisation. As obsessive-compulsive disorder is a very heterogeneous condition, more pragmatic trials of longer duration than have been done so far are needed to test the efficacy of existing interventions in patients encountered in daily clinical practice (including those with other comorbid conditions) and the augmenting effect of medications in addition to psychotherapy or vice versa in patients with treatment-refractory obsessive-compulsive disorder. Supplementary Material Supplementary appendix Acknowledgments This project was funded by the National Institute for Health Research Health Technology Assessment programme (project number 10/104/41). The views and opinions expressed herein are those of the authors and do not necessarily reflect those of the Health Technology Assessment programme, National Institute for Health Research, National Health Service, or Department of Health.

ealth Research Health Technology Assessment programme (project number 10/104/41). The views and opinions expressed herein are those of the authors and do not necessarily reflect those of the Health Technology Assessment programme, National Institute for Health Research, National Health Service, or Department of Health. Contributors PSk led the review, was responsible for managing the project, and drafted the report. DMC provided statistical support and did the network meta-analyses with PB. WH was a member of the review team, provided statistical support, and helped in the analysis. NAF was a member of the review team providing expert clinical advice (psychopharmacology) and helped in the writing of the report. PSa was a member of the review team providing expert clinical advice (psychotherapy). NJW provided statistical advice and helped with the statistical modelling. HB contributed to data extraction and the systematic review. DK was a member of the review team providing expert clinical advice. RC provided advice for the data searches, the systematic review, and methods. GL provided advice on the methods and systematic review and helped in writing of sections of the report. DMC, PB, NJW, and GL helped in interpretation of the results. WH, NAF, PSa, DK, RC, and GL made critical comments that helped in interpretation of the results.

the data searches, the systematic review, and methods. GL provided advice on the methods and systematic review and helped in writing of sections of the report. DMC, PB, NJW, and GL helped in interpretation of the results. WH, NAF, PSa, DK, RC, and GL made critical comments that helped in interpretation of the results. Declaration of interests PSk has received non-financial support from Lundbeck to attend a conference during the conduct of this study. GL is a board member of the National Institute for Health Research Efficacy and Mechanism Evaluation programme. NAF reports grants and non-financial support from the National Institute for Health Research during the conduct of this study; grants and personal fees from GlaxoSmithKline and AstraZeneca; non-financial support from Novartis; personal fees and non-financial support from the European College of Neuropsychopharmacology and Bristol-Myers Squibb; grants from the Medical Research Council and Wellcome Foundation; non-financial support from Janssen, the International College of Obsessive Compulsive Spectrum Disorders, the Journal of Behavioural Addiction, and WHO; grants, personal fees, non-financial support, and reimbursed registration to attend scientific meetings from Lundbeck; informal consultation without receiving payment for Transcept Pharmaceuticals; grants, personal fees, non-financial support, and reimbursed registration to attend scientific meetings from Servier; grants, non-financial support, and reimbursed registration to attend scientific meetings from Cephalon; personal fees, non-financial support, and reimbursed registration to attend scientific meetings from Jazz Pharmaceuticals; non-financial support and reimbursed registration to attend scientific meetings from the Royal College of Psychiatrists; and non-financial support and reimbursed registration to attend scientific meeting from the British Association for Psychopharmacology, all outside the submitted work. She is also medical lead to a National Health Service service that provides treatment for treatment-refractory obsessive-compulsive and related disorders, has been a council member for the British Association for Psychopharmacology, and sits on the Royal College of Psychiatrists Psychopharmacology Special Committee and the European College of Neuropsychopharmacology Education Committee and Research Network. DMC reports grants from the Medical Research Council Population Health Scientist fellowship during the conduct of this study.

chopharmacology, and sits on the Royal College of Psychiatrists Psychopharmacology Special Committee and the European College of Neuropsychopharmacology Education Committee and Research Network. DMC reports grants from the Medical Research Council Population Health Scientist fellowship during the conduct of this study. All other authors declare no competing interests. Figure 1 Study selection YBOCS=Yale-Brown Obsessive Compulsive Scale. Figure 2 Network diagram for efficacy analysis representing direct comparisons between individual treatments The size of each circle is proportional to the number of randomly allocated participants and the width of each line is proportional to the number of trials in each direct comparison. BT=behavioural therapy. CBT=cognitive behavioural therapy. CT=cognitive therapy. BTCLO=behavioural therapy and clomipramine. CBTFLV=cognitive behavioural therapy and fluvoxamine. CIT=citalopram. CLO=clomipramine. ESCIT=escitalopram. FLV=fluvoxamine. FLX=fluoxetine. HYP=hypericum. PAR=paroxetine. PL=placebo. PSYPL=psychological placebo. SER=sertraline. VEN=venlafaxine. WL=waiting list. Table 1 General characteristics of eligible studies

The size of each circle is proportional to the number of randomly allocated participants and the width of each line is proportional to the number of trials in each direct comparison. BT=behavioural therapy. CBT=cognitive behavioural therapy. CT=cognitive therapy. BTCLO=behavioural therapy and clomipramine. CBTFLV=cognitive behavioural therapy and fluvoxamine. CIT=citalopram. CLO=clomipramine. ESCIT=escitalopram. FLV=fluvoxamine. FLX=fluoxetine. HYP=hypericum. PAR=paroxetine. PL=placebo. PSYPL=psychological placebo. SER=sertraline. VEN=venlafaxine. WL=waiting list. Table 1 General characteristics of eligible studies All trials (n=64) Trials eligible for network meta-analysis (n=54) Eligible patients 7302 7014 Sample size 66 (31–159) 81 (40–168) Eligible arms 148 127 Number of arms Two 51 (80%) 42 (78%) Three 6 (9%) 5 (9%) Four 7 (11%) 7 (13%) Year of publication 1980–90 10 (16%) 4 (7%) 1991–2000 27 (42%) 23 (43%) 2001–12 27 (42%) 27 (50%) Type of intervention Medication only 38 (59%) 33 (61%) Psychotherapy only 18 (28%) 15 (28%) Both 8 (13%) 6 (11%) Duration (weeks) 12 (10–12) 12 (10–12) Continent North America 30 (47%) 26 (48%) Europe 19 (30%) 14 (26%) Asia 6 (9%) 6 (11%) Australia 3 (5%) 2 (4%) South America 3 (5%) 3 (6%) Multiple 3 (5%) 3 (6%) Characteristics of included patients Age (years) 36 (33–37) 36 (33–37) Disease severity (YBOCS score) NA 25 (24–26) Comorbid depression 27 (42%) 19 (35%) Pharmaceutical industry sponsorship* Yes 28/46 (61%) 25/39 (64%) No 15/46 (33%) 12/39 (31%) Unclear 3/46 (7%) 2/39 (5%) Allowed patients on antidepressant medication† Yes 13/18 (72%) 12/15 (80%) No 4/18 (22%) 2/15 (13%) Unclear 1/18 (6%) 1/15 (7%) Data are n, median (IQR), n (%), or n/N (%). YBOCS=Yale-Brown Obsessive Compulsive Scale. NA=Not applicable.

Yes 28/46 (61%) 25/39 (64%) No 15/46 (33%) 12/39 (31%) Unclear 3/46 (7%) 2/39 (5%) Allowed patients on antidepressant medication† Yes 13/18 (72%) 12/15 (80%) No 4/18 (22%) 2/15 (13%) Unclear 1/18 (6%) 1/15 (7%) Data are n, median (IQR), n (%), or n/N (%). YBOCS=Yale-Brown Obsessive Compulsive Scale. NA=Not applicable. * For pharmacological trials. † For psychotherapeutic trials. Table 2 Treatment efficacy compared with drug placebo

Yes 28/46 (61%) 25/39 (64%) No 15/46 (33%) 12/39 (31%) Unclear 3/46 (7%) 2/39 (5%) Allowed patients on antidepressant medication† Yes 13/18 (72%) 12/15 (80%) No 4/18 (22%) 2/15 (13%) Unclear 1/18 (6%) 1/15 (7%) Data are n, median (IQR), n (%), or n/N (%). YBOCS=Yale-Brown Obsessive Compulsive Scale. NA=Not applicable. * For pharmacological trials. † For psychotherapeutic trials. Table 2 Treatment efficacy compared with drug placebo Number of trials (n=54)* Number of patients (n=6652)* Mean YBOCS difference Full network (n=54) Excluding waiting list controlled trials (n=48) Drug placebo 23 1515 Reference Reference Waiting list 6 97 5·62 (0·91 to 10·26) NA Psychological placebo† 6 196 −4·15 (−8·65 to 0·49) −1·90 (−5·62 to 1·91) SSRIs (class effect) 37 3158 −3·49 (−5·12 to −1·81) −3·62 (−4·89 to −2·34) Fluoxetine 6 633 −3·46 (−5·27 to −1·58) −3·67 (−5·13 to −2·26) Fluvoxamine 13 521 −3·60 (−5·29 to −1·95) −3·66 (−4·96 to −2·37) Paroxetine 8 902 −3·42 (−5·10 to −1·61) −3·51 (−4·81 to −2·14) Sertraline 7 565 −3·50 (−5·30 to −1·63) −3·68 (−5·14 to −2·30) Citalopram 2 311 −3·49 (−5·62 to −1·31) −3·60 (−5·25 to −1·91) Escitalopram 1 226 −3·48 (−5·61 to −1·23) −3·59 (−5·25 to −1·86) Venlafaxine 2 98 −3·22 (−8·26 to 1·88) −3·21 (−7·01 to 0·69) Clomipramine 13 831 −4·72 (−6·85 to −2·60) −4·66 (−6·26 to −3·05) BT† 11 287 −14·48 (−18·61 to −10·23) −10·41 (−14·04 to −6·77) CBT† 9 231 −5·37 (−9·10 to −1·63) −7·98 (−11·02 to −4·93) Cognitive therapy† 6 172 −13·36 (−18·40 to −8·21) −9·45 (−13·76 to −5·19) Hypericum 1 30 −0·15 (−7·46 to 7·12) −0·13 (−5·93 to 5·68) CBT and fluvoxamine 1 6 −7·50 (−13·89 to −1·17) −8·81 (−13·75 to −3·88) BT and clomipramine 1 31 −12·97 (−19·18 to −6·74) −11·68 (−16·73 to −6·65) Data in parentheses are 95% credible intervals. YBOCS=Yale-Brown Obsessive Compulsive Scale. BT=behavioural therapy. CBT=cognitive behavioural therapy. NA=not applicable.

8) CBT and fluvoxamine 1 6 −7·50 (−13·89 to −1·17) −8·81 (−13·75 to −3·88) BT and clomipramine 1 31 −12·97 (−19·18 to −6·74) −11·68 (−16·73 to −6·65) Data in parentheses are 95% credible intervals. YBOCS=Yale-Brown Obsessive Compulsive Scale. BT=behavioural therapy. CBT=cognitive behavioural therapy. NA=not applicable. * Individual trials could be included in more than one treatment category. † Several patients randomly allocated into these psychotherapeutic interventions were allowed to take stable doses of antidepressants and remain on the same dose without further adjustments. Table 3 Efficacy of psychological and pharmacological interventions compared with SSRIs Mean YBOCS difference in full network (n=54) Mean YBOCS difference excluding waiting list controlled trials (n=48) SSRIs (class effect) Reference Reference Clomipramine −1·23 (−3·41 to 0·94) −1·05 (−2·73 to 0·63) BT* −10·99 (−15·14 to −6·75) −6·79 (−10·44 to −3·11) CBT* −1·88 (−5·52 to 1·76) −4·36 (−7·34 to −1·40) Cognitive therapy* −9·87 (−14·91 to −4·74) −5·83 (−10·17 to −1·51) CBT and fluvoxamine −4·03 (−10·36 to 2·21) −5·19 (−10·09 to −0·33) BT and clomipramine −9·48 (−15·78 to −3·14) −8·01 (−13·18 to −2·95) Data in parentheses are 95% credible intervals. YBOCS=Yale-Brown Obsessive Compulsive Scale. BT=behavioural therapy. CBT=cognitive behavioural therapy. * Several patients randomly allocated into these psychotherapeutic interventions were allowed to take stable doses of antidepressants and remain on the same dose without further adjustments.

Introduction 240 million people have a serious mental illness, with a broadly similar distribution worldwide.1 Serious mental illness is defined as a diagnosis of mental illness (eg, schizophrenia and schizoaffective disorders, bipolar disorder, or psychosis) that is persistent, disabling, and requiring specialised psychiatric treatment as an outpatient or inpatient admission.1 The point prevalence of serious mental illness is 4·6 cases per 1000 people, and 4·0% of people have a serious mental illness at some point during their life.1 With the increasing evidence that people with serious mental illness have significant health inequalities,2 increasing prominence has been given to physical health screening, health education, and improving access to treatment in primary and secondary care. However, sexual health needs, including screening for and prevention of sexually transmitted infections and blood-borne viruses, are neglected in this population. Of particular concern is the higher risk of blood-borne virus infections (HIV, hepatitis B virus, and hepatitis C virus), shown by prevalence studies done over the past 30 years.3, 4 These viruses are transmitted by parenteral contact with contaminated body fluids (blood and blood products; contaminated instruments and needles; semen and vaginal fluids). Transmission can also occur through unprotected sex (anal, vaginal, or oral), vertical transmission from mother to baby, and by sharing drug injecting equipment.

re transmitted by parenteral contact with contaminated body fluids (blood and blood products; contaminated instruments and needles; semen and vaginal fluids). Transmission can also occur through unprotected sex (anal, vaginal, or oral), vertical transmission from mother to baby, and by sharing drug injecting equipment. HIV, hepatitis B, and hepatitis C are serious infections, but can be treated. The prognosis is much improved by earlier detection and treatment. Prevalence studies3, 5, 6 have shown that serious mental illness is a risk factor for blood-borne virus infection. Many people with serious mental health problems engage in behaviours that increase their risk of infection with blood-borne viruses, including unprotected sex with multiple partners, sex work (or sex trading—performing sexual acts in exchange for a commodity), and intravenous drug use (or having a sexual partner who is an injecting drug user). Further risk can result from hypersexuality during an acute phase of mental illness, as well as co-occurring substance misuse problems that can lead to sexual risks while intoxicated. Finally, people with serious mental illness who live in shared accommodation might share personal equipment (eg, razors, toothbrushes), which might increase the risk of hepatitis B and hepatitis C transmission.

l illness, as well as co-occurring substance misuse problems that can lead to sexual risks while intoxicated. Finally, people with serious mental illness who live in shared accommodation might share personal equipment (eg, razors, toothbrushes), which might increase the risk of hepatitis B and hepatitis C transmission. In the UK, 93 000 men and 40 000 women are infected with HIV, and in 2013, there were 6000 new cases of HIV.7 In 2012, the incidence of hepatitis B in England was 1·4 cases per 100 000 people per year and prevalence was 0·1–0·5%.8 In 2009, the overall incidence of reported acute hepatitis B in the USA9 was 1·5 per 100 000 people per year and 800 000–1·4 million people in the USA have chronic hepatitis B virus infection. 786 000 people worldwide die from hepatitis B virus-related liver disease each year.10, 11 About 3% of the world's population are infected with hepatitis C virus, with about 4 million carriers in Europe alone12 and 214 000 in the UK.10 There were 17 000 new cases of hepatitis C in the USA in 2007, with 3·2 million people infected in total. Although previous reviews of blood-borne infection in people with serious mental illness have been published,3, 5, 6 no systematic reviews have been done and the reviews rarely reported the rate of HIV, hepatitis B, and hepatitis C.13 We did a systematic review and meta-analysis of prevalence studies to understand the global prevalence of HIV, hepatitis B virus, and hepatitis C virus in people with serious mental illness.

hed,3, 5, 6 no systematic reviews have been done and the reviews rarely reported the rate of HIV, hepatitis B, and hepatitis C.13 We did a systematic review and meta-analysis of prevalence studies to understand the global prevalence of HIV, hepatitis B virus, and hepatitis C virus in people with serious mental illness. Methods Search strategy We searched the Cochrane Library, Medline, Embase, PsycInfo, CINAHL, and DARE for studies published in English between Jan 1, 1980, and June 5, 2012, with the terms “hepatitis C”, ‘HCV’, “hepatitis B”, ‘HBV’, “HIV”, “human immunodeficiency virus”, “blood borne virus” cross-referenced with “bipolar”, “psychiatr*”, “schizophreni*”, “psychosis”, “schizoaffective”, “mental patient*”, “mental illness”, and “mental disorder*”. We also included eligible studies cited in reports identified by our database search. We repeated the search for June 5, 2012, to Jan 1, 2015, and identified two more papers. Data collection We systematically searched the scientific literature for observational cross-sectional studies that reported the seroprevalence of HIV, hepatitis B virus, or hepatitis C virus according to opt-in, opt-out, or anonymous unlinked blood or other bodily fluids research methods, in people aged older than 15 years, diagnosed with serious mental illness, and treated in a psychiatric setting. We excluded studies in which prevalence data were only obtained from case notes or only from self-report (not independently verified by testing). We did not include grey literature.

fluids research methods, in people aged older than 15 years, diagnosed with serious mental illness, and treated in a psychiatric setting. We excluded studies in which prevalence data were only obtained from case notes or only from self-report (not independently verified by testing). We did not include grey literature. After removing duplicates, SB screened the titles and abstracts using the eligibility criteria, with independent verification by EH and FM. For studies deemed suitable, we obtained the full text and they were again scrutinised against the eligibility criteria by SB and verified independently by EH and FM. Reports about which there was uncertainty were discussed by FM and EH with SB until a consensus about eligibility was achieved. We extracted data from eligible full-text articles including study characteristics, study date, publication date, location, diagnostic criteria, demographics (age, sex, ethnicity), consent, consent rate, ethics approval, post-test treatment, sample size, testing procedure, and prevalence.

bout eligibility was achieved. We extracted data from eligible full-text articles including study characteristics, study date, publication date, location, diagnostic criteria, demographics (age, sex, ethnicity), consent, consent rate, ethics approval, post-test treatment, sample size, testing procedure, and prevalence. We used the Quality Assessment Tool for Systematic Reviews of Observational Studies to assess the quality of the data.14 This instrument is reliable compared with other quality assessment tools.14 We modified the tool and each report was scored as follows: whether participants were clearly defined as representing the serious mental illness population (yes=1, no=0); participation rate (response rate >60%=1, response rate ≤60% or not reported=0); whether investigators controlled for confounding (eg, stratification, matching, restriction, adjustment) when analysing associations (controlled=1, only descriptive=0); and sample size (≥200 participants=1, <200 participants=0).

=1, no=0); participation rate (response rate >60%=1, response rate ≤60% or not reported=0); whether investigators controlled for confounding (eg, stratification, matching, restriction, adjustment) when analysing associations (controlled=1, only descriptive=0); and sample size (≥200 participants=1, <200 participants=0). Data analysis We did a meta-analysis to calculate combined estimates and 95% CIs for each continent separately. We did logistic regression to allow for the difficulties caused by proportions being unable to have values less than 0. We assumed random effects because there was clear clinical heterogeneity between the populations sampled. We did the calculations using Comprehensive Meta-Analysis 2. We transformed logits of estimated prevalence and their 95% CIs back to percentages. We prepared forest plots using Stata (version 12). We calculated relative weights for each continent, so that the weights for each continent sum to 100. We did a sensitivity analysis relating to quality scores using Stata 13. The outcome variable was the logit-transformation of prevalence. We did two such analyses: one for all studies with quality score as a quantitative predictor, and one for all studies using quality score as a quantitative predictor and region as a qualitative predictor. The results are presented as odds ratio (OR) per unit increase in quality score with 95% CIs.

mation of prevalence. We did two such analyses: one for all studies with quality score as a quantitative predictor, and one for all studies using quality score as a quantitative predictor and region as a qualitative predictor. The results are presented as odds ratio (OR) per unit increase in quality score with 95% CIs. Results Our literature search identified 373 reports, 169 of which were duplicates (figure 1). 41 publications were excluded because the full text was not available in English, followed by another 74 that were ineligible. With the addition of two papers from an updated search, we had 91 articles for quality assessment and meta-analysis. 44 studies assessed HIV infection (table 1), including a total of 21 071 patients. The pooled prevalence of HIV was highest in Africa (19%, 95% CI 14–25) and it was 2% in Europe and 6% in the USA (Table 1, Table 2, figure 2). Few data were available from Europe, Central and South America, and Asia, and the prevalence of HIV was very poorly recorded in these regions. Only three studies were done in India,51, 53 and only one study,38 done more than 20 years ago, was from Spain.

was 2% in Europe and 6% in the USA (Table 1, Table 2, figure 2). Few data were available from Europe, Central and South America, and Asia, and the prevalence of HIV was very poorly recorded in these regions. Only three studies were done in India,51, 53 and only one study,38 done more than 20 years ago, was from Spain. 19 studies reported prevalence of hepatitis B virus,53, 57, 59, 60, 61, 62 including a total of 8163 patients with serious mental illness tested for hepatitis B virus (table 3). The pooled prevalence of hepatitis B virus was 2·2% (95% CI 0·5–9·9) in North America, and 9·7% (95% CI 0·6–15·3) in Asia (table 2, figure 3). A study from Turkey60 reported 51% hepatitis B virus prevalence with 10% HBsAg positivity indicating active infection; the virus is highly prevalent in the general population of Turkey. A study from Taiwan61 reported an 18% prevalence of HBsAg, which is consistent with the general population: hepatitis B virus infection is endemic in Taiwan, with 80–90% of adults infected.

evalence with 10% HBsAg positivity indicating active infection; the virus is highly prevalent in the general population of Turkey. A study from Taiwan61 reported an 18% prevalence of HBsAg, which is consistent with the general population: hepatitis B virus infection is endemic in Taiwan, with 80–90% of adults infected. 28 studies tested 14 888 patients with serious mental illness for hepatitis C virus (Table 2, Table 4, figure 4). The prevalence of hepatitis C in people with serious mental illness was greatest in Turkey, perhaps a result of the high prevalence in the general population in Turkey. Pooled data from 13 studies from North America gave a prevalence of 17·4% (95% CI 13·2–22·6), which is higher than in the general population, of whom roughly 1% are infected (2·7 million).88 In Asia, pooled hepatitis C virus prevalence was 4·4% (95% CI 2·8–6·9). However, these data are from large and diverse geographical areas including southeast Asia and Turkey. Most studies consisted of convenience samples of people recruited from a particular treatment setting, typically inpatient psychiatric care. Although all the studies included patients with serious mental illness, the proportions of specific diagnoses in each sample varied. We assessed the effect of study quality on virus prevalence by meta-regression on quality score, for all studies combined and adjusting for geographical region. No analysis showed a significant effect of study quality on prevalence (table 5).

mental illness, the proportions of specific diagnoses in each sample varied. We assessed the effect of study quality on virus prevalence by meta-regression on quality score, for all studies combined and adjusting for geographical region. No analysis showed a significant effect of study quality on prevalence (table 5). Most of the studies had additional data on risk factors for blood-borne viruses, such as intravenous drug use and sexual behaviour, to test associations with infection. The reporting and the nature of these risk factors varied widely. Infomation on risk factors was mainly extracted from routine clinical case notes as opposed to using a standardised risk tool.89 Because of the variability of data quality and reporting consistency, we could not calculate adjusted prevalence after controlling for these risk factors. However, three common factors seem to increase the likelihood of infection with a blood-borne virus: first, being black and female;17, 26 second, injecting drug use;32 and third, engaging in risky sexual behaviour, including not using a condom, having multiple partners, sex trading, and unprotected sex with a partner who is infected with a blood-borne virus.

crease the likelihood of infection with a blood-borne virus: first, being black and female;17, 26 second, injecting drug use;32 and third, engaging in risky sexual behaviour, including not using a condom, having multiple partners, sex trading, and unprotected sex with a partner who is infected with a blood-borne virus. Discussion Our aim was to estimate the prevalence of blood-borne infection in people with serious mental illness. Most of the studies were of moderate to low quality, and based on convenience samples drawn from treatment settings rather than representative samples. This sampling method means that the prevalence reported was possibly overestimated. However, a study in Brazil, which used a representative sample drawn from the community as well as treatment settings, still showed that blood-borne infections are common in people with serious mental illness.57 The quality of defining the sample in terms of diagnoses of mental illness varied. Many studies used case note diagnoses rather than independently verified diagnoses made with gold standard diagnostic tools. Inpatient settings are likely to treat the most acutely ill people often with complex needs and a history of substance misuse.

defining the sample in terms of diagnoses of mental illness varied. Many studies used case note diagnoses rather than independently verified diagnoses made with gold standard diagnostic tools. Inpatient settings are likely to treat the most acutely ill people often with complex needs and a history of substance misuse. The prevalences of blood-borne viruses in people with serious mental illness were consistently higher than in the general population in regions with a low prevalence of blood-borne viruses, such as North America and Europe, and on par with the general population in regions with high general prevalence such as Africa for HIV and southeast Asia for hepatitis B virus and hepatitis C virus. The estimated prevalence of HIV in people with serious mental illness in the USA was 6% (95% CI 4·3–8·3), which is considerably higher than the 0·6% of the general population of the USA who have HIV.90 However, serious mental illness might not be an isolated risk factor for blood-borne virus infection, but might be better thought of as a potentially confounded association with poor socioeconomic background, drug and alcohol misuse, sex, and ethnic origin.

The prevalences of blood-borne viruses in people with serious mental illness were consistently higher than in the general population in regions with a low prevalence of blood-borne viruses, such as North America and Europe, and on par with the general population in regions with high general prevalence such as Africa for HIV and southeast Asia for hepatitis B virus and hepatitis C virus. The estimated prevalence of HIV in people with serious mental illness in the USA was 6% (95% CI 4·3–8·3), which is considerably higher than the 0·6% of the general population of the USA who have HIV.90 However, serious mental illness might not be an isolated risk factor for blood-borne virus infection, but might be better thought of as a potentially confounded association with poor socioeconomic background, drug and alcohol misuse, sex, and ethnic origin. Three USA studies18, 21, 26 included odds ratios adjusted for risk factors and showed that they significantly increased the risk of HIV and other blood-borne viral infections. However, these studies were done in settings where dual diagnosis of substance misuse and mental illness is very common. The samples were drawn from psychiatric inpatient and outpatient services in deprived urban areas with substantial social deprivation and health inequality, especially in those of non-white ethnic backgrounds.

hese studies were done in settings where dual diagnosis of substance misuse and mental illness is very common. The samples were drawn from psychiatric inpatient and outpatient services in deprived urban areas with substantial social deprivation and health inequality, especially in those of non-white ethnic backgrounds. Several studies, from both high prevalence and low prevalence locations, individually found a positive association between sex and infection. Women had a significantly higher risk of HIV infection than did men drawn from the same populations. One explanation might be that women with serious mental illness are more likely to experience exploitation and sexual assault, as well as power differentials, making them less empowered to negotiate condom use or to refuse sex.91 By contrast, men with serious mental illness were more likely to carry hepatitis B virus or hepatitis C virus, which could be because injecting drug use is more common in men. However, the causes of these sex differences were probably multifactorial, which we could not assess because of the heterogeneity of geography, demographics, and risk factors in the studies we included.

ely to carry hepatitis B virus or hepatitis C virus, which could be because injecting drug use is more common in men. However, the causes of these sex differences were probably multifactorial, which we could not assess because of the heterogeneity of geography, demographics, and risk factors in the studies we included. Many of the studies have been done in the USA, with fewer located in other countries. Of particular note is the paucity of research in Europe, and there have been no prevalence studies done in the UK. However, two articles suggest a potential problem in the UK. A hepatitis C virus screening and referral project done by an assertive outreach mental health team92 showed more than expected infections amongst users of the service. Of 76 users, ten (13%) were hepatitis C virus positive, and almost half had a history of intravenous drug use. Another article93 reported on the acceptability and feasibility of offering testing for blood-borne viruses in psychiatric inpatient settings. The results suggest more HIV, hepatitis B virus, and hepatitis C virus in patients who participated in the study. Overall, 18% of participants had current or past exposure to a blood-borne virus, one of whom was newly diagnosed with HIV and three were newly diagnosed with hepatitis B virus. Therefore, there is an urgent need to undertake high quality epidemiological studies of blood-borne virus infections and their associated risk behaviours in the people with serious mental illness in the UK and northern Europe.

e of whom was newly diagnosed with HIV and three were newly diagnosed with hepatitis B virus. Therefore, there is an urgent need to undertake high quality epidemiological studies of blood-borne virus infections and their associated risk behaviours in the people with serious mental illness in the UK and northern Europe. Few studies systematically collected data for risk factors directly from the participants. The risk data were mainly collected from case notes and routine clinical record systems. Sexual and drug use behaviours are probably under-reported in case notes, because there is evidence that mental health services do not consistently assess these behaviours in routine care.94, 95 Without accurate and consistent measurement of risks, we could not calculate the effect of the risk factors as mediators of infection in this population, and have merely mentioned the factors identified by individual studies that warrant more rigorous investigation. There is a need for a prospective longitudinal study of a cohort of people with serious mental illness, which can track risk behaviour and infections powered sufficiently to identify the mediating factors between serious mental illness and blood-borne virus infection.

l studies that warrant more rigorous investigation. There is a need for a prospective longitudinal study of a cohort of people with serious mental illness, which can track risk behaviour and infections powered sufficiently to identify the mediating factors between serious mental illness and blood-borne virus infection. We included cross-sectional studies. None of the studies included a matched comparison group of people without serious mental illness. Prospective cohort studies are needed that use representative samples alongside matched controls of people without serious mental illness. Such studies are the only way to accurately test whether the prevalence of blood-borne viruses is significantly elevated in people with serious mental illness compared with the general population. Comparing the estimated prevalence with available data for the country or region is limited but it does offer some indication that prevalence is higher in people with serious mental illness. The prevalence of HIV infection in the general population is much lower in the UK96 than in the USA, and therefore the assumption is that HIV infection is less of a risk for people with serious mental illness who live in the UK. However, hepatitis C virus is prevalent in drug users in the UK,97 and there could be a risk of hepatitis C virus infection and co-infections in people with serious mental illness as a result of substance misuse.98

mption is that HIV infection is less of a risk for people with serious mental illness who live in the UK. However, hepatitis C virus is prevalent in drug users in the UK,97 and there could be a risk of hepatitis C virus infection and co-infections in people with serious mental illness as a result of substance misuse.98 This meta-analysis estimated pooled prevalence of HIV, hepatitis B virus, and hepatitis C virus in people with serious mental illness. Our review included only published work, and therefore might have missed studies yet to be reported. In addition, the search strategy included only reports published in English, which might have biased our findings towards English-speaking countries.

B virus, and hepatitis C virus in people with serious mental illness. Our review included only published work, and therefore might have missed studies yet to be reported. In addition, the search strategy included only reports published in English, which might have biased our findings towards English-speaking countries. It is unclear why sexual health has been neglected as part of the physical health agenda for people with serious mental illness. One reason might be the perception that people with serious mental illness do not engage in activities that place them at risk, such as intravenous drug use or unprotected sex. However, 30–50% of people with serious mental illness have substance misuse disorders,98, 99 and, although intravenous drug use is rare, patients might have sexual partners who inject drugs, facilitating viral transmission. Additionally, as with the population as a whole, a substantial proportion of people with serious mental illness are sexually active and see intimate relationships as an important part of their lives.5, 100 The lack of attention of policy makers and educators has led to a lack of awareness and a failure to provide people with serious mental illness with access to assessment, screening, and education for sexually transmitted infections, including blood-borne viruses.

relationships as an important part of their lives.5, 100 The lack of attention of policy makers and educators has led to a lack of awareness and a failure to provide people with serious mental illness with access to assessment, screening, and education for sexually transmitted infections, including blood-borne viruses. A qualitative study in London101 documented that most people with psychosis were engaged in seeking and forming intimate relationships. Additionally, some had negative and harmful relationship experiences, including sexual exploitation and violence, yet these issues were rarely part of their routine consultation with their health-care providers. This lack of attention to sexual health and safety has also been reported in a review,91 which found that although women with serious mental illness were twice as likely to be exposed to severe domestic violence compared with women in the general population, these incidents were rarely detected by the health-care services they attended. A survey of psychiatrists in a Sydney, Australia, mental health service102 found poor knowledge of hepatitis C virus, and clinicians perceived their patients to be at lower risk than prevalence studies suggest.3 A survey of mental health staff at a London NHS service94 also showed that workers underestimated the risk of HIV in people with schizophrenia. Although they reported feeling comfortable discussing sexual health, this rarely happened in practice.94 In addition, a qualitative study of Australian mental health nurses95 showed that discussions of sex and sexuality were generally avoided.

ed that workers underestimated the risk of HIV in people with schizophrenia. Although they reported feeling comfortable discussing sexual health, this rarely happened in practice.94 In addition, a qualitative study of Australian mental health nurses95 showed that discussions of sex and sexuality were generally avoided. In summary, we show the high prevalence of blood-borne infections in people with serious mental illness, but more importantly we document the paucity of data on this topic. Although the physical health inequalities of people with serious mental illness have been identified and health policy is developing to ensure that these inequalities are addressed, little attention has been given to the sexual health and specifically risk factors facilitating transmission of blood-borne viruses in people with serious mental illness in the UK and worldwide. There is an urgent need for further robust epidemiological research using representative samples of people with serious mental illness to assess the relationship between lifestyle behaviour and risk of infections to more fully understand the relationship between serious mental illness and viral infection, and to inform preventive strategies in this population. This online publication has been corrected. The corrected version first appeared at thelancet.com/psychiatry on January 6, 2016 Acknowledgments This Article was funded by the Wellcome Trust (reference 097829) through the Centre for Chronic Diseases and Disorders at the University of York.

In summary, we show the high prevalence of blood-borne infections in people with serious mental illness, but more importantly we document the paucity of data on this topic. Although the physical health inequalities of people with serious mental illness have been identified and health policy is developing to ensure that these inequalities are addressed, little attention has been given to the sexual health and specifically risk factors facilitating transmission of blood-borne viruses in people with serious mental illness in the UK and worldwide. There is an urgent need for further robust epidemiological research using representative samples of people with serious mental illness to assess the relationship between lifestyle behaviour and risk of infections to more fully understand the relationship between serious mental illness and viral infection, and to inform preventive strategies in this population. This online publication has been corrected. The corrected version first appeared at thelancet.com/psychiatry on January 6, 2016 Acknowledgments This Article was funded by the Wellcome Trust (reference 097829) through the Centre for Chronic Diseases and Disorders at the University of York. Contributors EH, FM, and SG designed the study, SB did the literature search with support from EH and FM. MB analysed the data. FM, EH, and SG interpreted the data. EH, SB, and FM wrote the report, with input from SG and MB. Declaration of interests We declare no competing interests. Figure 1 Study selection Figure 2 Prevalence of HIV in people with serious mental illness

Contributors EH, FM, and SG designed the study, SB did the literature search with support from EH and FM. MB analysed the data. FM, EH, and SG interpreted the data. EH, SB, and FM wrote the report, with input from SG and MB. Declaration of interests We declare no competing interests. Figure 1 Study selection Figure 2 Prevalence of HIV in people with serious mental illness Figure 3 Prevalence of hepatitis B virus in people with serious mental illness Figure 4 Prevalence of hepatitis C virus in people with serious mental illness Table 1 Included studies of HIV in people with serious mental illness

Declaration of interests We declare no competing interests. Figure 1 Study selection Figure 2 Prevalence of HIV in people with serious mental illness Figure 3 Prevalence of hepatitis B virus in people with serious mental illness Figure 4 Prevalence of hepatitis C virus in people with serious mental illness Table 1 Included studies of HIV in people with serious mental illness Date Location Number of participants Prevalence of HIV (%) Quality score North America Clair et al15 1989 USA 1496 0·24 1 Hatem et al16 1990 USA 163 1·8 1 Cournos et al17 1991 USA 451 5·5 3 Volavka et al18 1991 USA 515 8·9 3 Lee et al19 1992 USA 135 16·3 0 Sacks et al20 1992 USA 87 3·4 1 Empfield et al21 1993 USA 203 6·4 2 Meyer et al22 1993 USA 199 4 2 Meyer et al23 1993 USA 87 5·75 1 Susser et al24 1993 USA 62 19·4 1 Stewart et al25 1994 USA 533 5·8 3 Cournos et al26 1994 USA 971 5·2 3 Silberstein et al27 1994 USA 118 22·9 2 Schwartz-Watts et al28 1995 USA 220 5·50 3 Doyle et al29 1997 USA 246 0 1 Krakow et al30 1998 USA 113 19 0 Klinkenberg et al31 2003 USA 204 6·2 2 Rosenberg et al32 2005 USA 755 3 3 Rothbard et al33 2009 USA 588 10 2 Jackson-Malik et al34 2011 USA 64 3·1 1 Himelhoch et al35 2011 USA 153 6·1 2 Europe Porta et al36 1990 Spain 139 0 0 Naber et al37 1994 Germany 623 4·8 1 Ayuso-Mateos et al38 1997 Spain 390 5·1 3 De Hert et al39 2009 Belgium 595 0·5 2 Kakisi et al40 2009 Greece 805 1 1 Africa Acuda et al41 1996 Zimbabwe 143 23·8 2 Mashaphu et al42 2007 South Africa 63 23·8 1 Collins et al43 2009 South Africa 151 26·5 3 Singh et al44 2009 South Africa 206 29·1 3 Omoregie et al45 2009 Nigeria 121 15·5 1 Henning et al46 2011 South Africa 195 12 2 Maling et al47 2011 Uganda 272 18·4 4 Lundberg et al48 2013 Uganda 602 11·3 4 Asia Dasananjali49 1994 Thailand 325 1·85 2 Chandra et al50 2003 India 2283 2·11 1 Tharyan et al51 2003 India 1160 1·03 3 Chen52 1994 Taiwan 834 0 2 Carey et al53 2007 India 948 1·7 4 Central and South America Rodgers-Johnson et al54 1996 Jamaica 201 2·5 2 Hutchinson et al55 1999 Trinidad and Tobago 1227 6·9 1 Alvarado-Esquivel et al56 2008 Mexico 105 0·9 0 Guimarães et al57 2009 Brazil 2238 0·8 3 Gibson et al58 2010 West Indies 82 7·3 1 Table 2 Pooled prevalence in people with serious mental illness

entral and South America Rodgers-Johnson et al54 1996 Jamaica 201 2·5 2 Hutchinson et al55 1999 Trinidad and Tobago 1227 6·9 1 Alvarado-Esquivel et al56 2008 Mexico 105 0·9 0 Guimarães et al57 2009 Brazil 2238 0·8 3 Gibson et al58 2010 West Indies 82 7·3 1 Table 2 Pooled prevalence in people with serious mental illness HIV Hepatitis B virus Hepatitis C virus Studies (n) Prevalence (95% CI) Studies (n) Prevalence (95% CI) Studies (n) Prevalence (95% CI) North America 21 6·0% (4·3–8·3) 2 2·2% (0·5–9·9) 13 17·4% (13·2–22·6) Europe 5 1·9% (0·8–4·8) 4 2·7% (1·8–3·9) 6 4·9% (3·0–7·9) Oceania 0 .. 0 .. 1 3·1% (1·0–9·3) Africa 8 19·2% (14·4–25·2) 0 .. 0 .. Asia 5 1·5% (1·0–2·4) 10 9·7% (0·6–15·3) 7 4·4% (2·8–6·9) Central and South America 5 2·7% (0·8–8·2) 3 2·6% (1·0–6·1) 2 3·0% (1·8–5·0) Table 3 Included studies of hepatitis B

·4% (13·2–22·6) Europe 5 1·9% (0·8–4·8) 4 2·7% (1·8–3·9) 6 4·9% (3·0–7·9) Oceania 0 .. 0 .. 1 3·1% (1·0–9·3) Africa 8 19·2% (14·4–25·2) 0 .. 0 .. Asia 5 1·5% (1·0–2·4) 10 9·7% (0·6–15·3) 7 4·4% (2·8–6·9) Central and South America 5 2·7% (0·8–8·2) 3 2·6% (1·0–6·1) 2 3·0% (1·8–5·0) Table 3 Included studies of hepatitis B Date Location Number of patients HbAg (%) Quality score North America Tabibian et al63 2008 USA 129 0·78 1 Wise et al64 2012 USA 115 4 2 Europe Gmelin et al65 1982 Germany 714 2·38 1 Porta et al36 1990 Spain 139 2·3 0 Di Nardo et al66 1995 Italy 206 4·8 1 Kakisi et al40 2009 Greece 805 2 1 Asia Tey et al67 1987 Singapore 71 12·7 1 Chaudhury et al68 1993 India 60 10 1 Chang et al61 1993 Taiwan 780 18·1 1 Chaudhury et al69 1994 India 100 11 1 Kimhi et al70 1997 Israel 121 32 0 Said et al71 2001 Jordan 192 7·29 1 Kuloglu et al60 2006 Turkey 255 10·1 3 Carey et al53 2007 India 948 3 4 Mamani et al72 2009 Iran 170 1·2 1 Hung et al12 2012 Taiwan 588 10·4 3 Central and South America de Souza et al73 2003 Brazil 433 1·6 3 Alvarado Esquivel et al59 2005 Mexico 99 7·1 3 Guimarães et al57 2009 Brazil 2238 1·6 3 Table 4 Included studies of hepatitis C virus in people with serious mental illness

ndia 948 3 4 Mamani et al72 2009 Iran 170 1·2 1 Hung et al12 2012 Taiwan 588 10·4 3 Central and South America de Souza et al73 2003 Brazil 433 1·6 3 Alvarado Esquivel et al59 2005 Mexico 99 7·1 3 Guimarães et al57 2009 Brazil 2238 1·6 3 Table 4 Included studies of hepatitis C virus in people with serious mental illness Year Location Number of patients Hepatitis C virus (%) Quality score North America Al Jurdi et al74 2003 USA 238 16 3 Klinkenberg et al31 2003 USA 204 30 2 Osher et al75 2003 USA 668 18 3 Dinwiddie et al76 2003 USA 1556 8·5 3 Butterfield et al77 2003 USA 376 18·9 2 Rosenberg et al32 2005 USA 755 14 3 Freudenreich et al78 2007 USA 98 8·2 2 Tabibian et al63 2008 USA 129 38 1 Goldberg et al79 2008 USA 100 31 1 Matthews et al80 2008 USA 112 12 0 Rothbard et al33 2009 USA 588 21 2 Sockalingam et al81 2010 Canada 110 2·7 1 Himelhoch et al35 2011 USA 153 24·8 1 Europe Di Nardo et al66 1995 Italy 206 10·7 1 Cividini et al82 1997 Italy 1180 6·7 2 Stroffolini et al83 2003 Italy 526 5·1 1 Raja et al84 2006 Italy 1492 2·8 3 De Hert et al39 2009 Belgium 595 0·7 2 Kakisi et al40 2009 Greece 805 9 1 Central and South America Alvarado-Esquivel et al56 2008 Mexico 99 4·8 0 Guimarães et al57 2009 Brazil 2238 2·63 3 Oceania Gunewardene et al85 2010 Australia 95 3·1 1 Asia Chang et al61 1993 Taiwan 780 6·8 3 Kimhi et al70 1997 Israel 121 4·13 0 Sawayama et al86 2000 Japan 196 10·2 2 Nakamura et al87 2004 Japan 455 6·15 3 Kuloglu et al60 2006 Turkey 255 2·7 3 Mamani et al72 2009 Iran 170 1·8 1 Hung et al12 2012 Taiwan 588 1·9 3 Table 5 Sensitivity analysis

l85 2010 Australia 95 3·1 1 Asia Chang et al61 1993 Taiwan 780 6·8 3 Kimhi et al70 1997 Israel 121 4·13 0 Sawayama et al86 2000 Japan 196 10·2 2 Nakamura et al87 2004 Japan 455 6·15 3 Kuloglu et al60 2006 Turkey 255 2·7 3 Mamani et al72 2009 Iran 170 1·8 1 Hung et al12 2012 Taiwan 588 1·9 3 Table 5 Sensitivity analysis Adjustment Odds ratio (95% CI) p value HIV None 1·00 (0·68–1·48) 0·99 HIV Region 0·90 (0·67–1·21) 0·49 Hepatitis B None 0·84 (0·46–1·55) 0·55 Hepatitis B Region 0·69 (0·44–1·10) 0·11 Hepatitis C None 0·92 (0·61–1·40) 0·69 Hepatitis C Region 0·86 (0·63–1·19) 0·36

Introduction Depression is common, familial, often recurrent, and one of the world's leading causes of disability burden.1 Offspring of parents with depression are at three-to-four times higher risk of developing a wide range of mental health disorders than are offspring of non-depressed parents, with adverse health, educational and social outcomes, including increased risk of suicide.1, 2 Mental health disorders are a global problem in children and adolescents,3 in whom they show persistence into adulthood and have lifelong consequences.4 Effective prevention of mental health disorders in this identifiable high-risk group is therefore important.5 One approach to improving outcomes is to ameliorate the risk to which young people are exposed. Trials of treatment of adult depression show potential benefits for offspring mental health,6 although not all parents respond to treatment and, even when they do, positive mental health effects on children are not always seen.7 An alternative approach is to provide preventive interventions for at-risk adolescent offspring themselves.8 Typically, prevention strategies are informed by observational research on risks and adverse outcomes.9 However, many at-risk offspring show remarkably positive mental health outcomes without intervention.10, 11 Understanding what explains young people's resilience in the context of familial risk is important for identifying additional prevention targets.12 Research in context Evidence before this study

Typically, prevention strategies are informed by observational research on risks and adverse outcomes.9 However, many at-risk offspring show remarkably positive mental health outcomes without intervention.10, 11 Understanding what explains young people's resilience in the context of familial risk is important for identifying additional prevention targets.12 Research in context Evidence before this study Previous research has typically examined risk mechanisms that explain increased psychopathology in children of parents with depression as compared to non-depressed parents. We undertook a systematic search on Dec 5, 2014, using the Web of Science database and the search terms “resilience” and (“maternal depression” or “paternal depression”, or “parental depression”) and (“child” or “adolescent”). We identified additional papers by checking citations and cited papers. We found only three studies on adolescent offspring of parents with depression that used longitudinal designs to test predictors of mental health disorder absence. These studies differed from the present study in that variation in severity of parental mental illness risk exposure was not taken into account, nor was type of offspring outcome examined. Added value of this study

Previous research has typically examined risk mechanisms that explain increased psychopathology in children of parents with depression as compared to non-depressed parents. We undertook a systematic search on Dec 5, 2014, using the Web of Science database and the search terms “resilience” and (“maternal depression” or “paternal depression”, or “parental depression”) and (“child” or “adolescent”). We identified additional papers by checking citations and cited papers. We found only three studies on adolescent offspring of parents with depression that used longitudinal designs to test predictors of mental health disorder absence. These studies differed from the present study in that variation in severity of parental mental illness risk exposure was not taken into account, nor was type of offspring outcome examined. Added value of this study The present study focused on factors that account for resilience in high-risk adolescents, and to our knowledge is the first to show that child, family, social, and lifestyle factors together contribute to adolescent mental health resilience. Crucially, these protective effects are not merely markers of parental depression severity—a caveat that has not been accounted for in previous studies. The study findings are also novel in that they reveal different contributors to mood and behavioural resilience. Two important findings are that emotional support from the healthy co-parent and youth physical exercise contribute to adolescent mood-related resilience even when parental depression severity is taken into account.

y findings are also novel in that they reveal different contributors to mood and behavioural resilience. Two important findings are that emotional support from the healthy co-parent and youth physical exercise contribute to adolescent mood-related resilience even when parental depression severity is taken into account. Implications of all the available evidence Adolescent mental health problems are common among offspring of parents who have recurrent depression, but our findings highlight that adolescent mental health problems in those at familial risk are not inevitable, and that interventions aimed at enhancing resilience will need to target and change multiple social and lifestyle factors. Evidence supports multimodal interventions for at-risk adolescents. Extension of family-focused aspects of interventions to include both parents may be of particular benefit. Providing information and support that encourages healthy lifestyles (including frequent exercise) and that encourages young people to capitalise on friendship networks also seem likely to be beneficial for maintaining good mental health.

ed aspects of interventions to include both parents may be of particular benefit. Providing information and support that encourages healthy lifestyles (including frequent exercise) and that encourages young people to capitalise on friendship networks also seem likely to be beneficial for maintaining good mental health. Mental health resilience has been conceptualised in different ways.13 Most studies have compared subgroups of at-risk individuals who either do or do not have mental health problems.11, 14, 15 An alternative approach operationalises resilience as showing lower symptom scores than those predicted by measures of risk. This approach has several important advantages as discussed by others.13, 16, 17 First, resilience is defined in terms of better than expected—rather than simply good—adaptation. This ensures that identified protective factors are not simply markers of lower severity of risk. Second, it permits distinction of protective factors for different mental health outcomes (eg, mood as well as behavioural). To our knowledge, no studies have used this approach to investigate resilience in offspring of parents with depression to date.

rotective factors are not simply markers of lower severity of risk. Second, it permits distinction of protective factors for different mental health outcomes (eg, mood as well as behavioural). To our knowledge, no studies have used this approach to investigate resilience in offspring of parents with depression to date. Family, social, and cognitive factors suggested to be associated with mental health resilience in young people include good-quality relationships with the parent with depression, support provided by other family members and friends, and adolescents' own self-appraisal.10, 11, 12, 13, 14, 15 Some national guidelines (eg, such as those of the National Institute for Health and Care Excellence19) also highlight potential protective effects of frequent physical exercise for depression.18, 19 However, we do not know whether these factors simply reflect lower levels of familial risk exposure. Also, although promoting mental health resilience in young people at familial risk is an internationally recognised priority,5 is it enough for prevention programmes to focus on a single domain of functioning? This study examines adolescent offspring of parents with recurrent depression, studied prospectively in adolescence. We first examined the subgroup of at-risk individuals who exhibited no mental health problems for the duration of the study; then we used a residual scores method to assess family, social, and cognitive predictors of better-than-expected mood and behaviour outcomes beyond that accounted for by severity of parental depression.

t examined the subgroup of at-risk individuals who exhibited no mental health problems for the duration of the study; then we used a residual scores method to assess family, social, and cognitive predictors of better-than-expected mood and behaviour outcomes beyond that accounted for by severity of parental depression. Methods Study design and participants The Early Prediction of Adolescent Depression study (EPAD) is a prospective longitudinal study of offspring of parents with recurrent depression.20 Families were recruited primarily from general practices across South Wales, UK. The presence of at least two previous episodes of parent DSM-IV major depressive disorder was confirmed at baseline interview. The youngest child in the age range 9–17 years was selected for inclusion. All selected children were biologically related to and living with the affected parent. We excluded offspring with an intelligence quotient lower than 50, children with serious physical illnesses, and parents with psychosis, bipolar disorder, or mania or hypomania. Parents and young people provided written informed consent (≥16 years of age) or assent (<16 years). The Multicentre Research Ethics Committee for Wales provided ethical approval.

gence quotient lower than 50, children with serious physical illnesses, and parents with psychosis, bipolar disorder, or mania or hypomania. Parents and young people provided written informed consent (≥16 years of age) or assent (<16 years). The Multicentre Research Ethics Committee for Wales provided ethical approval. Procedures Assessments were undertaken on three occasions (referred to as waves), 12–18 months apart, over a 4-year period (2007–11). Trained, supervised research psychologists assessed families at home using semi-structured research diagnostic interviews and self-report questionnaires to measure risk exposure, mental health of adolescents, and family, social, cognitive, and health behaviour protective factors. For risk exposure, we looked at the severity and course of parental depression. Parents were interviewed at each of the three waves with the Schedules for Clinical Assessment in Neuropsychiatry (SCAN)21 to assess past month episodes of depression and to collect information about additional episodes between assessments. Interviews at baseline ascertained parents' age at first episode, periods of hospital admission for depression, impairment of the worst two episodes using Global Assessment of Functioning scores,22 and additional family history of depression (in adolescents' siblings, parents, and grandparents). We also retrospectively obtained information about depression during pregnancy and the postnatal period (up to 1 year after birth) with the index child from the mothers in the sample.

l Assessment of Functioning scores,22 and additional family history of depression (in adolescents' siblings, parents, and grandparents). We also retrospectively obtained information about depression during pregnancy and the postnatal period (up to 1 year after birth) with the index child from the mothers in the sample. To assess the mental health of the adolescent, we used the Child and Adolescent Psychiatric Assessment (CAPA) interview and a questionnaire. The CAPA is a well validated semi-structured diagnostic interview.23 It assesses psychiatric symptoms and disorders over the preceding 3 months according to DSM-IV criteria.22 Sections on “mood” (depressive disorders), “behaviour” (oppositional defiant disorder and conduct disorder), anxiety disorders, attention-deficit hyperactivity disorder, bipolar disorder, cyclothymia, and eating disorders were completed independently with parents and young people at each wave. Two child and adolescent psychiatrists reviewed the diagnosed and sub-threshold cases. CAPA interviews also generated mood and behaviour disorder symptom totals. Inter-rater reliabilities for symptoms were excellent (average κ=0·94). Suicidality or self-harm was coded if parents or adolescents reported suicidal ideation or behaviour in the CAPA or endorsed the item “thought about killing self” on a well validated child and adolescent depression measure, the Mood and Feelings Questionnaire.24

liabilities for symptoms were excellent (average κ=0·94). Suicidality or self-harm was coded if parents or adolescents reported suicidal ideation or behaviour in the CAPA or endorsed the item “thought about killing self” on a well validated child and adolescent depression measure, the Mood and Feelings Questionnaire.24 Parents and adolescents also completed the well validated Strengths and Difficulties Questionnaire (SDQ),25 a 25-item screen for common emotional and behavioural problems allowing a direct comparison with UK population norms.26 The family, social, cognitive, and health behaviour protective factors were assessed at baseline unless otherwise specified. For family functioning, we looked at four measurements. First, we assessed index parent-rated warmth towards the adolescent using the Iowa Youth and Families Project (IYFP) parental warmth subscale (six items, range 6–42, α=0·93).27 Second, we recorded “five minute expressed emotion” interviews, and trained researchers coded positive expressed emotions of index parents about adolescents according to tone and content of speech samples (range 0–5), as previously validated.28 We imputed partial missing data for this particular task at baseline using expressed emotion data at first follow-up. Third, adolescents rated co-parent emotional support using the interviewer-administered Perceived Social Support scale (eg, “this person listens if I need to talk about worries”; range 0–6; α=0·95).29 Finally, we measured sibling warmth using the IYFP family interaction rating scales (six items, range 6–30, α=0·84).27

Third, adolescents rated co-parent emotional support using the interviewer-administered Perceived Social Support scale (eg, “this person listens if I need to talk about worries”; range 0–6; α=0·95).29 Finally, we measured sibling warmth using the IYFP family interaction rating scales (six items, range 6–30, α=0·84).27 For social relationships and friendships, we looked at four measurements. First, we used the parent-rated five-item SDQ peer subscale, which is a measure commonly used in epidemiological surveys to assess positive and negative aspects of young people's social relationships (eg, “liked by other children”, “has at least one good friend”).25 Negative items were reverse coded. Higher total scores indicated better quality social relationships (range 0–10; α=0·68). Second, we used the adolescent-rated SDQ peer subscale coded in the same way (range 0–10, α=0·55). Third, we assessed adolescent-perceived friendship quality using a ten-item questionnaire assessing social esteem and peer inclusion (eg, “other children think I am a nice person”, “other children want to be my friend”; α=0·83).30 Finally, parents also reported attendance at clubs or other organised out-of-school activities (at least monthly).

cent-perceived friendship quality using a ten-item questionnaire assessing social esteem and peer inclusion (eg, “other children think I am a nice person”, “other children want to be my friend”; α=0·83).30 Finally, parents also reported attendance at clubs or other organised out-of-school activities (at least monthly). For adolescent exercise, we assessed the frequency of exercise using an adolescent questionnaire at baseline with two items: “how often do you exercise (intense enough to be out of breath)?” and “how often do you play sport?”. Ratings were combined into a single dichotomous indicator (intense exercise or sport more than once a week vs less often).19 We first assessed adolescent-reported self-efficacy at wave 2 using the ten-item Generalized Self Efficacy Scale (α=0·98). This measure assesses young people's perceived ability to overcome problems, cope with adversity and achieve difficult tasks (eg, “if I am in trouble, I can usually think of a solution”).31

ten).19 We first assessed adolescent-reported self-efficacy at wave 2 using the ten-item Generalized Self Efficacy Scale (α=0·98). This measure assesses young people's perceived ability to overcome problems, cope with adversity and achieve difficult tasks (eg, “if I am in trouble, I can usually think of a solution”).31 Statistical analysis Logistic regression analyses examined associations between protective factors and sustained good mental health in offspring. This outcome variable was defined as the absence at all three waves of any DSM-IV disorder diagnosis, of elevated CAPA interview depression or behaviour disorder symptoms (both three or more), or of suicidal ideation or self-harm. Interactions with gender and age were also tested. Cumulative effects models tested the extent to which significant predictors jointly contributed to sustained good mental health. For this purpose, we dichotomised protective factors using standard cutpoints or otherwise median splits. We repeated the analyses for subgroups of families in which parent severity differed: depressive episode recurrence over the study, yes or no; past severe depressive episode (Global Assessment of Functioning score <30 or admitted to hospital), yes or no.

Statistical analysis Logistic regression analyses examined associations between protective factors and sustained good mental health in offspring. This outcome variable was defined as the absence at all three waves of any DSM-IV disorder diagnosis, of elevated CAPA interview depression or behaviour disorder symptoms (both three or more), or of suicidal ideation or self-harm. Interactions with gender and age were also tested. Cumulative effects models tested the extent to which significant predictors jointly contributed to sustained good mental health. For this purpose, we dichotomised protective factors using standard cutpoints or otherwise median splits. We repeated the analyses for subgroups of families in which parent severity differed: depressive episode recurrence over the study, yes or no; past severe depressive episode (Global Assessment of Functioning score <30 or admitted to hospital), yes or no. We created the continuous outcomes of mood resilience and behavioural resilience using residual scores generated via regression analysis. Adolescent mood disorder and behaviour disorder symptom counts at final follow-up were regressed onto the predictor variables indexing parent illness-related risks (parent depression age at onset, parent depression severity, family history of depression). Negative residual scores indicate better than predicted offspring mood and behaviour (resilience) at follow-up and allow for variability in the level of parental depression-related risk.16, 17 Univariate and multivariate models tested associations between hypothesised protective factors and the two derived continuous outcome measures of mood and behavioural resilience.

cted offspring mood and behaviour (resilience) at follow-up and allow for variability in the level of parental depression-related risk.16, 17 Univariate and multivariate models tested associations between hypothesised protective factors and the two derived continuous outcome measures of mood and behavioural resilience. Role of the funding source The funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of the 469 families screened between February and June, 2007, we included 337 index parents with recurrent major depressive disorder (315 women and 22 men) and adolescent offspring who were biologically related to and living with the affected parent (figure 1). Six families were excluded at follow-up because of bipolar disorder in the parent with depression (n=2) or because adolescents had not been exposed to episodes of parental depression in their lifetime (n=4), so the final number of families included in the eligible sample was 331 (194 girls and 137 boys, mean age 12·4 years [SD 2·0] at baseline). Table 1 shows the demographic characteristics of the eligible sample at baseline.

cause adolescents had not been exposed to episodes of parental depression in their lifetime (n=4), so the final number of families included in the eligible sample was 331 (194 girls and 137 boys, mean age 12·4 years [SD 2·0] at baseline). Table 1 shows the demographic characteristics of the eligible sample at baseline. Full information on offspring mental health was available for 262 (79%) of the 331 eligible baseline sample. We present the results for the analysis using the complete case sample (n ≤262). Sensitivity analyses used multiple imputation (see the appendix for information on characteristics of sample retained vs not retained, details of the imputation, and imputed results). Results for imputed data were closely similar. Overall, 103 (39%) of 262 adolescents met criteria for a psychiatric diagnosis, 118 (45%) had elevated depression symptoms, 182 (70%) elevated behaviour symptoms, and 73 (28%) exhibited suicidal ideation or self-harm on at least one occasion. 53 (20%) exhibited none of these mental health problems across the study period and thus met study criteria for sustained good mental health (table 2). The likelihood of sustained good mental health did not differ by adolescent gender or age, but was lower for adolescents whose parents had a history of severe depressive episodes (table 3). SDQ screen scores indicated equivalent or better mental health for those with no mental health problems than for the UK population for this age group (appendix).

d mental health did not differ by adolescent gender or age, but was lower for adolescents whose parents had a history of severe depressive episodes (table 3). SDQ screen scores indicated equivalent or better mental health for those with no mental health problems than for the UK population for this age group (appendix). Family, social, and adolescent cognitive or health behaviour factors were associated with sustained good mental health in offspring (table 4). Positive expressed emotion in index parents, co-parent support, parent-rated peer relationship quality, adolescent self-efficacy, and frequent exercise were all associated with good mental health. Index-parent warmth, sibling warmth, out-of-school activities, and adolescent-rated peer relationship and friendship quality were not significant predictors. These bivariate associations did not differ according to gender or age (all interactions non-significant, p>0·05). The likelihood of sustained good mental health in the offspring increased with the total number of significant protective factors present across family, social, and adolescent cognitive or health behaviour domains (odds ratio, OR=2·27 [1·62–3·19], p<0·0001; figure 2). The proportion of adolescents with sustained good mental health ranged from 3·8% for zero or one protective factor to 48·0% for five protective factors.

gnificant protective factors present across family, social, and adolescent cognitive or health behaviour domains (odds ratio, OR=2·27 [1·62–3·19], p<0·0001; figure 2). The proportion of adolescents with sustained good mental health ranged from 3·8% for zero or one protective factor to 48·0% for five protective factors. As a sensitivity check, analyses investigated whether protective effects varied between families where parents had (n=167) or had not (n=93) experienced an episode recurrence by follow-up, or had (n=73) or had not (n=186) experienced a past severe episode of depression (Global Assessment of Functioning score <30 or hospital admission). The number of protective factors was associated with offspring mental health irrespective of parent depression episode recurrence (association within subgroups: recurrence, OR=2·45 [1·56–3·83], p<0·0001; no recurrence, OR=1·98 [1·17–3·37], p=0·011; interaction: OR=1·23 [0·62–2·47], p=0·56). The association between number of protective factors and offspring sustained mental health was significant for those not exposed to a severe parent depression episode (OR=2·23 [1·53–3·24], p<0·0001) but not in the subgroup exposed to a severe episode (OR=2·67 [0·91–7·87], p=0·075), although the ORs did not differ significantly between the two subgroups (interaction OR=1·20 (0·38–3·76), p=0·76).

mental health was significant for those not exposed to a severe parent depression episode (OR=2·23 [1·53–3·24], p<0·0001) but not in the subgroup exposed to a severe episode (OR=2·67 [0·91–7·87], p=0·075), although the ORs did not differ significantly between the two subgroups (interaction OR=1·20 (0·38–3·76), p=0·76). Table 5 shows univariate tests of association with mood and behavioural resilience. Negative residuals suggested lower than expected mood and behaviour symptom scores given parental depression severity. Baseline co-parent support (but not index parent factors), better parent-rated and adolescent-rated social relationships, and adolescent self-efficacy were associated both with mood and behavioural resilience at the final assessment. Frequent exercise was associated with mood resilience only, whereas index parent warmth and positive expressed emotion were associated with behavioural resilience only. Two multivariate models of mood and behaviour resilience were then examined taking forward significant univariate protective factors (appendix). The first model found that co-parent support (β=–0·19, p=0·004), adolescent self-efficacy (β=–0·19, p=0·004), and adolescent exercise (β=–0·17, p=0·01) independently predicted mood resilience, with a marginal effect of parent-rated peer relationship quality (β=–0·14, p=0·05). The second showed that parent-rated peer relationship quality (β=–0·16, p=0·04) and adolescent self-efficacy (β=–0·21, p=0·004) independently predicted behavioural resilience.

e (β=–0·17, p=0·01) independently predicted mood resilience, with a marginal effect of parent-rated peer relationship quality (β=–0·14, p=0·05). The second showed that parent-rated peer relationship quality (β=–0·16, p=0·04) and adolescent self-efficacy (β=–0·21, p=0·004) independently predicted behavioural resilience. When we did our sensitivity analyses, we saw that all results were comparable when excluding male index parents (n=19) from the sample with sustained mental health information (data not shown). When additionally excluding offspring not living with their father at baseline (n=73) findings were similar for associations between co-parent (ie, paternal) support with offspring sustained mental health (OR=1·93 [1·31–2.81], p=0·001) and mood resilience at follow-up (β=–0·22, p=0·006). However, there was no longer evidence of an association between paternal support and behavioural resilience (β=–0·15, p=0·07). The multivariate model results for mood and behavioural resilience remained identical using alternative forward and backward step-wise regression that included all predictor variables. Finally, analyses were repeated using multiple imputation to address missing data. Results were closely comparable (appendix).

p=0·07). The multivariate model results for mood and behavioural resilience remained identical using alternative forward and backward step-wise regression that included all predictor variables. Finally, analyses were repeated using multiple imputation to address missing data. Results were closely comparable (appendix). Discussion Our findings show that, as a group, offspring of parents with recurrent depression experienced high rates of mental health problems. Despite this finding, about one in five adolescents had sustained good mental health across all three waves of assessment. Index-parent positive expressed emotion, support from co-parents, good quality social relationships, youth self-efficacy, and regular physical exercise all predicted sustained good mental health. Protective factors also predicted lower than expected youth mood and behaviour symptoms at follow-up given the severity of their parents' depression. The findings of this study are important because they highlight several novel protective factors associated with resilience in adolescents at high risk of psychiatric problems due to recurrent parental depression. The findings advance our understanding by directly taking account of variation in the severity of parents' depression and by considering protective effects across multiple domains including co-parent support and adolescent physical exercise that have not been previously investigated. The results are also novel in highlighting that multiple protective factors are required, and in showing differential protection for mood and behavioural resilience.

dering protective effects across multiple domains including co-parent support and adolescent physical exercise that have not been previously investigated. The results are also novel in highlighting that multiple protective factors are required, and in showing differential protection for mood and behavioural resilience. Depressive disorder in adults who are parents is common, as is cross-generational transmission of mental health problems. However, our findings have shown that mental health problems in this high-risk group of young people are not inevitable, even when parents have experienced multiple episodes of major depressive disorder. Against a background of significantly elevated risk for psychopathology, including depression, behaviour problems, and suicidality, a subgroup of offspring (20%) was characterised by a pattern of sustained good mental health in adolescence and remained problem free for the duration of this study. This subgroup showed better or equivalent levels of mental health when compared with age-appropriate general population norms for the SDQ.

nd suicidality, a subgroup of offspring (20%) was characterised by a pattern of sustained good mental health in adolescence and remained problem free for the duration of this study. This subgroup showed better or equivalent levels of mental health when compared with age-appropriate general population norms for the SDQ. The identification of factors associated with good mental health in adolescents who are at high familial risk has important implications for treatment and prevention. Findings from this study show that family, social, cognitive, and health behaviour factors contributed to resilience, when defined as sustained good adolescent mental health, but that multiple protective factors were required. Even in this clinically derived sample of parents with recurrent depression there was substantial variation in parent illness course and severity. Crucially, we were able to also show that protective effects did not simply reflect variation in parental depression severity. Specifically we accounted for parent age at onset and severity of depression as well as family history of depression to assess better than expected functioning given the severity of risk exposure. Our results showed that co-parent support, social factors, and adolescents' self-efficacy each predicted both mood and behavioural resilience. The study findings are also novel in that they reveal that contributors to mood and behavioural resilience differ. Interestingly, frequent physical exercise and co-parent emotional support were found to have a specific association with mood resilience only.

-efficacy each predicted both mood and behavioural resilience. The study findings are also novel in that they reveal that contributors to mood and behavioural resilience differ. Interestingly, frequent physical exercise and co-parent emotional support were found to have a specific association with mood resilience only. Co-parent emotional support emerged as a particularly strong predictor of mental health resilience in this high-risk sample. Most of the index parents in this sample were mothers, and the present findings show the importance of father involvement in moderating the potential impact of maternal depression. The reasons why most of our sample consisted of mothers with depression may include gender differences in prevalence of adult depression32 and in help-seeking and presentation in primary care, as well as a greater likelihood for mothers with depression to opt in to our study once approached. These findings highlight the potential benefits of including the wider family in prevention programmes for adolescent depression.12 Research has shown protective effects of good-quality social relationships in relation to psychosocial adversities such as maltreatment.33 We showed that there were also strong associations with mental health resilience in offspring of parents with depression.

Co-parent emotional support emerged as a particularly strong predictor of mental health resilience in this high-risk sample. Most of the index parents in this sample were mothers, and the present findings show the importance of father involvement in moderating the potential impact of maternal depression. The reasons why most of our sample consisted of mothers with depression may include gender differences in prevalence of adult depression32 and in help-seeking and presentation in primary care, as well as a greater likelihood for mothers with depression to opt in to our study once approached. These findings highlight the potential benefits of including the wider family in prevention programmes for adolescent depression.12 Research has shown protective effects of good-quality social relationships in relation to psychosocial adversities such as maltreatment.33 We showed that there were also strong associations with mental health resilience in offspring of parents with depression. Extending previous findings,14, 15 self-efficacy was not only a strong predictor of sustained mental health, but also of mood and behaviour resilience at follow-up. Belief in one's ability to successfully deal with adversity might be especially important in the context of parental depression if it allows young people to better rationalise their parent's illness, and exert greater control over their own responses to stressors that result from parent illness.13, 15

nce at follow-up. Belief in one's ability to successfully deal with adversity might be especially important in the context of parental depression if it allows young people to better rationalise their parent's illness, and exert greater control over their own responses to stressors that result from parent illness.13, 15 Finally, frequent physical exercise was associated with lower than predicted depression problems. This supports National Institute for Health and Care Excellence guidance that regular intense exercise is advised to ameliorate or prevent depression, evidence that has thus far been lacking for young people.19 The findings suggest that there are potential modifiable targets for preventative interventions that might help interrupt the intergenerational transmission of risk for psychopathology between parents with depression and their children. Treatment of parents is a priority but might be insufficient on its own to prevent psychopathology in offspring. Treatment studies have shown that remission of parental depression is associated with reductions in some types of offspring symptoms. However, not all parents with depression seek treatment, and treatment is not always successful. Furthermore, likelihood of success is associated with parental depression severity (controlled for in this study).7, 8

that remission of parental depression is associated with reductions in some types of offspring symptoms. However, not all parents with depression seek treatment, and treatment is not always successful. Furthermore, likelihood of success is associated with parental depression severity (controlled for in this study).7, 8 The study highlights several additional targets for preventative intervention beyond risk reduction. These include the facilitation of support from co-parents and young people's social relationships. Cognitive behavioural programmes that enhance adolescents' sense of efficacy together with programmes to promote healthy lifestyles, including exercise, are also likely to be important.

intervention beyond risk reduction. These include the facilitation of support from co-parents and young people's social relationships. Cognitive behavioural programmes that enhance adolescents' sense of efficacy together with programmes to promote healthy lifestyles, including exercise, are also likely to be important. A number of cognitive-behaviourally orientated prevention programmes already exist. These seem to be effective for some children and adolescents in high-income and low-income countries, but importantly previous evidence also suggests such prevention programmes are ineffective when a parent is currently depressed.8 Careful thought is required in deciding about which programmes to target at which young people.34 One important message from the present study is that efforts to target single protective factors in isolation for at-risk adolescents are probably insufficient. Good mental health outcomes were typically achieved only if adolescents reported a combination of multiple protective factors across family, peer, and cognitive or behavioural domains. The results support the idea that multimodal interventions that simultaneously address multiple systems (eg, home, school, and the young person themselves) hold the greatest promise for preventing mental ill health.35, 36 Our findings are also relevant to adult mental health services. Raising awareness of the importance of early prevention of mental health problems in the offspring of parents with depression is important, as is enhancing effective links between adult and youth mental health services.

A number of cognitive-behaviourally orientated prevention programmes already exist. These seem to be effective for some children and adolescents in high-income and low-income countries, but importantly previous evidence also suggests such prevention programmes are ineffective when a parent is currently depressed.8 Careful thought is required in deciding about which programmes to target at which young people.34 One important message from the present study is that efforts to target single protective factors in isolation for at-risk adolescents are probably insufficient. Good mental health outcomes were typically achieved only if adolescents reported a combination of multiple protective factors across family, peer, and cognitive or behavioural domains. The results support the idea that multimodal interventions that simultaneously address multiple systems (eg, home, school, and the young person themselves) hold the greatest promise for preventing mental ill health.35, 36 Our findings are also relevant to adult mental health services. Raising awareness of the importance of early prevention of mental health problems in the offspring of parents with depression is important, as is enhancing effective links between adult and youth mental health services. The study used a large longitudinal sample of parents with depression and adolescents, well validated multi-informant assessments of psychiatric disorders and subthreshold problems, together with careful characterisation of parental depression risk and hypothesised resilience promoting factors.

Our findings are also relevant to adult mental health services. Raising awareness of the importance of early prevention of mental health problems in the offspring of parents with depression is important, as is enhancing effective links between adult and youth mental health services. The study used a large longitudinal sample of parents with depression and adolescents, well validated multi-informant assessments of psychiatric disorders and subthreshold problems, together with careful characterisation of parental depression risk and hypothesised resilience promoting factors. This study has also limitations. First, further follow-up will be needed to determine whether resilience is sustained into early adulthood. Second, even carefully designed longitudinal observational studies cannot unambiguously identify causal influences, because of the possible role of other unmeasured confounders or reverse causation. Finally, we cannot rule out the contribution of genetic factors that were not indexed by parent illness severity. In conclusion, depression is familial, and, although psychiatric problems among adolescents whose parents have recurrent depression are common, they are not inevitable. Some young people show unexpectedly positive outcomes. The study identified several potentially modifiable protective factors that together seem to promote adolescents' mental health resilience. These findings now need to be taken forward by refining existing preventive interventions, developing new ones, and testing through randomised trials. Supplementary Material Supplementary appendix

In conclusion, depression is familial, and, although psychiatric problems among adolescents whose parents have recurrent depression are common, they are not inevitable. Some young people show unexpectedly positive outcomes. The study identified several potentially modifiable protective factors that together seem to promote adolescents' mental health resilience. These findings now need to be taken forward by refining existing preventive interventions, developing new ones, and testing through randomised trials. Supplementary Material Supplementary appendix Acknowledgments We are very grateful to all the families and GPs who helped with this study. We thank the assistant psychologists who helped with data collection. The work was supported by the Jules Thorn Charitable Trust (JTA/06) and the Economic and Social Research Council (ES/J011657/1). Contributors SC, MJO, NC, AKT, GTH, FR, and AT conceived of the research question and/or designed the study. SC, NC, AKT, GTH, FR and AT oversaw the data acquisition. SC oversaw the analysis. GH, LM and RS analysed the data. SC, GH, LM, RS, NC, AKT, FR, GTH and AT interpreted the data. SC, GH, and AT drafted the report. All authors provided critical revisions to the report, important intellectual content and final approval. Declaration of interests We declare no competing interests. Figure 1 Retention of participants in the Early Prediction of Adolescent Depression study (EPAD)

Contributors SC, MJO, NC, AKT, GTH, FR, and AT conceived of the research question and/or designed the study. SC, NC, AKT, GTH, FR and AT oversaw the data acquisition. SC oversaw the analysis. GH, LM and RS analysed the data. SC, GH, LM, RS, NC, AKT, FR, GTH and AT interpreted the data. SC, GH, and AT drafted the report. All authors provided critical revisions to the report, important intellectual content and final approval. Declaration of interests We declare no competing interests. Figure 1 Retention of participants in the Early Prediction of Adolescent Depression study (EPAD) Initial telephone screening identified 469 families. 116 families withdrew before baseline assessment, or were withdrawn for other reasons (incomplete baseline assessments, bipolar diagnosis at baseline assessment; n=16). The baseline eligible sample thus consisted of 337 participants. *Participants were recruited primarily from 62 general practitioner surgeries (263 of 337 of eligible baseline sample), from a database of previously identified adults with recurrent unipolar depression from the community (64 of 337), and via other methods (posters in local health centres, and depression alliance newsletter; 10 of 337). †Numbers vary in main analyses from 209 to 260 due to missing data on individual protective factors. Figure 2 Cumulative influences on sustained good mental health (n=220)

Initial telephone screening identified 469 families. 116 families withdrew before baseline assessment, or were withdrawn for other reasons (incomplete baseline assessments, bipolar diagnosis at baseline assessment; n=16). The baseline eligible sample thus consisted of 337 participants. *Participants were recruited primarily from 62 general practitioner surgeries (263 of 337 of eligible baseline sample), from a database of previously identified adults with recurrent unipolar depression from the community (64 of 337), and via other methods (posters in local health centres, and depression alliance newsletter; 10 of 337). †Numbers vary in main analyses from 209 to 260 due to missing data on individual protective factors. Figure 2 Cumulative influences on sustained good mental health (n=220) Likelihood of sustained good mental health in offspring according to total number of identified protective factors: positive index parent expressed emotion (high or very high), coparent support (median split; score >3), good-quality social relationships (parent Strengths and Difficulties Questionnaire peer subscale in normal range), adolescent efficacy (median split; General Self Efficacy Scale >28), physical exercise (intense exercise or sport more often than once per week) Table 1 Parent and child demographic characteristics of the eligible sample at baseline

Likelihood of sustained good mental health in offspring according to total number of identified protective factors: positive index parent expressed emotion (high or very high), coparent support (median split; score >3), good-quality social relationships (parent Strengths and Difficulties Questionnaire peer subscale in normal range), adolescent efficacy (median split; General Self Efficacy Scale >28), physical exercise (intense exercise or sport more often than once per week) Table 1 Parent and child demographic characteristics of the eligible sample at baseline Number of individuals (%) or mean (SD), n=331 Parent characteristics Female 309 (93%) Age at baseline, years 41·6 (5·4) Single parent 95 (29%) Family income below £20 000 91 (30%) Child characteristics Female 194 (59%) Age at baseline, years 12·4 (2·0) IQ 94·9 (12·9) IQ=intelligence quotient. Table 2 Offspring without mental health problems at each wave and by gender Number of participants (%) Boys Baseline assessment (n=131) 48 (37%) First follow-up (n=111) 43 (39%) Second follow-up (n=115) 45 (39%) Across the study period (n=105) 21 (20%) Girls Baseline assessment (n=188) 61 (32%) First follow-up (n=172) 70 (41%) Second follow-up (n=162) 59 (36%) Across the study period (n=157) 32 (20%) Table 3 Offspring of parents with depression: mental health according to differences in risk exposure and adolescent characteristics

ss the study period (n=105) 21 (20%) Girls Baseline assessment (n=188) 61 (32%) First follow-up (n=172) 70 (41%) Second follow-up (n=162) 59 (36%) Across the study period (n=157) 32 (20%) Table 3 Offspring of parents with depression: mental health according to differences in risk exposure and adolescent characteristics Sustained good mental health (n=53) No sustained good mental health (n=209) OR (95% CI)* p value Index parent depression severity Age at first episode, years 29·16 (9·33) 25·68 (8·21) 1·49 (1·10–2·02) 0·010 Worst episode hospitalisation or GAF <30 5/52 (10%) 68/207 (33%) 0·22 (0·08–0·57) 0·002 Recurrence during study period 31/53 (58%) 136/207 (66%) 0·74 (0·40–1·36) 0·33 Antenatal depression 3/46 (7%) 22/194 (11%) 0·55 (0·16–1·91) 0·34 Postnatal depression 18/46 (39 %) 84/194 (43%) 0·84 (0·44–1·62) 0·61 Additional family history of depression Two or more first-degree or second-degree relatives 20/53 (38%) 83/209 (40%) 0·92 (0·50–1·71) 0·79 Adolescent characteristics Age, years 12·32 (2·09) 12·39 (2·03) 0·97 (0·72–1·31) 0·83 Female 32/53 (60%) 125/209 (60%) 1·02 (0·55–1·90) 0·94 Data are n (%) or mean (SD). Total numbers vary because of occasional missing data for measures of index parent severity measures. OR=odds ratio. GAF=global assessment of functioning. * For analyses of parent age at onset and offspring age, ORs indicate change in odds of sustained good mental health per one SD change in mean age. Table 4 Univariate associations of family, social and adolescent cognitive/health behaviour factors with sustained good adolescent mental health

Sustained good mental health (n=53) No sustained good mental health (n=209) OR (95% CI)* p value Index parent depression severity Age at first episode, years 29·16 (9·33) 25·68 (8·21) 1·49 (1·10–2·02) 0·010 Worst episode hospitalisation or GAF <30 5/52 (10%) 68/207 (33%) 0·22 (0·08–0·57) 0·002 Recurrence during study period 31/53 (58%) 136/207 (66%) 0·74 (0·40–1·36) 0·33 Antenatal depression 3/46 (7%) 22/194 (11%) 0·55 (0·16–1·91) 0·34 Postnatal depression 18/46 (39 %) 84/194 (43%) 0·84 (0·44–1·62) 0·61 Additional family history of depression Two or more first-degree or second-degree relatives 20/53 (38%) 83/209 (40%) 0·92 (0·50–1·71) 0·79 Adolescent characteristics Age, years 12·32 (2·09) 12·39 (2·03) 0·97 (0·72–1·31) 0·83 Female 32/53 (60%) 125/209 (60%) 1·02 (0·55–1·90) 0·94 Data are n (%) or mean (SD). Total numbers vary because of occasional missing data for measures of index parent severity measures. OR=odds ratio. GAF=global assessment of functioning. * For analyses of parent age at onset and offspring age, ORs indicate change in odds of sustained good mental health per one SD change in mean age. Table 4 Univariate associations of family, social and adolescent cognitive/health behaviour factors with sustained good adolescent mental health Sustained good mental health (n=53) No sustained good mental health (n=209) OR (95% CI)* p value N n (%) or mean (SD) N n (%) or mean (SD) Family factors Index parent warmth 53 36·74 (5·75) 200 35·84 (6·08) 1·19 (0·84–1·69) 0·34 Index parent positive expressed emotion 52 3·85 (0·75) 206 3·33 (1·00) 1·91 (1·31–2·79) 0·0008 Co-parent support to adolescent 52 3·79 (2·73) 208 2·08 (2·61) 1·90 (1·38–2·62) <0·0001 Sibling warmth 40 17·00 (4·86) 169 16·38 (4·92) 1·14 (0·80–1·61) 0·48 Social factors Parent-reported peer relationship quality 53 9·00 (1·43) 202 7·91 (2·12) 2·07 (1·35–3·18) 0·001 Adolescent-reported peer relationship quality 51 8·45 (1·47) 201 7·97 (1·82) 1·36 (0·96–1·93) 0·08 Out of school activities (monthly) 50 33 (66%) 197 114 (58%) 1·41 (0·74–2·71) 0·30 Adolescent perceived friendships 52 28·42 (5·02) 197 27·04 (5·81) 1·30 (0·94–1·81) 0·12 Adolescent self-efficacy and exercise Self efficacy 48 29·19 (3·02) 186 27·46 (5·06) 1·49 (1·05–2·11) 0·03 Frequent physical exercise 52 45 (87%) 200 137 (69%) 2·96 (1·26–6·92) 0·01 Data are mean (SD) or n (%), unless otherwise specified. For scale scores, ORs indicate change in odds per one SD change in mean scale score.

lescent self-efficacy and exercise Self efficacy 48 29·19 (3·02) 186 27·46 (5·06) 1·49 (1·05–2·11) 0·03 Frequent physical exercise 52 45 (87%) 200 137 (69%) 2·96 (1·26–6·92) 0·01 Data are mean (SD) or n (%), unless otherwise specified. For scale scores, ORs indicate change in odds per one SD change in mean scale score. * Ns for predictor variables range from 209 to 260. Data from multiple imputation models shown in supplementary table 1. Table 5 Univariate associations of family, social and adolescent cognitive or health behaviour factors with mood and behaviour resilience at final follow-up Standardised residuals Mood resilience Standardised residuals Behavioural resilience N β p value N β p value Family factors Index-parent warmth 260 −0·06 0·33 256 −0·17 0·007 Index-parent positive expressed emotion 261 −0·11 0·08 257 −0·16 0·01 Co-parent support 268 −0·23 0·0001 264 −0·14 0·03 Sibling warmth 211 0·06 0·43 208 −0·10 0·15 Social Factors Parent-reported peer relationship quality 260 −0·17 0·006 256 −0·23 0·0002 Adolescent-reported peer relationship quality 256 −0·17 0·005 253 −0·16 0·01 Out of school activities 251 −0·15 0·02 248 −0·10 0·12 Adolescent perceived friendships 253 −0·13 0·03 250 −0·15 0·02 Adolescent cognition or behaviour Self-efficacy 228 −0·22 0·001 224 −0·25 0·0001 Frequent physical exercise 256 −0·22 0·0004 253 −0·001 0·99 Ns for predictor variables and outcome data range from 208 to 268. Data from multiple imputation models shown in supplementary table 2.

Introduction Many mental disorders emerge during adolescence and continue into adulthood.1 In depressive disorders, younger onset is associated with more depressive episodes, longer episode duration, increased comorbidity, suicidality, and admission to hospital.2 Among individuals with a diagnosed depressive disorder, adolescents are more likely than adults to delay contact with mental health services, thereby increasing episode duration and risk of recurrence. Clearly, early identification and treatment of mental disorders during adolescence would contribute to reduction and perhaps prevention of adverse sequelae. Measurement of the treatment gap—the discrepancy between disorder prevalence and proportion treated—is a prerequisite to enable policy makers to prevent such adverse sequelae from arising. To predict service need, a clearly recognised cutoff for mental disorder, such as meeting DSM diagnostic criteria, is desirable. Our review of international studies that report DSM-IV disorder and past-year contact with mental health services for those with a disorder (appendix pp 1, 2), found that 12–25% of adolescents have a mental disorder, of whom only 34–56% access mental health services. Previous surveys3, 4 in the UK report much higher proportions of contact with mental health services (71% of children or adolescents with a mental disorder); however, unlike most studies, these estimates classify seeking help from a teacher as a mental health service contact. Other studies5, 6 report 12–19% lower service use rates for anxiety than for depression.

gher proportions of contact with mental health services (71% of children or adolescents with a mental disorder); however, unlike most studies, these estimates classify seeking help from a teacher as a mental health service contact. Other studies5, 6 report 12–19% lower service use rates for anxiety than for depression. The association between adolescents' contact with mental health services and subsequent mental health remains unclear in community samples, but is vital to clarify if adolescent mental health services are to compete for health-care funding. Findings from studies7, 8 using broad definitions of mental health problems without a cutoff for service need have shown that use of mental health services had little effect on subsequent mental health problems. However, results are more promising if adolescents are at greater risk of, or already have, a mental disorder. In adolescents who witnessed community violence, use of mental health services reduced depressive symptoms.9 Adolescents with fearful spells or panic attacks were more likely to develop diagnosable panic disorder and depression if they had not used mental health services.10 Patients treated for emotional disorders at Child and Adolescent Mental Health Services (CAMHS) showed significant improvement compared with controls,11 yet this change was not clinically meaningful. Finally, among DSM-diagnosed adolescents, users of specialist mental health services had reduced symptoms compared with those who were untreated, but only if eight or more sessions were attended.12 However, none of the studies that showed a positive association between service contact and mental health addressed non-randomisation or attrition. Only one study9 adequately addressed confounding variables (ie, those associated with both predictor and outcome, which could bias the association between service use and subsequent mental health), and only one study10 showed significant effects that were clinically relevant.

h addressed non-randomisation or attrition. Only one study9 adequately addressed confounding variables (ie, those associated with both predictor and outcome, which could bias the association between service use and subsequent mental health), and only one study10 showed significant effects that were clinically relevant. Research in context Evidence before this study In 2015, a task force in the UK noted the paucity of good quality national information regarding Child and Adolescent Mental Health Services (CAMHS) outcomes. To identify previous published work, with no language restrictions, that assessed the association between CAMHS use and subsequent mental health, we searched PubMed (* denotes wildcard) for articles published in the past 16 years (from Jan 1, 2000, to July 5, 2016) for the terms (service* OR help-seek*) AND (psychopatholog* OR mental* OR psychiatric*) AND (observation* OR community OR survey OR cohort OR epidemiolog*) AND (longitudinal[Title] OR prospective[Title] OR change[Title] OR reduc*[Title] OR improve*[Title] OR effectiveness[Title] OR outcome[Title]) AND (adolescen*[Title] OR youth*[Title] OR young*[Title]). We required studies to reflect treatment-as-usual mental health service use, and have a non-service using comparison group. We identified additional papers by checking citations.

le] OR reduc*[Title] OR improve*[Title] OR effectiveness[Title] OR outcome[Title]) AND (adolescen*[Title] OR youth*[Title] OR young*[Title]). We required studies to reflect treatment-as-usual mental health service use, and have a non-service using comparison group. We identified additional papers by checking citations. We identified six studies that yielded mixed findings regarding the association of service contact with subsequent mental health. Two studies that assessed change in all service users without a clearly recognised cutoff for service need, such as DSM, showed that mental health service use had little effect on subsequent total mental health problems over and above that to be expected from natural remission. The four remaining studies assessed adolescents at greater risk of a mental disorder or those with a DSM diagnosis. These studies showed an improvement in mental health following service contact, but none addressed non-randomisation of service contact or attrition, only one adequately addressed confounding variables, and only one showed significant effects that were clinically relevant. None of these studies were from the UK (three were from the USA and three were from Europe). Added value of this study

We identified six studies that yielded mixed findings regarding the association of service contact with subsequent mental health. Two studies that assessed change in all service users without a clearly recognised cutoff for service need, such as DSM, showed that mental health service use had little effect on subsequent total mental health problems over and above that to be expected from natural remission. The four remaining studies assessed adolescents at greater risk of a mental disorder or those with a DSM diagnosis. These studies showed an improvement in mental health following service contact, but none addressed non-randomisation of service contact or attrition, only one adequately addressed confounding variables, and only one showed significant effects that were clinically relevant. None of these studies were from the UK (three were from the USA and three were from Europe). Added value of this study To our knowledge, this study is the first of its kind in the UK, and the first to support the association of mental health service contact and the improvement of mental health by late adolescence, while addressing non-randomisation of service contact and attrition. In addition to propensity score weighting (which balances treatment and control groups on confounders, similar to a randomised control trial) to adjust for participants' initial likelihood to access services, and multiple imputation to deal with missing data, we used a clinically relevant cutoff and adjusted for a wide range of time-varying confounding variables. These adjustments give greater confidence than previous studies to the notion that mental health service contact is related to meaningful improvements in subsequent mental health. This study is also the first we are aware of that shows that the association of mental health with previous treatment is attenuated if that treatment was irrespective of service need.

han previous studies to the notion that mental health service contact is related to meaningful improvements in subsequent mental health. This study is also the first we are aware of that shows that the association of mental health with previous treatment is attenuated if that treatment was irrespective of service need. Implications of all the available evidence The spending of the UK National Health Service (NHS) on children's mental health services has fallen by 5·4% in real terms since 2010 (£41 million), despite an increase in demand. The present findings support the positive role played by mental health services in a cohort before these NHS cuts, illustrating to policy makers the validity of increasing the availability of child mental health services to at least 2010 levels. That positive findings became non-significant upon inclusion of all mental health service users irrespective of disorder underscores the importance of clinical assessment when making referral decisions. These findings support training of service referrers (eg, in primary care or schools) in detection of the presenting features of mental disorders, to increase the proportion of referrals of individuals with a clear need who could be more responsive to treatment.

ce of clinical assessment when making referral decisions. These findings support training of service referrers (eg, in primary care or schools) in detection of the presenting features of mental disorders, to increase the proportion of referrals of individuals with a clear need who could be more responsive to treatment. In the present study, we used a longitudinal repeated-measures design on a community ascertained cohort to assess change in adolescent depressive symptoms from ages 14 years to 17 years after contact with mental health services. For the outcome, we used depressive symptoms as a valid identifier of major depressive disorders,13, 14 which are highly prevalent5 and predictive of future morbidity.2 To extend this previous work, the design controls for differences in symptoms and background factors among service users and non-users at baseline and over time, in individuals with and without a DSM-defined mental disorder. We hypothesised that self-reported depression scores would be reduced to a greater extent in adolescents who contacted mental health services than in those with no contact, but that these effects would be stronger in the subsample with a clearly defined need for mental health services, based on the presence of a diagnosable mental disorder. We hypothesised that these findings would remain when we addressed attrition, confounding variables, non-randomisation of mental health service contact, and clinical relevance.

ld be stronger in the subsample with a clearly defined need for mental health services, based on the presence of a diagnosable mental disorder. We hypothesised that these findings would remain when we addressed attrition, confounding variables, non-randomisation of mental health service contact, and clinical relevance. Methods Study design and participants As part of the ROOTS longitudinal cohort study15 of mental health, between April 28, 2005, and March 17, 2010, we recruited 1238 adolescents and primary caregivers (1134 [92%] were the biological mother of the adolescent) from 27 secondary schools in Cambridgeshire, UK. 18 secondary schools approached agreed to participate, with letters of invitation sent until the sample size reached a prespecified cutoff of 1000 participants. Of a possible 3762 students, 1238 agreed to participate. Participants were interviewed separately and completed questionnaires at mean ages 14·5 years (timepoint 1 [T1]), 16 years (timepoint 2 [T2]), and 17·5 (timepoint 3 [T3]) years (T1–3 means, SDs 0·3). Written informed consent was obtained from adolescents and caregivers before participation. Cambridgeshire 2 Research Ethics Committee local ethics committee provided ethics approval.

naires at mean ages 14·5 years (timepoint 1 [T1]), 16 years (timepoint 2 [T2]), and 17·5 (timepoint 3 [T3]) years (T1–3 means, SDs 0·3). Written informed consent was obtained from adolescents and caregivers before participation. Cambridgeshire 2 Research Ethics Committee local ethics committee provided ethics approval. Procedures At T1, trained researchers assessed adolescents' mental health status using the Schedule for Affective Disorders and Schizophrenia for School-Age Children–Present and Lifetime version (K-SADS-PL)16 to establish DSM-IV17 axis 1 diagnoses. Adolescents with a high clinical index (subthreshold for diagnosis, but exhibiting four symptoms and clinical impairment) were classified as diagnosed. Experienced psychiatrists (IMG, PBJ) trained interviewers and conducted consensus meetings regarding all K-SADS assessments. Inter-rater agreement for diagnosis was high (95%). Disagreements were settled by clinical consensus meetings between clinical psychiatry experts.

inical impairment) were classified as diagnosed. Experienced psychiatrists (IMG, PBJ) trained interviewers and conducted consensus meetings regarding all K-SADS assessments. Inter-rater agreement for diagnosis was high (95%). Disagreements were settled by clinical consensus meetings between clinical psychiatry experts. Mental health service contact was defined as an adolescent's assessment or treatment of a mental health problem by a primary care provider (ie, general practitioner) or a mental health specialist from any sector. Caregiver and adolescent responses were obtained by use of various measures (items in appendix pp 3–5). T1 past-year mental health service contact (no or yes) was generated as an exposure variable, and any mental health services after T1 (no or yes, post-T1–3) as a confounder. Caregivers reported contacts with adolescent mental health services at T1 from a semi-structured interview, with high inter-rater agreement on core indicators (κ=0·7–0·9; Cambridge Early Experiences Interview18) and from a self-reported questionnaire at T1 and T3. Adolescents were interviewed at T3 about mental health service contact before K-SADS-PL assessment. Adolescents also reported how often they had seen a doctor or other health professional regarding depressive symptoms in the past month (Kessler's Psychological Distress Scale19). We combined adolescent and caregiver responses with either response if one was missing, or with the positive response if sources disagreed (considering differential recall and caregivers potentially unaware of adolescent service use).

ing depressive symptoms in the past month (Kessler's Psychological Distress Scale19). We combined adolescent and caregiver responses with either response if one was missing, or with the positive response if sources disagreed (considering differential recall and caregivers potentially unaware of adolescent service use). A combined variable was derived at T1 that defined participants with current mental disorder (yes or no) and past-year mental health service contact. This variable resulted in four levels: unaffected (no current disorder or past-year service contact), service contact only, disorder only, and disorder and service contact. We assessed the Mood and Feelings Questionnaire (MFQ)20 at T1–3. This 33-item adolescent self-report of current or past 2 weeks' depressive symptoms covers DSM criteria for major depressive disorders. The MFQ has shown prognostic validity in clinic and non-clinic samples,13, 14 yielding high internal consistency (α=0·92–0·94) in the present sample. Higher sum scores indicate more symptoms.

-item adolescent self-report of current or past 2 weeks' depressive symptoms covers DSM criteria for major depressive disorders. The MFQ has shown prognostic validity in clinic and non-clinic samples,13, 14 yielding high internal consistency (α=0·92–0·94) in the present sample. Higher sum scores indicate more symptoms. We chose 18 putative confounders that covered sociodemographic, environmental, individual, mental health, and diagnostic domains (appendix pp 7, 8) based on a previous association with mental health service contact, or depression. For example, family structure, functioning and mental health problems, peer support, maltreatment, stressful events, socioeconomic status, gender, past referrals for mental health problems, current diagnosis type, severity, and comorbidity have all been related to current mental health service contact.21 We assessed seven confounders at multiple timepoints (appendix p 9). Statistical analysis We did primary analyses on an imputed dataset (appendix pp 7, 8) of individuals with complete data for T1 past-year mental health service contact and current mental disorder.

We chose 18 putative confounders that covered sociodemographic, environmental, individual, mental health, and diagnostic domains (appendix pp 7, 8) based on a previous association with mental health service contact, or depression. For example, family structure, functioning and mental health problems, peer support, maltreatment, stressful events, socioeconomic status, gender, past referrals for mental health problems, current diagnosis type, severity, and comorbidity have all been related to current mental health service contact.21 We assessed seven confounders at multiple timepoints (appendix p 9). Statistical analysis We did primary analyses on an imputed dataset (appendix pp 7, 8) of individuals with complete data for T1 past-year mental health service contact and current mental disorder. Imputed longitudinal MFQ scores were the outcome in multilevel mixed-effects linear regression models with maximum likelihood estimation, implemented in STATA 13.0. This analysis nests correlated data, thereby accounting for violations in the assumption of independence. For the present data, repeated assessments over time were nested within individuals (the random effect). Fixed effects (ie, predictors in the regression) included linear, quadratic, and categorical effects of age, and confounders (appendix p 7). We assessed categorical effects of T1 disorder and services (unaffected or disorder only or disorder and services) and this variable's interaction with age. We did not include the services-only group of individuals in the primary analysis because without a mental disorder their need for services was less clear. We explored the effects of nesting by school by adding school as a further random-effect.

r only or disorder and services) and this variable's interaction with age. We did not include the services-only group of individuals in the primary analysis because without a mental disorder their need for services was less clear. We explored the effects of nesting by school by adding school as a further random-effect. We did receiver operating characteristic (ROC) analysis to determine the ability of MFQ to classify affective disorder. In ROC analysis the true positive rate (sensitivity) is plotted against the false positive rate (1–specificity). We estimated the area under the curve (AUC) and used it as an index of diagnostic accuracy; a higher AUC reflects greater accuracy. The MFQ has previously been shown to have good-to-high diagnostic accuracy with this method.13, 14 Additionally, MFQ scores above the 75th percentile are an established behavioural marker for clinical diagnosis of major depression.22 The Youdin Index was calculated to determine the clinical cutoff point, because it maximises sensitivity and specificity,23 thereby increasing correct classification of individuals with and without depression.

above the 75th percentile are an established behavioural marker for clinical diagnosis of major depression.22 The Youdin Index was calculated to determine the clinical cutoff point, because it maximises sensitivity and specificity,23 thereby increasing correct classification of individuals with and without depression. To address the absence of randomisation of mental health service use, a propensity score was generated to weigh the outcome model. A propensity score is the individual probability of attending or receiving a service or treatment conditional on observed baseline covariates. The score is designed to balance confounders between a treatment and control group, as is done in a randomised control trial.24 The primary propensity-adjusted analyses comprised data from adolescents with a mental disorder, because those in the disorder-only group were the most appropriate for comparison with the disorder-and-services group (appendix pp 7, 8 provide further details of propensity score). To reduce estimate bias, we first did analyses of the full sample with a disorder, then we restricted the sample to the region of common support—the range of propensity scores which were observed in both treated and untreated individuals.25 We estimated the propensity score weighted outcome models with generalised linear modelling (GLM) with a logit link, with adjustment for post-baseline confounding variables. A robust estimator accounted for the sample weighting.

range of propensity scores which were observed in both treated and untreated individuals.25 We estimated the propensity score weighted outcome models with generalised linear modelling (GLM) with a logit link, with adjustment for post-baseline confounding variables. A robust estimator accounted for the sample weighting. To address the importance of use of a clearly defined need for mental health services based on the presence of a mental disorder, we reanalysed data including all service users, irrespective of disorder. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

To address the importance of use of a clearly defined need for mental health services based on the presence of a mental disorder, we reanalysed data including all service users, irrespective of disorder. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of the 1238 participants recruited, 1190 adolescents had data for T1 current mental disorder and past-year mental health service contact (appendix p 6). The number of respondents with complete data for all outcomes and covariates at all timepoints was 995 (84%) for T1, 778 (65%) for T2, and 806 (68%) for T3. 64 (5%) adolescents made past-year contact with mental health services; 126 (11%) had a current mental disorder. Among individuals with a disorder, 48 (38%) reported past-year service contact and 46 (96%) of these contacts were based on T1 past-year recall; 36 (84%) of 43 of these adolescents attended five or more sessions (n=5 had missing data for treatment length). In the disorder-and-services group (n=48), disorders were affective (n=16 [33%]), anxiety (n=10 [21%]), behavioural (n=25 [52%]), and other (n=5 [10%]); 14 (29%) of these participants had a comorbid K-SADS diagnosis (appendix p 9).

adolescents attended five or more sessions (n=5 had missing data for treatment length). In the disorder-and-services group (n=48), disorders were affective (n=16 [33%]), anxiety (n=10 [21%]), behavioural (n=25 [52%]), and other (n=5 [10%]); 14 (29%) of these participants had a comorbid K-SADS diagnosis (appendix p 9). Overall, 16 (25%) of 64 service users had no disorder, and differed from the disorder-and-services group: baseline MFQ scores were lower in the no-disorder group, although with no significant difference between groups (coefficient −7·64, 95% CI −15·30 to 0·02; p=0·051), and MFQ scores did not change over time (coefficient 1·22, −1·01 to 3·44; p=0·28). Adolescents with a disorder predominantly accessed CAMHS, whereas unaffected adolescents mostly accessed a school counsellor (appendix p 11). Unaffected service users were less antisocial than service users with a disorder (coefficient −3·20, 95% CI 1·10 to 5·29; p=0·0034); remaining covariates p>0·062 (means in appendix p 9). Adolescents with a disorder were substantially more impaired than unaffected adolescents across all domains of confounders (appendix p 9). When we compared adolescents with a disorder by mental health service contact, individuals varied mainly in diagnostic factors (appendix p 9). 1002 (84%) of 1190 service contacts were reported by both adolescents and caregivers, showing 98% agreement and high chance-corrected agreement (κ=0·78, 95% CI 0·71–0·84). The remaining service contacts were based on either adolescent or caregiver report.

Adolescents with a disorder were substantially more impaired than unaffected adolescents across all domains of confounders (appendix p 9). When we compared adolescents with a disorder by mental health service contact, individuals varied mainly in diagnostic factors (appendix p 9). 1002 (84%) of 1190 service contacts were reported by both adolescents and caregivers, showing 98% agreement and high chance-corrected agreement (κ=0·78, 95% CI 0·71–0·84). The remaining service contacts were based on either adolescent or caregiver report. Findings from adjusted multilevel mixed-effects regression analysis revealed that at T1, individuals in both the disorder only and disorder-and-services groups had significantly higher MFQ scores than did those in the unaffected group, but scores between the disorder only and disorder-and-services groups did not differ significantly (table 1, figure). MFQ scores in both these groups improved over time compared with the unaffected group, in which scores remained stable; however, scores improved more quickly among the disorder-and-services group than the disorder-only group (table 1, figure). By T3, scores in the disorder-and-services group had improved (reported reduced symptoms) to the levels of those in the unaffected group (table 1, figure). By contrast, at T3, patients in the disorder-only group reported significantly more symptoms than did those in both the disorder-and-services group and the unaffected group (table 1). Analyses repeated on complete case data yielded similar results (table 1; appendix p 12 shows imputed and complete-case analysis results from unadjusted models). Nesting by school did not affect complete-case results; thus, we did not do clustering during imputation. All data we present for comparability are non-nested results.

repeated on complete case data yielded similar results (table 1; appendix p 12 shows imputed and complete-case analysis results from unadjusted models). Nesting by school did not affect complete-case results; thus, we did not do clustering during imputation. All data we present for comparability are non-nested results. ROC analysis revealed MFQ as an excellent discriminator of affective disorder (AUC=0·93, 95% CI 0·90–0·96). The Youden Index indicated an MFQ clinical cutoff point of 22, with 94% sensitivity and 79% specificity, greater than previously obtained in a similar sample measured with differing cutoff point methodology.14

repeated on complete case data yielded similar results (table 1; appendix p 12 shows imputed and complete-case analysis results from unadjusted models). Nesting by school did not affect complete-case results; thus, we did not do clustering during imputation. All data we present for comparability are non-nested results. ROC analysis revealed MFQ as an excellent discriminator of affective disorder (AUC=0·93, 95% CI 0·90–0·96). The Youden Index indicated an MFQ clinical cutoff point of 22, with 94% sensitivity and 79% specificity, greater than previously obtained in a similar sample measured with differing cutoff point methodology.14 We included nine baseline covariates in the propensity score weighting (table 2). Propensity score weighted GLM revealed that among adolescents with a mental disorder, those without contact with mental health services at T1 had nearly four times the odds of being depressed by T3 compared with those in the disorder-and-services group (table 2). Inclusion of post-baseline confounding variables increased odds by more than five times, and in the common support sample, to more than seven times (table 2). Data for propensity score covariates were missing for five (4%) of 124 adolescents with a disorder. To assess the effect of MFQ imputation and missing covariate data on findings, we did unweighted GLM with mental health service contact at T1 predicting T3 MFQ clinical cutoff (adjusted by T1 MFQ only) in three separate models: model A (raw MFQ [n=95]), model B (imputed MFQ [n=124]), and model C (imputed MFQ with missing data from propensity score weighted covariates [n=119]). Effect sizes (calculated from odds ratios26) for mental health service contact in these models were similar (0·44 for model A, 0·46 for model B, and 0·45 for model C), indicating no effect of imputation or missing data.

24]), and model C (imputed MFQ with missing data from propensity score weighted covariates [n=119]). Effect sizes (calculated from odds ratios26) for mental health service contact in these models were similar (0·44 for model A, 0·46 for model B, and 0·45 for model C), indicating no effect of imputation or missing data. We repeated analyses by expanding the treatment group to include all adolescents who had made past-year contact with mental health services at T1, including 16 individuals with no T1 mental disorder. Comparison groups remained the same as before. The multilevel mixed-effects regression required the same confounding variables as the primary analyses, yielding equivalent results for the unaffected group compared with the other groups. Although this treatment group had the equivalent T1 MFQ to the disorder-only group (coefficient −0·94, 95% CI −3·81 to 1·93; p=0·52) as in the primary analyses, the two groups did not differ in their rate of change over time (linear coefficient −0·68, −2·07 to 0·70; p=0·33; quadratic coefficient −0·27, −0·74 to 0·20; p=0·26). Results did not differ significantly with propensity score weighted GLMs (table 2, appendix pp 7, 8).

−3·81 to 1·93; p=0·52) as in the primary analyses, the two groups did not differ in their rate of change over time (linear coefficient −0·68, −2·07 to 0·70; p=0·33; quadratic coefficient −0·27, −0·74 to 0·20; p=0·26). Results did not differ significantly with propensity score weighted GLMs (table 2, appendix pp 7, 8). Discussion To our knowledge, this study is the first in adolescents to support the role of contact with mental health services in improving mental health by late adolescence, while addressing non-randomisation and attrition. Four similar studies9, 10, 11, 12 did not address these issues; only one study9 adequately controlled for confounding variables, and one other study10 showed significant effects that were clinically relevant. Two studies11, 12 only assessed specialist mental health services, and one study12 reported effects of services only if eight or more sessions were attended. In the present study, we considered mental health services from all sectors irrespective of treatment length, we multiply imputed missing data, used propensity score weighting to adjust for participants' initial likelihood to access services, and data yielded clinically relevant results robust to a wide range of confounds. Contact with mental health services appeared to be of such value that after 3 years the levels of depressive symptoms of service users with a mental disorder were similar to those of unaffected individuals. Among adolescents with a mental disorder at age 14 years, the odds of those without past-year contact with mental health services having clinical depression by age 17 years were more than seven times greater than for service users who had been similarly depressed at baseline. Recruitment of participants from the general population, who vary in diagnosis type, severity, and treatment type, and the absence of strict inclusion criteria as in randomised controlled trials also increases the external validity of our study, especially for public mental health and policy makers in the field of community and specialised youth services.

general population, who vary in diagnosis type, severity, and treatment type, and the absence of strict inclusion criteria as in randomised controlled trials also increases the external validity of our study, especially for public mental health and policy makers in the field of community and specialised youth services. Our findings are in contrast with the null8 or negative7 association of mental health services reported with longitudinal total emotional and behavioural problems, with no diagnostic threshold. These studies defined mental health services in a similar manner to the present study; one study7 implemented propensity matching to address the absence of randomisation. However, measurement of total problems irrespective of clinical typology might mask potential influences of mental health services on emotional or internalising symptoms. Previous null findings can also be explained by a disregard to service need. The present study's findings became non-significant when all users of mental health services were included in the treatment group irrespective of disorder. This outcome underscores the importance of assessment, and supports training of service referrers (eg, in primary care or schools) in the presenting features of mental disorders, to increase the proportion of referrals of adolescents with a clear need who could be more responsive to treatment. Our findings suggest that adolescents accessing mental health services without a mental disorder might be less antisocial, but with fewer symptoms they could be less likely to improve from treatment. Future work should further elucidate this group.

dolescents with a clear need who could be more responsive to treatment. Our findings suggest that adolescents accessing mental health services without a mental disorder might be less antisocial, but with fewer symptoms they could be less likely to improve from treatment. Future work should further elucidate this group. Our study has some limitations. First, verification of the self-report of mental health service use against medical records would have been beneficial; however, findings are supported by high caregiver–adolescent agreement and similar proportions reported in comparable studies in other countries—eg, in adolescents with a DSM-IV diagnosis, 34–56% had past-year contact with any mental health services and 19–25% had contact with specialist mental health services (appendix p 1); the proportions in our study were 38% and 22%, respectively (appendix p 11). Second, heterogeneous treatment makes speculation about a mechanism for improvement difficult. However, common features across treatments could have a role; for example, listening, advice giving, problem solving, being non-judgmental, and being supportive. Larger studies assessing service use separately by treatment type might reveal relative effectiveness, to aid policy makers in determining which services to support. Third, sample size prohibited a focus on participants with a depressive diagnosis; thus, we included adolescents with any DSM diagnosis. However, because adolescents without depression are less likely to show change in depression related to service contact, inclusion of all diagnoses biases the findings to the null. Furthermore, because of numbers of participants, we could not do analyses by varying treatment lengths. However, the intention-to-treat assumption also biases findings to the null; therefore, it is noteworthy that an effect of service use was found. Finally, although addition of covariates and propensity score weighting helped us to address confounding variables, our study had no pretreatment baseline. A larger study with more longitudinal assessments could allow analysis of adolescents initiating service use in a naturalistic setting.12

of service use was found. Finally, although addition of covariates and propensity score weighting helped us to address confounding variables, our study had no pretreatment baseline. A larger study with more longitudinal assessments could allow analysis of adolescents initiating service use in a naturalistic setting.12 Although our findings are an encouragement to policy makers and commissioners that CAMHS helps to improve mental health, such findings cannot be cause for complacency. Figures published in 2015 show that National Health Service (NHS) spending on children's mental health services in the UK has fallen by 5·4% in real terms since 2010 (£41 million), despite an increase in demand.27 Audits have shown a resultant increase in referrals and waiting times; providers report increasingly complex and severe presenting problems, associated with longer stays in inpatient facilities.28 The present study occurred in a cohort before these NHS cuts, illustrating to UK policy makers the importance of increasing availability of CAMHS to at least the 2010 levels. Globally, in high-income countries, total mental health spending represents no more than 6% of governmental health expenditures; in many other countries, this figure is less than 1%,29 despite mental disorders being one of the leading causes of non-communicable disease burden worldwide.30 More studies assessing the effectiveness of CAMHS are needed28 for children's mental health to compete for government funds.

governmental health expenditures; in many other countries, this figure is less than 1%,29 despite mental disorders being one of the leading causes of non-communicable disease burden worldwide.30 More studies assessing the effectiveness of CAMHS are needed28 for children's mental health to compete for government funds. When mental health services are ramped-up, care needs to be taken to reach individuals with mental health needs who would typically not access services, comprising more than 60% of those with a mental disorder in our sample. This approach could include increasing community-based services, and ensuring a clear access point to mental health services, such as a designated individual in every school and primary care practice.28 Focused training of such individuals in identification of mental disorders could help to prioritise access to mental health services for young people with a clearly defined need, to the betterment of their mental health and wellbeing. Supplementary Material Supplementary appendix Acknowledgments This study was funded by The Wellcome Trust (grant number 074296), and the National Institute for Health Research Collaboration for Leadership in Applied Health Research & Care for Cambridgeshire and Peterborough. We thank the participants, and the schools and research assistants who helped with recruitment and data collection.

y was funded by The Wellcome Trust (grant number 074296), and the National Institute for Health Research Collaboration for Leadership in Applied Health Research & Care for Cambridgeshire and Peterborough. We thank the participants, and the schools and research assistants who helped with recruitment and data collection. Contributors IMG, PBJ, and TJC conceived and designed the ROOTS study. VJD, PBJ, TJC, and IMG organised the conduct of, and carried out the ROOTS study (including acquiring study data). SASN coded health services data, and SASN and VJD resolved any queries. SASN analysed the data, and SASN, IMG, PBJ, and VJD interpreted the data. SASN drafted the manuscript, and IMG, PBJ, and VJD criticised the manuscript for intellectual content. All authors have read and approved the final version of the manuscript. SASN and IMG are the guarantors for the study. Declaration of interests We declare no competing interests. Figure Adolescent MFQ scores by T1 current mental disorder and past-year contact with mental health services Disorder and services variable; imputed and adjusted results. Error bars represent SDs. Adjustments made as for table 1. MFQ=Mood and Feelings Questionnaire. T1=timepoint 1 (age 14·5 years). Table 1 Longitudinal change in MFQ by current mental disorder and past-year contact with mental health services at T1

Figure Adolescent MFQ scores by T1 current mental disorder and past-year contact with mental health services Disorder and services variable; imputed and adjusted results. Error bars represent SDs. Adjustments made as for table 1. MFQ=Mood and Feelings Questionnaire. T1=timepoint 1 (age 14·5 years). Table 1 Longitudinal change in MFQ by current mental disorder and past-year contact with mental health services at T1 Imputed sample Complete case sample n Coefficient (95% CI) p value n Coefficient (95% CI) p value MFQ all timepoints Main effects Disorder and services variable 3302 1·10 (0·47 to 1·72) 0·0011 2469 1·56 (0·95 to 2·17) <0·0001 Age (linear) 3302 −0·11 (−0·34 to 0·12) 0·34 2469 −0·24 (−0·45 to −0·02) 0·032 Age2 (quadratic) 3302 0·05 (−0·23 to 0·34) 0·72 2469 −0·36 (−0·62 to −0·10) 0·0075 Disorder and services variable × age Unaffected vs disorder only 3302 −1·00 (−1·92 to −0·08) 0·034 2469 −0·34 (−1·25 to 0·57) 0·46 Unaffected vs disorder and services 3302 −2·68 (−3·96 to −1·40) <0·0001 2469 −2·89 (−4·12 to −1·66) <0·0001 Disorder only vs disorder and services 3302 −1·68 (−3·22 to −0·14) 0·033 2469 −2·54 (−4·04 to −1·04) <0·0001 Disorder and services variable × age2 Unaffected vs disorder only 3302 −0·28 (−0·58 to 0·027) 0·074 2469 −0·08 (−0·38 to 0·22) 0·60 Unaffected vs disorder and services 3302 −0·83 (−1·26 to −0·41) <0·0001 2469 −0·99 (−1·39 to −0·58) <0·0001 Disorder only vs disorder and services 3302 −0·56 (−1·08 to −0·03) 0·037 2469 −0·91 (−1·40 to −0·41) <0·0001 Categorical analysis of age Unaffected T1–2 2965 0·02 (−0·70 to 0·74) 0·96 2257 0·59 (−0·07 to 1·24) 0·078 T2–3 2965 0·22 (−0·52 to 0·96) 0·56 2257 −0·95 (−1·63 to −0·27) 0·0063 T1–3 2965 0·24 (−0·49 to 0·96) 0·52 2257 −0·36 (−1·00 to 0·28) 0·27 Disorder only T1–2 202 −2·38 (−5·64 to 0·88) 0·15 140 −0·90 (−4·11 to 2·30) 0·58 T2–3 202 −0·81 (−4·32 to 2·69) 0·65 140 −1·42 (−4·75 to 1·89) 0·40 T1–3 202 −3·19 (−6·44 to 0·05) 0·053 140 −2·32 (−5·53 to 0·88) 0·15 Disorder and services T1–2 126 −4·07 (−9·12 to 0·98) 0·11 72 −1·29 (−6·76 to 4·21) 0·65 T2–3 126 −3·55 (−9·30 to 2·20) 0·23 72 −7·85 (−14·55 to −1·15) 0·022 T1–3 126 −7·62 (−12·82 to −2·42) 0·0037 72 −9·13 (−14·81 to −3·44) 0·0016 T1 MFQ Unaffected vs disorder only 1115 5·56 (3·58 to 7·54) <0·0001 983 5·03 (2·85 to 7·20) <0·0001 Unaffected vs disorder and services 1115 5·61 (2·95 to 8·27) <0·0001 983 7·52 (4·63 to 10·42) <0·0001 Disorder only vs disorder and services 1115 −0·05 (−3·23 to 3·13) 0·98 983 2·50 (−0·96 to 5·95) 0·16 T3 MFQ Unaffected vs disorder only 1084 2·80 (0·23 to 5·37) 0·033 769 4·20 (1·73 to 6·67) <0·0001 Unaffected vs disorde

affected vs disorder and services 1115 5·61 (2·95 to 8·27) <0·0001 983 7·52 (4·63 to 10·42) <0·0001 Disorder only vs disorder and services 1115 −0·05 (−3·23 to 3·13) 0·98 983 2·50 (−0·96 to 5·95) 0·16 T3 MFQ Unaffected vs disorder only 1084 2·80 (0·23 to 5·37) 0·033 769 4·20 (1·73 to 6·67) <0·0001 Unaffected vs disorde r and services 1084 −1·94 (−5·41 to 1·53) 0·27 769 −1·20 (−4·67 to 2·27) 0·50 Disorder only vs disorder and services 1084 −4·74 (−8·80 to −0·68) 0·022 769 −5·40 (−9·47 to −1·34) 0·0085 Data were adjusted as follows: gender, sociodemographics (ethnic origin, Index of Multiple Deprivation, adolescent living with biological parents), environmental factors (number of stressful life events in the past year, current family dysfunction and friendships, any family-focused adversities by T1), and mental health factors (any past Schedule for Affective Disorders and Schizophrenia for School-Age Children diagnosis, any mental health services after T1, any emotional problems in a family member [past 3 years or present], current antisocial traits). Variables not included were any mental health service referral age 0–13 years (p=0·19 in base model) and pubertal status (not a true confounder as p>0·10 and ρ<0·10 with predictor). MFQ=Mood and Feelings Questionnaire. T1=timepoint 1 (age 14·5 years). T2=timepoint 2 (age 16 years). T3=timepoint 3 (age 17·5 years). Table 2 MFQ clinical cutoff point at T3 predicted by propensity score weighted mental health service contact at T1

Variables not included were any mental health service referral age 0–13 years (p=0·19 in base model) and pubertal status (not a true confounder as p>0·10 and ρ<0·10 with predictor). MFQ=Mood and Feelings Questionnaire. T1=timepoint 1 (age 14·5 years). T2=timepoint 2 (age 16 years). T3=timepoint 3 (age 17·5 years). Table 2 MFQ clinical cutoff point at T3 predicted by propensity score weighted mental health service contact at T1 Propensity score weighted only* Propensity score weighted and adjusted for post-baseline confounds Post-baseline confounds OR (95% CI) p value OR (95% CI) p value Adolescents with a T1 mental disorder: service contact vs none Full propensity score sample (n=119) 3·70 (1·40–9·82) 0·0086 5·23 (1·47–18·63) 0·011 T2 MFQ; T3 family dysfunction, stressful life events Common support sample (n=98) 4·36 (1·41–13·47) 0·011 7·38 (1·73–31·50) 0·0069 T2 MFQ; T3 stressful life events, family dysfunction, living with biological parents All with T1 mental health service contact vs T1 mental disorder but no services Full propensity score sample (n=134) 1·78 (0·81–3·92) 0·15 2·41 (0·92–6·32) 0·073 T1 MFQ;† mental health service contact after T1; T2 friendships; T3 stressful life events, living with biological parents Common support sample (n=94) 2·36 (0·93–6·02) 0·072 2·65 (0·88–7·97) 0·085 T1 MFQ; mental health service contact after T1; T3 stressful life events, living with biological parents, family dysfunction OR=odds ratio. T1=timepoint 1 (age 14·5 years). T2=timepoint 2 (age 16 years). MFQ=Mood and Feelings Questionnaire. T3=timepoint 3 (age 17·5 years).

2·36 (0·93–6·02) 0·072 2·65 (0·88–7·97) 0·085 T1 MFQ; mental health service contact after T1; T3 stressful life events, living with biological parents, family dysfunction OR=odds ratio. T1=timepoint 1 (age 14·5 years). T2=timepoint 2 (age 16 years). MFQ=Mood and Feelings Questionnaire. T3=timepoint 3 (age 17·5 years). * Variables used in the propensity score model are ethnic origin, gender, pubertal status, mental health referrals aged 0–13 years, past Schedule for Affective Disorders and Schizophrenia for School-Age Children diagnosis, current behavioural diagnosis, and environmental factors (current friendships and family dysfunction, past-year stressful life events). † T1 MFQ was used if more strongly related to predictor and outcome than T2 MFQ.

highlighted by an American Psychiatric Association taskforce.5 To address the need for a scalable and valid tool to assess violence risk in patients with schizophrenia spectrum and bipolar disorder, we describe the derivation of a score based on routinely collected factors and present findings from external validation. Research in context Evidence before this study

highlighted by an American Psychiatric Association taskforce.5 To address the need for a scalable and valid tool to assess violence risk in patients with schizophrenia spectrum and bipolar disorder, we describe the derivation of a score based on routinely collected factors and present findings from external validation. Research in context Evidence before this study We searched MEDLINE from Jan 1, 1946, until Feb 6, 2017, with no language or date restrictions, for systematic reviews comparing violence risk assessment tools in individuals with a diagnosis of severe mental illness (schizophrenia, schizophrenia spectrum disorder, bipolar disorder, and other psychotic illnesses). We used the following search terms: “violen*” AND (“risk” OR “assess*” OR “predict*” OR “tool*” OR “instrument*”) AND (“mental*” OR “psychiatr*”) AND “systematic review”. We identified two systematic reviews of violence risk instruments in psychosis. The first review, published in 2011, was based on ten commonly used tools for violence risk in psychiatric populations and identified only two instruments validated in 861 patients with psychosis. These two instruments reported areas under the curves that ranged from 0·60 to 0·77, with little other information about performance. A second review, published in 2016, was based on violence risk tools in Chinese patients with psychiatric disorders, and the authors found typically poor-to-moderate performance, with areas under the curves of between 0·67 and 0·72 using tools developed in high-income countries. Current thinking in the field accepts that group data can be informative in decision making about individual cases and separates risk assessment from risk reduction or management. A scalable approach to violence risk assessment (Oxford Risk of Recidivism [OxRec] tool) has been published for released prisoners, which includes modifiable risk factors.

Introduction Major depressive disorder (MDD) typically follows a recurrent course 1. On average, individuals with a history of depression who respond to treatment have a 30-50% chance of relapse within one year2 and they will have five to nine separate episodes in their lifetime3. The risk of relapse is reduced by maintenance interventions including pharmacotherapy4 or psychosocial treatments5. Clinical trials evaluating relapse prevention approaches generally attempt to reduce the proportion of patients who relapse within a pre-determined time period (i.e., 4-6 months), where relapse is defined as surpassing a cut-point on an aggregate severity scale (i.e., Hamilton Depression Scale (HAMD) score ≥ 14). However, it has been noted that this transformation of continuous data to categorical data (i.e. “relapse” or “non-relapse”) can amplify small mean differences, which may obscure the evaluation of the clinical importance of therapeutic interventions6-8.

everity scale (i.e., Hamilton Depression Scale (HAMD) score ≥ 14). However, it has been noted that this transformation of continuous data to categorical data (i.e. “relapse” or “non-relapse”) can amplify small mean differences, which may obscure the evaluation of the clinical importance of therapeutic interventions6-8. With this in mind, there has been a growing interest in using trajectory-based approaches to analyze clinical trial data, particularly in trials attempting to produce an initial remission of depression symptoms6-9. Trajectory-based models (e.g. latent class models10, and growth mixture models11) capture heterogeneity in the development of clinical outcomes during an intervention, and this more sensitive approach can result in trial outcomes that differ from traditional endpoint measures8,12. Additionally, they enable the identification of distinct classes of time-dependent treatment responses and the evaluation of treatment effects upon trajectory membership. This approach has identified distinct classes of antidepressant response trajectory, including rapid or gradual improvement8, transient improvement followed by symptom worsening9, or “non-responders” who do more poorly on medication than placebo6.

nses and the evaluation of treatment effects upon trajectory membership. This approach has identified distinct classes of antidepressant response trajectory, including rapid or gradual improvement8, transient improvement followed by symptom worsening9, or “non-responders” who do more poorly on medication than placebo6. Far less is known, however, about trajectories of relapse for patients who initially responded to treatment and who either continued or discontinued medication treatment. To study this issue, we applied growth mixture modeling to identify distinct trajectories of HAMD scores using individual patient level data pooled from four randomized double-blind placebo-controlled discontinuation trials of duloxetine and fluoxetine. In particular, we explored whether similar or different trajectory classes exist for patients who continued active treatment or who discontinued active treatment, and tested whether there were clinical predictors of trajectory class membership. Applied in this context, these methods provided new insights into the stability of clinical response and the trajectory of relapse to depression.

sses exist for patients who continued active treatment or who discontinued active treatment, and tested whether there were clinical predictors of trajectory class membership. Applied in this context, these methods provided new insights into the stability of clinical response and the trajectory of relapse to depression. Methods Sample We analyzed data from four randomized, multicenter, double-blind, placebo-controlled discontinuation clinical trials of duloxetine and fluoxetine for MDD conducted by Eli Lilly prior to 2012. Table 1 describes the studies, arms, sample sizes and duration. Four different protocols were followed (protocol identifiers HCIZ, HCEX, HMBC and HMDI). All studies incorporated an open-label acute treatment phase with either duloxetine or fluoxetine and a double-blind discontinuation phase where patients continued their medication or received placebo. Two of the studies (HCIZ and HMBC) had an optional rescue phase that was not included in this analysis. Flow-charts of the protocols and a summary of inclusion/exclusion criteria are included in the supplemental materials (Figures S1-S4). Results from time-to-relapse analyses are published elsewhere13-16 (Table S1). We modelled trajectories of relapse up to 26 weeks during double-blind treatment. Data were aligned so that the following time points were used: weeks 0, 2, 4, 10, 16, 22 and 26 (Table S2).

ccepts that group data can be informative in decision making about individual cases and separates risk assessment from risk reduction or management. A scalable approach to violence risk assessment (Oxford Risk of Recidivism [OxRec] tool) has been published for released prisoners, which includes modifiable risk factors. Added value of this study We have developed a 16-item tool using mostly routinely collected information, with good measures of calibration and discrimination (c-index of 0·89 [95% CI 0·85–0·93] in external validation), which has been translated into a freely available online calculator. The tool should be used solely by clinicians and has specific questions about previous diagnoses and treatments. Implications of all the available evidence Violence risk assessment with use of an online tool can serve as an adjunct to clinical decision making in general adult psychiatry, accurately identify patients at low risk of violent offending, and help inform decisions about additional risk management. General adult psychiatric services could use the tool as part of their routine violence risk assessment for new patients rather than local tools, which have no evidence base.

uded in the supplemental materials (Figures S1-S4). Results from time-to-relapse analyses are published elsewhere13-16 (Table S1). We modelled trajectories of relapse up to 26 weeks during double-blind treatment. Data were aligned so that the following time points were used: weeks 0, 2, 4, 10, 16, 22 and 26 (Table S2). Statistical analysis methods The outcome variable was total score on the 17-item HAMD scale. We used growth mixture modeling11 to identify distinct trajectories of HAMD scores during treatment discontinuation. We first fitted models to the entire sample and then fitted separate models to the placebo and active arms separately. The latter analyses were used to evaluate whether different classes would emerge for subjects in the active arms and in the placebo arms. We considered linear and quadratic trends over time and up to four trajectory classes. The selection of the best model was based on the Schwartz-Bayesian Information Criterion (BIC) and on the Lo-Mendell-Rubin likelihood ratio test (LMR)17. The LMR test compares the fit of a model with two or more classes to a model with one fewer class in order to identify the optimal number of classes. We only considered models where the smallest class had at least 5% of the total subjects so that the resulting model would be meaningful clinically and stable numerically. Classification confidence was assessed using the entropy value ranging between 0 and 1, with higher values corresponding to higher confidence in latent class assignments.

els where the smallest class had at least 5% of the total subjects so that the resulting model would be meaningful clinically and stable numerically. Classification confidence was assessed using the entropy value ranging between 0 and 1, with higher values corresponding to higher confidence in latent class assignments. To evaluate whether the resulting trajectories were consistent across the different trials, we also performed separate trajectory analyses by protocol. In this secondary analysis, treatment was included as a predictor in the entire sample and by protocol in order to assess whether there were significant treatment effects on trajectory membership. Trajectories during discontinuation were classified as “relapse” vs. “non-relapse”. We assessed the association between the most likely trajectory classification of the individuals and a simple categorical indicator of relapse (HAMD ≥ 14 at the last available assessment point) using Fisher's exact tests and conditional probabilities.

ring discontinuation were classified as “relapse” vs. “non-relapse”. We assessed the association between the most likely trajectory classification of the individuals and a simple categorical indicator of relapse (HAMD ≥ 14 at the last available assessment point) using Fisher's exact tests and conditional probabilities. Weighted logistic regression was performed to assess the effects of treatment (during open-label and during the double-blind discontinuation phase), length of time with clinical response and subject characteristics on membership in a particular class. Study protocol was not included because it was confounded with treatment and is not useful as a predictor outside of these data. Interactions between treatment and the covariates were considered in order to assess potential moderating effects, but were dropped from the final model because they were not statistically significant. We calculated the length of time with clinical response as the number of weeks between randomization and time when HAMD score fell below 10 during the open-label phase. Other characteristics included sex, age, age of onset of first episode, number of previous episodes (0, 1 or 2, 3 or 4, 5 or more, missing) and CGI-severity score at randomization to discontinuation treatment. The weights were the posterior probabilities of membership in the assigned class. The association of each predictor with trajectory membership was also tested one at a time using t-tests, chi-square tests or Fisher's exact tests. Odds ratios and 95% confidence intervals were used to estimate effect sizes for the different predictors.

e the posterior probabilities of membership in the assigned class. The association of each predictor with trajectory membership was also tested one at a time using t-tests, chi-square tests or Fisher's exact tests. Odds ratios and 95% confidence intervals were used to estimate effect sizes for the different predictors. We also performed weighted logistic regression with the same predictors but with the simple clinical definition of relapse (HAMD score of 14 or higher) in order to assess the robustness of predictors of relapse to the definition of relapse. Identification of latent trajectory classes was performed using MPlus9 and all other analyses were conducted in SAS. Results In the entire sample, as well as in the samples receiving active medication and placebo during the discontinuation phase, we selected the models with three trajectory classes (Table 2). The models with three trajectory classes fit better according to both the BIC and the LMR statistic than the models with fewer classes in all analyses. Models with more than three classes could not be estimated reliably (i.e., the best-likelihood value could not replicated, the estimated variance-covariance matrix in one or more classes was not positive definite or the number of subjects per class was less than 5%) and hence are not presented. Separate analyses of the studies also identified three trajectory classes with similar shapes over time (Figures S5-S8).

alue could not replicated, the estimated variance-covariance matrix in one or more classes was not positive definite or the number of subjects per class was less than 5%) and hence are not presented. Separate analyses of the studies also identified three trajectory classes with similar shapes over time (Figures S5-S8). Figure 1 shows the estimated and sample means for the three trajectory classes over time for the samples on active medication and on placebo. The trajectory classes in the entire sample were very similar. The two classes on the bottom of both figure panels show flat HAMD trajectories over time well below the symptomatic range (HAMD scores below 5) with slightly more separation between the two classes on active medication than on placebo. We refer to these classes as “non-relapse” classes. They differ slightly in their mean scores but also there are more fluctuations in scores over time in the higher non-relapse class than in the lower relapse class (Figure S9). The third class shows rapidly increasing HAMD scores (to above 10) during the discontinuation phase with slightly higher scores on placebo but the shape of these trajectories in subjects on active medication and on placebo are very similar. We refer to this class as the “relapse” class. HAMD data on subjects after they meet clinical criteria for relapse in the studies are not reported, as they entered “rescue” treatment. As a result, the sample mean trajectories for the “relapse” class are somewhat below the estimated mean trajectories for the same class in all analyses. However, it is unlikely that missing data influences the reported findings substantially (specifically the separation of relapse vs. non-relapse trajectories), as growth mixture models provide valid results under the assumption that data are randomly missing and trajectory up until relapse predicts the loss of data. Sensitivity analysis using pattern-mixture models9 investigating stability of latent classes under missing not at random assumptions failed to identify stable trajectory classes.

e models provide valid results under the assumption that data are randomly missing and trajectory up until relapse predicts the loss of data. Sensitivity analysis using pattern-mixture models9 investigating stability of latent classes under missing not at random assumptions failed to identify stable trajectory classes. The estimated probability of membership in the relapse class is 45.8% on placebo and 33.1% on active medication. Almost all (944 out of 947, 99.7%) of patients who were classified as non-relapsers based on the trajectory analysis with trajectories 1 and 2 combined, did not relapse according to the simpler clinical relapse criterion of a HAMD score of 14 or higher. The percentages were almost the same when calculated by treatment group: 99.5% (660 out of 663) on active medication and 100% (all 284 individuals) on placebo. More than two thirds (365 out of 515, 70.9%) of the individuals most likely to follow the “relapse” trajectory relapsed according to the simpler clinical definition of relapse, with a higher rate for individuals on placebo (75.2%, 164 out of 218) than for individuals on active medication (67.7%, 201 out of 297). Thus almost a third of individuals on active medication (32.3%, 96 out of 297) and about a quarter of the subjects on placebo (24.8%, 54 out of 218) who were following the relapse trajectory did not meet traditional clinical definitions of relapse. Those individuals had on average lower mean depression scores than those who relapsed according to both definitions (Figure S10).

2.3%, 96 out of 297) and about a quarter of the subjects on placebo (24.8%, 54 out of 218) who were following the relapse trajectory did not meet traditional clinical definitions of relapse. Those individuals had on average lower mean depression scores than those who relapsed according to both definitions (Figure S10). Univariate associations between the trajectory classes identified in the joint analysis of active and placebo arms (grouped as “relapse” vs. “non-relapse”) and treatments, study protocol and covariates are provided in Table 3. When adjusting for uncertainty in trajectory membership and other covariates, active treatment during discontinuation halved the odds of following the relapse trajectory (OR=0.47, 95% CI: (0.37, 0.61)) while female gender (OR=1.56, 95% CI: (1.23, 2.06)), shorter length of time with clinical response by 1 week (OR=1.10, 95% CI: (1.06, 1.15)) and higher CGI severity by 1 (OR=1.28, 95% CI: (1.01, 1.62)) significantly increased the odds of following the relapse trajectory. Accuracy in predicting whether a patient would be in the relapse trajectory or not was reasonable (AUC = 66%, Figure S11), especially given the small number of baseline predictors available for analysis. The results from the weighted logistic regression with simple HAMD remission definition (HAMD score of 14 or more) were very similar (see Table 4).

a patient would be in the relapse trajectory or not was reasonable (AUC = 66%, Figure S11), especially given the small number of baseline predictors available for analysis. The results from the weighted logistic regression with simple HAMD remission definition (HAMD score of 14 or more) were very similar (see Table 4). Discussion The protective effect of antidepressant medications on depressive relapse is a cornerstone of psychiatry and one that has yielded the recommendation that patients with recurrent depression remain on antidepressant treatment for the remainder of their lives18,19. This study analyzed data from four clinical trials aimed at evaluating the risk for relapse when patients who had responded to treatment with fluoxetine or duloxetine were blindly maintained on their medication or switched to placebo. The principal finding was that trajectory-based analyses revealed the same three response trajectories in patients who stayed on their medications or were switched to placebo. This suggests that there is no specific relapse signature associated with antidepressant discontinuation.

eir medication or switched to placebo. The principal finding was that trajectory-based analyses revealed the same three response trajectories in patients who stayed on their medications or were switched to placebo. This suggests that there is no specific relapse signature associated with antidepressant discontinuation. The first two trajectories we identified constituted the majority of patients, showed sustained clinical response over 26 weeks, and respected traditional symptom thresholds for remission extremely closely. Individuals in the lowest severity trajectory had low scores and low variability of scores from visit to visit. The middle severity trajectory showed more score instability and slightly higher depression scores (still in the subclinical range). Since both groups had good outcomes we did not explore differences in characteristics between them in this study. Our main focus was on the relapse trajectory of increasing depression scores, in which about 46% of patients treated with placebo and 33% of patients treated with active medication were categorized.

range). Since both groups had good outcomes we did not explore differences in characteristics between them in this study. Our main focus was on the relapse trajectory of increasing depression scores, in which about 46% of patients treated with placebo and 33% of patients treated with active medication were categorized. Within the relapse trajectory, over 70% of patients also met symptomatic criteria for relapse but close to 30% did not. Trajectories of relapse may be informative even if clinical relapse criteria are not met, since prediction of trajectory membership could occur early on and since clinical relapse criteria are somewhat arbitrary. Patients who follow a relapse trajectory but do not meet criteria may have effectively relapsed nonetheless or may be at an increased risk of relapse in the future. Although on average this group had lower depression scores than the group of individuals identified as relapsers by both the trajectory and clinical criteria in this study, the absence of longer term follow-up data precluded us from comparing their longer term outcomes. Future studies are needed to evaluate this question.

hough on average this group had lower depression scores than the group of individuals identified as relapsers by both the trajectory and clinical criteria in this study, the absence of longer term follow-up data precluded us from comparing their longer term outcomes. Future studies are needed to evaluate this question. The high rate of relapse suggests that short-term antidepressant response is not very stable and the similarity of relapse trajectories on active medication and on placebo indicates that the temporal dynamic of mood regulation is not altered by SRI treatment. Approximately a third of the patients in this study followed the relapse trajectory, which is consistent with the findings of other studies5,20. Nevertheless, this should not be interpreted as evidence to downplay the benefit of SRIs during the initial episode. Furthermore, SRIs appeared to protect against the natural tendency to relapse during maintenance, i.e. they make patients more resilient. The likelihood of relapse was also related to length of time with clinical response 21, level of residual symptoms21, and was greater for women than men22. One possible explanation for these observations suggests that the efficacy of SRI antidepressants continues to be consolidated long after initial symptom reductions have occurred. At the moment, we do not understand this consolidation process, although structural neurobiological changes might be one part of it. It is striking that short-term antidepressant response is not particularly stable but that demonstration of long-term antidepressant efficacy is not required for approval by the U.S. Food and Drug Administration (FDA). One wonders whether it would be valuable to expect evidence of long-term efficacy when new antidepressants are approved by the FDA. This concern is somewhat reduced by evidence that early antidepressant response is a relatively strong predictor of later response 23.

roval by the U.S. Food and Drug Administration (FDA). One wonders whether it would be valuable to expect evidence of long-term efficacy when new antidepressants are approved by the FDA. This concern is somewhat reduced by evidence that early antidepressant response is a relatively strong predictor of later response 23. The current study suggests that SRI antidepressants have only a modest protective effect against relapse relative to placebo, as reflected in an approximately 13% difference in the likelihood of being in the relapse trajectory. This suggests that SRI treatment by itself leaves many patients at risk and specific strategies for preventing relapse should be more widely implemented in depression treatment. In the future, one hopes that this research can be extended by identifying moderators of treatment effects. That is, to ultimately identify the type and intensity of treatment that would maximize the probability of a desired outcome for that specific patient. Until then, non-specific predictors are still useful for setting prior expectations about clinically relevant outcomes, e.g. relapse or initial treatment response 23,24. The application of machine learning methods to a much broader array of predictive markers has proven successful in other areas of psychiatry, particularly predicting treatment outcomes23,25

still useful for setting prior expectations about clinically relevant outcomes, e.g. relapse or initial treatment response 23,24. The application of machine learning methods to a much broader array of predictive markers has proven successful in other areas of psychiatry, particularly predicting treatment outcomes23,25 In light of the current data, it may be important to develop new and more cost effective psychosocial treatments to reduce depression relapse in order to ensure widespread implementation. Interpersonal psychotherapy (IPT), for example, decreases depression relapse26, but it does not appear to reduce treatment costs19 and it is less effective in preventing relapse in patients who did not respond to IPT alone, but did respond when pharmacotherapy was added27. The development of more effective pharmacologic relapse prevention strategies might also improve outcomes for patients with unipolar depression. Lithium, for example, reduces relapse for mood disorders overall, but does not show clear efficacy in preventing relapse for patients with unipolar depression28. More broadly, the strategic integration of psychotherapy and medication for relapse prevention is a critical issue for patients and for the field29, especially in avoiding common clinical problems associated with long-term antidepressant treatment30.

ficacy in preventing relapse for patients with unipolar depression28. More broadly, the strategic integration of psychotherapy and medication for relapse prevention is a critical issue for patients and for the field29, especially in avoiding common clinical problems associated with long-term antidepressant treatment30. One potential limitation of three of the four protocols is the absence of rigorous measures of antidepressant withdrawal symptoms. Withdrawal symptoms for duloxetine and fluoxetine most commonly include dizziness, nausea, headache, but may also include worsening of anxiety or depression31. Some withdrawal symptoms may persist for several weeks after antidepressant discontinuation32. Fluoxetine has among the longest half-life from among the SRI medications, i.e., up to 3 days which may protect patients from some withdrawal symptoms, while duloxetine has a much shorter elimination half-life (approximately 10 hours) and would therefore be viewed as having more potential to produce withdrawal symptoms. However, we did not observe a trajectory consistent with the emergence and abatement of withdrawal symptoms on placebo. This may be at least partially attributable to the fact that some studies tapered off antidepressant medications over several weeks, which could have further limited the impact of post-discontinuation symptoms on the results. Indeed, this makes the results more reflective of clinical practice. These findings reduce our concern that the results were substantially contaminated by the appearance of antidepressant withdrawal symptoms.

eral adult psychiatry, accurately identify patients at low risk of violent offending, and help inform decisions about additional risk management. General adult psychiatric services could use the tool as part of their routine violence risk assessment for new patients rather than local tools, which have no evidence base. Methods Study design and patients We did a cohort study of individuals aged 15–65 years with a diagnosis of severe mental illness (here defined as schizophrenia spectrum or bipolar disorder) through linkage of population-based registers in Sweden. We identified a cohort of 75 158 individuals with 574 018 recorded patient episodes (424 842 [74%] outpatient episodes) between Jan 1, 2001, and Dec 31, 2008. The final study cohort consisted of a single inpatient or outpatient visit for each patient, selected at random (with use of the random number generator in Stata version 12), with equal probability. We excluded repeat visits because they complicate model fitting and interpretation. The study was approved by the Regional Ethics Committee at Karolinska Institutet (2009/939-31/5).

ver several weeks, which could have further limited the impact of post-discontinuation symptoms on the results. Indeed, this makes the results more reflective of clinical practice. These findings reduce our concern that the results were substantially contaminated by the appearance of antidepressant withdrawal symptoms. This study has other limitations. Firstly, there is a potential for expectancy bias since the analyses are based on discontinuation studies which may carry some greater expectation of relapse. Secondly, we did not have data available about these patients to adjust for intercurrent major life stress, which contributes to relapse while treated with medications33. Thirdly, patients exited the study upon relapse (to move into a rescue phase), which is likely why predicted trajectories for the relapse class are slightly inflated relative to the observed mean trajectories9. Fourthly, quadratic models do not capture the curvature in the relapse trajectory very well. More complicated models such as latent basis growth mixture models could provide better fit34. Lastly, we had a limited number of predictors to relate to relapse trajectories thus there might be much stronger predictors that might be useful in reducing the probability of relapse35. In particular, other biological factors may contribute to the association between elevated CGI, gender and depression. Important future work will be to identify additional predictors of relapse, and other clinical features associated with these relapse trajectories in line with the NIMH RDoC framework, as well as eventually advancing our understanding of neurobiological mechanisms related to relapse.

levated CGI, gender and depression. Important future work will be to identify additional predictors of relapse, and other clinical features associated with these relapse trajectories in line with the NIMH RDoC framework, as well as eventually advancing our understanding of neurobiological mechanisms related to relapse. Conclusion The similarity of trajectories on active medication and on placebo suggests that there is no specific relapse signature associated with antidepressant discontinuation. The current study supports the continued prescription of SRI antidepressants to protect against relapse of depression. However, it suggests that this protective effect is less than one might have expected. Patients and providers should be prepared for the possibility that as many as one of three patients who initially respond to an antidepressant will worsen over the subsequent six months, which justifies a more widespread effort at preventing relapse in patients with unipolar major depression. Supplementary Material supplement Funding: National Institutes of Health Grants K05AA14906, 2P50AA012870, UL1RR024139, the U.S. Department of Veterans Affairs Alcohol Research Center and National Center for Post Traumatic Stress Disorder. Non-author Contributions: We acknowledge Apurva Prakash from Eli Lilly and Company and Sarah Lipsius from Inventiv Health Clinical for help with identifying the data sets and obtaining the data.

Supplementary Material supplement Funding: National Institutes of Health Grants K05AA14906, 2P50AA012870, UL1RR024139, the U.S. Department of Veterans Affairs Alcohol Research Center and National Center for Post Traumatic Stress Disorder. Non-author Contributions: We acknowledge Apurva Prakash from Eli Lilly and Company and Sarah Lipsius from Inventiv Health Clinical for help with identifying the data sets and obtaining the data. Funding/Support: We acknowledge support from the National Institute on Alcohol Abuse and Alcoholism, K05 AA 14906, the National Institute on Alcohol Abuse and Alcoholism, 2P50 AA 012870, the U.S. Department of Veterans Affairs Alcohol Research Center, National Center Post Traumatic Stress Disorder, Clinical Neurosciences Division, West Haven, CT, the CTSA Grant Number UL1 RR024139 from the National Center for Research Resources (NCRR), a component of the National Institutes of Health (NIH), and the NIH roadmap for Medical Research. The contents of the manuscript are solely the responsibility of the authors and do not necessarily represent the official view of NIAAA, NCRR, NIH or NCPTSD. The funder/sponsor supported the effort of the authors but had no role 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. Trial Registration: ClinicalTrials.gov numbers NCT00105989 (HMDI) and NCT00036309 (HMBC). Studies HCEX and HCIZ were completed prior to the requirements for registration.

Funding/Support: We acknowledge support from the National Institute on Alcohol Abuse and Alcoholism, K05 AA 14906, the National Institute on Alcohol Abuse and Alcoholism, 2P50 AA 012870, the U.S. Department of Veterans Affairs Alcohol Research Center, National Center Post Traumatic Stress Disorder, Clinical Neurosciences Division, West Haven, CT, the CTSA Grant Number UL1 RR024139 from the National Center for Research Resources (NCRR), a component of the National Institutes of Health (NIH), and the NIH roadmap for Medical Research. The contents of the manuscript are solely the responsibility of the authors and do not necessarily represent the official view of NIAAA, NCRR, NIH or NCPTSD. The funder/sponsor supported the effort of the authors but had no role 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. Trial Registration: ClinicalTrials.gov numbers NCT00105989 (HMDI) and NCT00036309 (HMBC). Studies HCEX and HCIZ were completed prior to the requirements for registration. Authors' Contributions: Ralitza Gueorguieva had access to all data in the study and takes, overall responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Krystal, Gueorguieva Acquisition, analysis, or interpretation of data: All authors Drafting of the manuscript: All authors Critical revision of the manuscript for important intellectual content: All authors Statistical Analysis: Gueorguieva

Authors' Contributions: Ralitza Gueorguieva had access to all data in the study and takes, overall responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Krystal, Gueorguieva Acquisition, analysis, or interpretation of data: All authors Drafting of the manuscript: All authors Critical revision of the manuscript for important intellectual content: All authors Statistical Analysis: Gueorguieva Obtained Funding: All authors Administrative/technical/material support: All authors

Study concept and design: Krystal, Gueorguieva Acquisition, analysis, or interpretation of data: All authors Drafting of the manuscript: All authors Critical revision of the manuscript for important intellectual content: All authors Statistical Analysis: Gueorguieva Obtained Funding: All authors Administrative/technical/material support: All authors Conflict of Interest Disclosure: Dr. Krystal has been a consultant to the following companies: Amgen, LLC, AstraZeneca Pharmaceuticals, Biogen, Biomedisyn Corporation, Forum Pharmaceuticals, Janssen Pharmaceuticals, Orsuka America Pharmaceutical, Sunovion Pharmaceuticals, Takeda Industries, Taisho Pharmaceutical Co. He is on the Scientific Advisory Board of Biohaven Pharmaceuticals, Blackthorn Therapeutics, Lohocla Research Coreportation, Luc Therapeutices, Pfizer Pharmaceuticals, TRImaran Pharma. He holds stock in ArRETT Neuroscience and Biohaven Pharmaceuticals Medical Sciences and stock options in Blackthorn Therapeutics and Luc Therapeutics. Dr. Krystal is the editor of Biological Psychiatry and has the following patents and inventions: (1) Seibyl JP, Krystal JH, Charney DS. Dopamine and noradrenergic reuptake inhibitors in treatment of schizophrenia. Patent #:5,447,948.September 5, 1995 (2) a co-inventor with Dr. Gerard Sanacora on a filed patent application by Yale University related to targeting the glutamatergic system for the treatment of neuropsychiatric disorders (PCTWO06108055A1) (3)

pamine and noradrenergic reuptake inhibitors in treatment of schizophrenia. Patent #:5,447,948.September 5, 1995 (2) a co-inventor with Dr. Gerard Sanacora on a filed patent application by Yale University related to targeting the glutamatergic system for the treatment of neuropsychiatric disorders (PCTWO06108055A1) (3) Charney D, Krystal JH, Manji H, Matthew S, Zarate C., - Intranasal Administration of Ketamine to Treat Depression United States Application No. 14/197,767 filed on March 5, 2014; United States application or PCT International application No. 14/306,382 filed on June 17, 2014 (4) Arias A, Petrakis I, Krystal JH. – Composition and methods to treat addiction. Provisional Use Patent Application no.61/973/961. April 2, 2014, filed by Yale University Office of Cooperative Research and (5) Chekroud, A., Gueorguieva, R., & Krystal, JH. “Treatment Selection for Major Depressive Disorder” [filing date 3rd June 2016, USPTO docket number Y0087.70116US00]. Provisional patent submission by Yale University. Mr Chekroud holds equity in Spring Care Inc., a startup company developing tools for treatment selection in major depression. He is lead inventor on a provisional patent submission by Yale University titled “Treatment selection for Major Depressive Disorder” [filing date 3rd June 2016, USPTO docket number Y0087.70116US00].

hekroud holds equity in Spring Care Inc., a startup company developing tools for treatment selection in major depression. He is lead inventor on a provisional patent submission by Yale University titled “Treatment selection for Major Depressive Disorder” [filing date 3rd June 2016, USPTO docket number Y0087.70116US00]. Dr. Gueorguieva discloses consulting fees for Palo Alto Health Sciences and Mathematica Policy Research and a provisional patent submission by Yale University: Chekroud, AM., Gueorguieva, R., & Krystal, JH. “Treatment Selection for Major Depressive Disorder” [filing date 3rd June 2016, USPTO docket number Y0087.70116US00]. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Figure 1 A:Trajectories of HAMD scores in the active arms. B:Trajectories of HAMD scores in the placebo arms. Table 1 Discontinuation clinical trials for patients with major depression.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Figure 1 A:Trajectories of HAMD scores in the active arms. B:Trajectories of HAMD scores in the placebo arms. Table 1 Discontinuation clinical trials for patients with major depression. Protocol Duration of acute phase open label treatment Treatment during acute phase Duration of double-blind discontinuation phase Total Sample Size Arms and sample sizes per arm during discontinuation phase HCIZ 13 weeks Fluox 20mg/day 25 weeks 501 Fluox 20mg/day 189 Fluox 90mg/week** 190 Placebo 122 HCEX 12 weeks Fluox 20mg/day 14-50 weeks (see arms) 395 Fluox 20mg QD 50 weeks 102 Fluox 20mg QD 38 weeks 100 Fluox 20mg QD 14 weeks 97 Placebo 96 HMBC 12 weeks Dulox 60mg QD 26 weeks 288 Dulox 60mg QD 136 Placebo 142 HMDI 10 + 24 weeks* Dulox 60-120mg QD 52 weeks 278 Dulox 60mg QD 64 Dulox 90mg QD 45 Dulox 120mg QD 37 Placebo 142 * Patients underwent up to 10 weeks open label acute therapy phase during which dose was optionally increased to 90 or 120mg QD in the case of non-responses. Patients who met response criteria during week 4 to 10 of the acute open-label treatment were moved directly to the 24 week open-label continuation therapy phase.

tpatient visit for each patient, selected at random (with use of the random number generator in Stata version 12), with equal probability. We excluded repeat visits because they complicate model fitting and interpretation. The study was approved by the Regional Ethics Committee at Karolinska Institutet (2009/939-31/5). Procedures We followed up each individual from the day of their patient episode until the first time that they committed a violent offence, death, emigration, or end of follow-up (12 months after their patient episode). We linked individuals to national registers to obtain information about risk factors, with unique personal identification numbers enabling accurate linkage. We obtained sociodemographic factors and information about previous violent crime conviction, psychiatric diagnoses, and dispensed medication and identified parents and siblings of patients to extract historical information (ie, before the current patient episode; appendix pp 1–2).

* Patients underwent up to 10 weeks open label acute therapy phase during which dose was optionally increased to 90 or 120mg QD in the case of non-responses. Patients who met response criteria during week 4 to 10 of the acute open-label treatment were moved directly to the 24 week open-label continuation therapy phase. ** Enteric coated fluoxetine 90mg once a week. Table 2 Results from model selection for the entire sample and for the subsamples of subjects on active medication and on placebo. Likelihood BIC Entropy Lo-Mendell-Rubin test Proportion of Individuals In Class* 2log-lik. P-value 1 2 3 4 All arms 1-class model -22782.46 45695.13 1.00 -- -- -- 2-class model -19695.07 39499.45 0.82 327.46 <.0001 0.56 0.44 -- -- 3-class model -19363.41 38879.85 0.78 672.5 <0.0001 0.44 0.37 0.19 -- Only placebo arms 1-class model -7024.31 14092.15 1.00 -- -- -- 2-class model -6535.92 13114.47 0.83 954.84 <.0001 0.51 0.49 -- -- 3-class model -6445.19 13020.96 0.79 177.40 0.05 0.46 0.39 0.15 -- Only active arms 1-class model -14176.91 28401.88 1.00 -- -- -- 2-class model -12992.38 26101.49 0.77 1795.52 <0.0001 0.54 0.46 -- -- 3-class model -12883.43 25911.06 0.78 486.51 0.0002 0.47 0.33 0.20 -- * Classes are ordered by class size and do not necessarily correspond to one another across samples. Four-class models could not be reliably estimated. Table 3 “Trajectory relapsers” and “trajectory non-relapsers” by treatment, study and baseline characteristics.

Likelihood BIC Entropy Lo-Mendell-Rubin test Proportion of Individuals In Class* 2log-lik. P-value 1 2 3 4 All arms 1-class model -22782.46 45695.13 1.00 -- -- -- 2-class model -19695.07 39499.45 0.82 327.46 <.0001 0.56 0.44 -- -- 3-class model -19363.41 38879.85 0.78 672.5 <0.0001 0.44 0.37 0.19 -- Only placebo arms 1-class model -7024.31 14092.15 1.00 -- -- -- 2-class model -6535.92 13114.47 0.83 954.84 <.0001 0.51 0.49 -- -- 3-class model -6445.19 13020.96 0.79 177.40 0.05 0.46 0.39 0.15 -- Only active arms 1-class model -14176.91 28401.88 1.00 -- -- -- 2-class model -12992.38 26101.49 0.77 1795.52 <0.0001 0.54 0.46 -- -- 3-class model -12883.43 25911.06 0.78 486.51 0.0002 0.47 0.33 0.20 -- * Classes are ordered by class size and do not necessarily correspond to one another across samples. Four-class models could not be reliably estimated. Table 3 “Trajectory relapsers” and “trajectory non-relapsers” by treatment, study and baseline characteristics. “Relapsers” (N=515, Row %=35.2%) “Non-relapsers” (N=947, Row %=64.8%) Total (N=1462) Drug during open label treatment .004 Duloxetine 174 (30.7%) 392 (69.3%) Fluoxetine 341 (38.1%) 555 (61.9%) Protocol <.0001 HCIZ 214 (42.7%) 150 (57.3%) HCEX 127 (32.2%) 86 (67.9%) HMBC 120 (43.2%) 90 (56.8%) HMDI 54 (18.8%) 155 (81.3%) Drug during discontinuation <.0001 Active 297 (30.9%) 663 (69.1%) Placebo 218 (43.4%) 284 (56.6%) Gender 0.0005 Female 390 (38.1%) 634 (61.9%) Male 125 (28.5%) 313 (71.5%) Number of previous episodes 0.0002 0 39 (32.2%) 82 (67.8%) 1-2 136 (37.5%) 227 (62.5%) 3-4 111 (27.1%) 298 (72.9%) 5 or more 120 (37.3%) 202 (62.7%) missing 109 (44.1%) 361 (55.9%) Baseline CGI 0.09 1 311 (33.3%) 623 (66.70%) 2 202 (38.9%) 318 (61.15%) 3 2 (28.6%) 5 (71.43%) Quantitative predictors Mean (SD) Mean (SD) Length of time with clinical response 6.50 (3.17) 7.58 (3.36) <.0001 Age 42.18 (11.27) 43.53 (12.08) 0.04 Age of onset 21.48 (14.94) 22.99 (17.01) 0.08 Table 4 Odds ratios and 95% confidence intervals for the effects of predictors on relapse trajectory membership and relapse defined as HAMD score of 14 or more.

an (SD) Length of time with clinical response 6.50 (3.17) 7.58 (3.36) <.0001 Age 42.18 (11.27) 43.53 (12.08) 0.04 Age of onset 21.48 (14.94) 22.99 (17.01) 0.08 Table 4 Odds ratios and 95% confidence intervals for the effects of predictors on relapse trajectory membership and relapse defined as HAMD score of 14 or more. Relapse trajectory HAMD score of 14 or more Categorical predictors Drug during discontinuation Active vs. Placebo 0.47 (0.37, 0.61) 0.45 (0.35, 0.59) Drug during open label treatment Duloxetine vs. Fluoxetine 0.88 (0.64, 1.21) 0.83 (0.59, 1.18) Gender Female vs. Male 1.59 (1.23, 2.06) 1.71 (1.28, 2.29) Number of previous episodes 1-2 vs. 0 1.28 (0.78, 2.10) 1.33 (0.78, 2.28) 3-4 vs. 0 1.00 (0.60, 1.67) 1.06 (0.61, 1.87) 5 or more vs. 0 1.40 (0.86, 2.29) 1.34 (0.78, 2.30) Missing vs. 0 1.86 (1.11, 3.12) 1.58 (0.90, 2.77) Baseline CGI 2 or 3 vs. 1 1.30 (1.02, 1.66) 1.34 (1.03, 1.74) Quantitative predictors Shorter vs. longer duration of response by one week 1.10 (1.06, 1.15) 1.12 (1.07, 1.17) Older vs. younger age by 1 year 1.00 (0.99, 1.00) 1.00 (0.99, 1.01) Older vs. younger age of onset by 1 year 1.00 (0.99, 1.01) 1.00 (0.99, 1.01) Research in Context Evidence before this study Discontinuation trials, in which patients are randomized to stay on an effective treatment or to blindly discontinue this treatment, provide unique insights into the stability of antidepressant response. We searched PubMed from inception to Jan 10, 2017 with the terms (“depression” OR “major depressive disorder”) AND “discontinuation” AND “trial” in any field, with no language restrictions. We retrieved and scanned 833 articles, then focused on the 346 articles in which (“depression” OR “major depressive disorder”) was in the title. All articles that we deemed not to be relevant on the basis of their titles were excluded. Abstracts of the remaining articles were reviewed to identify potentially relevant articles, and, on the basis of this selection, we read full-text articles.

n which (“depression” OR “major depressive disorder”) was in the title. All articles that we deemed not to be relevant on the basis of their titles were excluded. Abstracts of the remaining articles were reviewed to identify potentially relevant articles, and, on the basis of this selection, we read full-text articles. Epidemiological and clinical evidence indicates that major depressive disorder (MDD) typically follows a recurrent course, with one-third to one-half of patients relapsing within one year of discontinuation. Clinical trials that examine relapse prevention approaches generally attempt to reduce the proportion of patients who relapse within a pre-determined time period (i.e., 4-6 months), where relapse is defined as surpassing a cut-point on an aggregate severity scale (i.e., Hamilton Depression Scale (HAMD) score ≥ 14). However, it has been noted that this transformation of continuous data to categorical data (i.e. “relapse” or “non-relapse”) can amplify small mean differences, which may obscure the evaluation of the clinical importance of therapeutic interventions.

e severity scale (i.e., Hamilton Depression Scale (HAMD) score ≥ 14). However, it has been noted that this transformation of continuous data to categorical data (i.e. “relapse” or “non-relapse”) can amplify small mean differences, which may obscure the evaluation of the clinical importance of therapeutic interventions. Trajectory-based models (i.e. latent class based approaches) provide a data-driven method to identify distinct classes of time-dependent treatment responses, and for evaluating the effect of treatment upon trajectory membership. This approach has identified distinct classes of antidepressant response trajectories, including rapid or gradual improvement, transient improvement followed by symptom worsening, or “non-responders” who do more poorly on medication than placebo. However, to our knowledge this approach has not been used to identify distinct trajectories during discontinuation clinical trials.

ssant response trajectories, including rapid or gradual improvement, transient improvement followed by symptom worsening, or “non-responders” who do more poorly on medication than placebo. However, to our knowledge this approach has not been used to identify distinct trajectories during discontinuation clinical trials. Added value of this study The objective of the study was to estimate trajectories of relapse in responders to antidepressant treatment for major depression who remained on active treatment or were switched to placebo. We analyzed individual-patient level data from four double-blind randomized placebo controlled discontinuation clinical trials of fluoxetine or duloxetine. We identified the same three patterns over time across multiple double-blind treatment continuation or discontinuation studies, i.e. we found no evidence that there are distinct trajectories of relapse during discontinuation on active medication and on placebo. Compared to a simple clinical definition of relapse (depression severity of 14 points or above on the HAMD), the trajectory approach identified individuals likely to follow a relapse trajectory who do not meet simpler criteria for relapse.

jectories of relapse during discontinuation on active medication and on placebo. Compared to a simple clinical definition of relapse (depression severity of 14 points or above on the HAMD), the trajectory approach identified individuals likely to follow a relapse trajectory who do not meet simpler criteria for relapse. The protective effect of antidepressant continuation treatment was modest, with only about 13% difference between the estimated proportion of individuals following a relapse trajectory on active medication (33%) vs. placebo (46%). In addition to treatment, female sex, shorter time with clinical response and poorer Clinical Global Impression score at baseline all predicted that a patient would be in the “relapse” trajectory. In summary, this study identified trajectories of relapse, predictors of patterns of increasing HAMD scores, and provided evidence for a statistically significant but a modest protective effect of antidepressant treatment.

l Impression score at baseline all predicted that a patient would be in the “relapse” trajectory. In summary, this study identified trajectories of relapse, predictors of patterns of increasing HAMD scores, and provided evidence for a statistically significant but a modest protective effect of antidepressant treatment. Implications of all the available evidence Clinical implications It is important for providers and consumers of depression treatment to understand the actual benefits of antidepressant treatments. For example, the STAR*D study suggested that one in three patients will have a full clinical response to their initial antidepressant. The current study suggests that about one-third of clinical responders will relapse even if they continue with the medication that produced their initial clinical response, while nearly 50% of patients will experience a return of depression symptoms if they stop their medications. However, the protective effect of continued medication is much smaller, only 13%, than one might have expected or hoped for. These findings suggest that strategies for reducing or forestalling the return of depression symptoms need to be developed and widely implemented in depression treatment.

stop their medications. However, the protective effect of continued medication is much smaller, only 13%, than one might have expected or hoped for. These findings suggest that strategies for reducing or forestalling the return of depression symptoms need to be developed and widely implemented in depression treatment. Methodological implications Latent class techniques can be used for data-driven identification of patterns of depression severity during acute and discontinuation treatment. Our results confirm previous research using simple dichotomous definitions of response that the majority of patients maintain clinical response regardless of whether they continue active medication. Future studies to identify predictors of trajectories of relapse are indicated.

Introduction The association between severe mental illness (defined here as schizophrenia, schizoaffective disorders, and bipolar affective disorders) and excess mortality has been well established worldwide. Such mortality is not restricted to suicide mortality but also encompasses mortality from natural causes, including cardiovascular and respiratory diseases.1 Few studies have assessed the nature of this inequality by ethnicity. In most studies, ethnicity is treated as a confounder or the sample size has been too small to allow stratified analysis. This concern is noteworthy because many mortality risk factors implicated in severe mental illness, such as cardiovascular disease and diabetes,2, 3 are also known to be more prevalent in some ethnic minority groups relative to white British, European, and non-Hispanic white American populations. Contrary to expectation, results from a 2015 study from the USA4 implicated lower all-cause, natural-cause, and unnatural-cause mortality in ethnic minority groups (black non-Hispanic, Hispanic, and other ethnic groups) with schizophrenia than in the white non-Hispanic group with schizophrenia. These findings are yet to be replicated outside the USA.

from a 2015 study from the USA4 implicated lower all-cause, natural-cause, and unnatural-cause mortality in ethnic minority groups (black non-Hispanic, Hispanic, and other ethnic groups) with schizophrenia than in the white non-Hispanic group with schizophrenia. These findings are yet to be replicated outside the USA. In this study, we aimed to estimate the risk of all-cause, natural-cause, and unnatural-cause mortality for black Caribbean, black African, south Asian, Irish, and white British people with severe mental illness relative to the general population of England and Wales, and to assess the association of ethnicity and other clinical and sociodemographic factors with mortality risk. The present analysis forms part of a larger investigation into ethnic minority inequalities in severe mental illness.5 Methods Study setting and participants In this longitudinal cohort study, we included individuals with a valid diagnosis of severe mental illness from the case registry of the South London and Maudsley Trust (SLaM), a large secondary care mental health trust serving roughly 1·36 million people in an ethnically diverse catchment area of London, UK.6, 7 Since 2006, electronic health records have been used for routine patient care in the Trust. The SLaM Clinical Record Interactive Search system, established in 2008, allows search and retrieval of fully anonymised patient records for secondary analysis.6, 7 Electronic health records predating its establishment have also been incorporated into the system.7 Research in context Evidence before this study

on numbers enabling accurate linkage. We obtained sociodemographic factors and information about previous violent crime conviction, psychiatric diagnoses, and dispensed medication and identified parents and siblings of patients to extract historical information (ie, before the current patient episode; appendix pp 1–2). Outcomes The primary outcome was the occurrence of any violent offending within 1 year of hospital discharge for inpatients or clinical contact with psychiatric services for outpatients (patient episode). We did not consider repeat offences by an individual within 1 year. We used conviction data because the Swedish criminal code determines that individuals are convicted as guilty regardless of mental disorder and no plea bargaining is permitted at conviction. We defined violent crime as homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), or illegal threats and harassment.

Methods Study setting and participants In this longitudinal cohort study, we included individuals with a valid diagnosis of severe mental illness from the case registry of the South London and Maudsley Trust (SLaM), a large secondary care mental health trust serving roughly 1·36 million people in an ethnically diverse catchment area of London, UK.6, 7 Since 2006, electronic health records have been used for routine patient care in the Trust. The SLaM Clinical Record Interactive Search system, established in 2008, allows search and retrieval of fully anonymised patient records for secondary analysis.6, 7 Electronic health records predating its establishment have also been incorporated into the system.7 Research in context Evidence before this study To assess the evidence for excess mortality in severe mental illness (defined here as schizophrenia, schizoaffective disorders, and bipolar affective disorders) in ethnic minority groups, we systematically searched MEDLINE, PsycINFO, and Embase from database inception to Aug 23, 2016, using the search term “((psychosis OR schizo* OR bipolar) AND (ethnic* OR race OR racial) AND mortality)” and no language restrictions. 294 abstracts were retrieved. After screening by abstract, 42 potentially relevant publications remained. Of these, 12 studies directly assessed the association of severe mental illness with mortality outcomes in individuals with schizophrenia and specifically provided detail on mortality outcomes in severe mental illness by racial or ethnic group. Only one study presented findings relating to cause-specific mortality, and most of the studies focused on suicide or all-cause mortality. With the exception of two studies, most had few individuals from ethnic minority groups, affecting the ability to report on associations for these populations. No adequately powered studies have been done outside the USA that assessed all-cause, natural-cause, and unnatural-cause mortality in ethnic minority groups with severe mental illness.

o studies, most had few individuals from ethnic minority groups, affecting the ability to report on associations for these populations. No adequately powered studies have been done outside the USA that assessed all-cause, natural-cause, and unnatural-cause mortality in ethnic minority groups with severe mental illness. Added value of this study Our findings suggest that people with severe mental illness, irrespective of ethnicity, have excess mortality compared with the general population. This observation was evident for all mortality outcomes, including suicide mortality, non-suicide unnatural-cause mortality, respiratory mortality, cardiovascular mortality, and cancer mortality. However, among people with severe mental illness, those from ethnic minority groups (black African, black Caribbean, and south Asian) had lower mortality than the reference white British population. Unlike previous work, we adjusted for the possibility that first-generation migrants might return to their country of origin prior to death (which could erroneously give the impression of so-called healthy migrant effects), since we assessed the effect of emigrations out of the cohort as a competing risk in sensitivity analyses. However, despite taking this possibility into account, observed differences in mortality persisted. Implications of all the available evidence

Our findings suggest that people with severe mental illness, irrespective of ethnicity, have excess mortality compared with the general population. This observation was evident for all mortality outcomes, including suicide mortality, non-suicide unnatural-cause mortality, respiratory mortality, cardiovascular mortality, and cancer mortality. However, among people with severe mental illness, those from ethnic minority groups (black African, black Caribbean, and south Asian) had lower mortality than the reference white British population. Unlike previous work, we adjusted for the possibility that first-generation migrants might return to their country of origin prior to death (which could erroneously give the impression of so-called healthy migrant effects), since we assessed the effect of emigrations out of the cohort as a competing risk in sensitivity analyses. However, despite taking this possibility into account, observed differences in mortality persisted. Implications of all the available evidence Excess mortality is a concern in all people with severe mental illness irrespective of ethnicity. The reduced mortality in black African, black Caribbean, and south Asian groups relative to the white British group in this UK-based population with severe mental illness deserves further investigation. Our findings are consistent with results from a US study of mortality outcomes in schizophrenia, which also indicated reduced mortality in black, Hispanic, and other ethnic groups compared with non-Hispanic white Americans for most causes of death, excluding unnatural causes. These findings might indicate differential factors that might have relevance in improving mortality outcomes in people with severe mental illness.

also indicated reduced mortality in black, Hispanic, and other ethnic groups compared with non-Hispanic white Americans for most causes of death, excluding unnatural causes. These findings might indicate differential factors that might have relevance in improving mortality outcomes in people with severe mental illness. Mental health teams within SLaM are required to assign psychiatric diagnoses according to ICD-10 for all patients.6, 8 Searches for diagnoses were done within structured fields and supplemented by a natural language-processing application developed with Generalised Architecture for Text Engineering9 to identify diagnosis of mental disorders according to ICD-10 in case notes and correspondence, including schizophrenia spectrum disorders (F2*) and bipolar disorders (F30 and F31). We included individuals older than age 15 years at the time of diagnosis who had any contact with SLaM services (including inpatient, outpatient, or emergency department contacts). Individuals with comorbid dementia before their diagnosis of severe mental illness were excluded. The observation period for the study was from Jan 1, 2007, to Dec 31, 2014. At-risk periods in the study were from the date of severe mental illness diagnosis to the date of death or emigration (whichever came first) or to the censor date (Dec 31, 2014) for individuals who were still alive.

ental illness were excluded. The observation period for the study was from Jan 1, 2007, to Dec 31, 2014. At-risk periods in the study were from the date of severe mental illness diagnosis to the date of death or emigration (whichever came first) or to the censor date (Dec 31, 2014) for individuals who were still alive. Permission to conduct secondary analysis of the Clinical Record Interactive Search system was granted by the Oxfordshire Research Ethics Committee C (reference 08/H0606/71+5). Separate approvals to examine linked mortality data with approved researcher status were obtained from the UK Health & Social Care Information Centre.

sion to conduct secondary analysis of the Clinical Record Interactive Search system was granted by the Oxfordshire Research Ethics Committee C (reference 08/H0606/71+5). Separate approvals to examine linked mortality data with approved researcher status were obtained from the UK Health & Social Care Information Centre. Measures Information on mortality for the general population of England and Wales was linked to the cohort using data from the UK Office for National Statistics with the National Health Service (NHS) number, which is a unique patient identifier for all NHS health records within the UK. The NHS unique patient identifier was also used to link to all records relating to emigrations and so-called cancelled ciphers. The cancelled cipher code is ascribed to individuals who have not consulted with a general practitioner within a 3 year period if younger than 75 years or within a 1 year period if aged 75 years or above.10, 11 Individuals are sent a letter to the last known postal address and if they do not respond within 6 months their registration is cancelled.10 Methodological work undertaken by the Office of National Statistics using deregistration data from the Longitudinal Study indicates that health authority deregistrations are a reasonable proxy for emigration, with 95% of members from the study who were deregistered on or before census day in 2001 not found in the census.10, 11 Case-tracing procedures were done until the end of the observation period. Causes of death on death certificates were classified by the Office for National Statistics according to ICD-108 and were grouped into all-cause mortality (A00–R99; U00–Y89), natural-cause mortality (A00–Q99), unnatural or external causes of mortality (U509, V01–Y89), and deaths not elsewhere classified (R00–R99). Cause-specific mortality was further categorised to include deaths from respiratory diseases (J00–J99), circulatory diseases (I00–I99), and cancers (C00–D48). Unnatural-cause mortality included deaths from suicide (X60–X84 and Y10–Y34), intentional self-harm (X60–X84), and events of undetermined intent (Y10–Y34). Unnatural-cause mortality that was not classified as suicide, self-harm, or events of undetermined intent was classified as deaths from accidents or external causes.

nnatural-cause mortality included deaths from suicide (X60–X84 and Y10–Y34), intentional self-harm (X60–X84), and events of undetermined intent (Y10–Y34). Unnatural-cause mortality that was not classified as suicide, self-harm, or events of undetermined intent was classified as deaths from accidents or external causes. We included the following demographic indicators: date of birth, sex, and marital status. Postcodes of participants were linked to area-level indicators of deprivation (Index of Multiple Deprivation) at the level of Lower Super Output Area,12 which comprises areas with a mean of roughly 1500 residents. The Index of Multiple Deprivation takes into account deprivation across multiple domains—income, employment, health, education, barriers to housing and services, living environment, and crime—with specific weightings.12 Self-ascribed ethnicity was classified according to the following categories defined by the Office for National Statistics: white British, Irish, black Caribbean, and black African. Indian, Pakistani, and Bangladeshi individuals were classed into one group (south Asians) because the number of individuals was too small for separate analyses. The ICD-10 diagnostic codes F30 and F31 were categorised as affective disorders, and all other F2* diagnoses were coded as non-affective disorders. Individuals who also had a clinician-ascribed diagnosis of ICD-10 code F10–F19 (“mental and behavioural disorders due to psychoactive substance use”) were classified as having comorbid alcohol or substance misuse.8

re categorised as affective disorders, and all other F2* diagnoses were coded as non-affective disorders. Individuals who also had a clinician-ascribed diagnosis of ICD-10 code F10–F19 (“mental and behavioural disorders due to psychoactive substance use”) were classified as having comorbid alcohol or substance misuse.8 Statistical analysis Deaths by cause and ethnicity in the cohort over the observation period were indirectly standardised by age and sex to their counterparts (resident population and deaths) from England and Wales for 2011 (mid-point of the observation period) to derive standardised mortality ratios (SMRs) with 95% CIs. For the purposes of standardisation, age was determined as the mid-point of the observation period (Jan 1, 2011) or the date of diagnosis of mental disorder if the diagnosis occurred after the mid-point. We categorised age into 10 year bands (15–24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84, and ≥85 years) corresponding to the reference population age groups. Because mortality of the standard population was given for 1 year but our target population was observed for 8 years, weights to account for the length of follow-up were derived by taking the mean observation period contributed by individuals within each corresponding age and sex band in the cohort. Each weight was then multiplied by the number of deaths recorded in each corresponding band for the standard population to provide an estimation of the expected number of deaths in the observation period.

ng the mean observation period contributed by individuals within each corresponding age and sex band in the cohort. Each weight was then multiplied by the number of deaths recorded in each corresponding band for the standard population to provide an estimation of the expected number of deaths in the observation period. We used Cox proportional hazards regression to estimate crude and adjusted hazard ratios (HRs) for the association of ethnicity and other covariates (sex, diagnosis, marital status, comorbid alcohol or substance misuse, and area-level deprivation) with mortality. Lexis expansion was used to derive the time-varying covariates of age (15–44, 45–64, and ≥65 years) and time since diagnosis (0–3, 3–7, and >7 years). Proportional hazards assumptions were checked by assessing interactions with survival time, examining Schoenfeld residual plots, and testing for a zero-slope in scaled residuals.13 Likelihood ratio tests were used to assess statistical interactions.

t individuals are convicted as guilty regardless of mental disorder and no plea bargaining is permitted at conviction. We defined violent crime as homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), or illegal threats and harassment. Statistical analysis We derived the model with logistic regression (appendix pp 3–4). On the basis of existing evidence of criminal history and sociodemographic and clinical factors,12, 13 we grouped variables a priori on the anticipated strength of association with the outcome in decreasing levels of priority (appendix p 5).14, 15 We excluded covariates with more than 30% missing data. We made an exception for the recent treatment variables, which were unavailable before 2006 only because the Prescribed Drug Register was not available: the missingness mechanism was thus known and thought to be unrelated to the missing values themselves. We imputed missing data via multiple imputation using chained equations.

–64, and ≥65 years) and time since diagnosis (0–3, 3–7, and >7 years). Proportional hazards assumptions were checked by assessing interactions with survival time, examining Schoenfeld residual plots, and testing for a zero-slope in scaled residuals.13 Likelihood ratio tests were used to assess statistical interactions. To assess associations with natural-cause and unnatural-cause mortality outcomes, we used a modified Cox regression approach taking into account competing risks.14 These models take into account the likelihood of the competing event occurring (events that remove study participants from being at risk of the event of interest—eg, through death from another cause).14 We first assessed the association of independent variables with natural-cause mortality, competing with unnatural-cause mortality risks. We next assessed the association of independent variables with unnatural-cause mortality, competing with natural-cause mortality risk. Sub-HRs with 95% CIs based on robust SE estimations were generated. Wald tests were used for hypothesis testing. First-generation ethnic minorities might migrate back to their country of origin when they are unwell or before death, which might lead to a biased under-estimation of mortality risk through numerator–denominator mismatch.15 Therefore, we did a sensitivity analysis to re-assess associations between ethnicity and all-cause mortality by using competing-risks regression, specifying emigration out of the cohort as a competing event as opposed to a censored event. Statistical analyses were done in Stata, version 12.

First-generation ethnic minorities might migrate back to their country of origin when they are unwell or before death, which might lead to a biased under-estimation of mortality risk through numerator–denominator mismatch.15 Therefore, we did a sensitivity analysis to re-assess associations between ethnicity and all-cause mortality by using competing-risks regression, specifying emigration out of the cohort as a competing event as opposed to a censored event. Statistical analyses were done in Stata, version 12. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results After excluding individuals with incomplete data and those from ethnic groups other than the ones studied here, 18 201 were included in the cohort, contributing a median follow-up of 6·36 years (IQR 3·26–9·92) and a total of 1767 deaths by the study censor date (Dec 31, 2014; figure). 5041 (27·7%) individuals had an affective diagnosis at baseline and roughly half of the cohort belonged to an ethnic minority group (table 1).

re, 18 201 were included in the cohort, contributing a median follow-up of 6·36 years (IQR 3·26–9·92) and a total of 1767 deaths by the study censor date (Dec 31, 2014; figure). 5041 (27·7%) individuals had an affective diagnosis at baseline and roughly half of the cohort belonged to an ethnic minority group (table 1). The main causes of death were from circulatory disease (including cardiovascular and cerebrovascular disease; 474 [27%] deaths), cancers (308 [17%]), respiratory disease (290 [16%]), and external causes (including suicides and non-suicide causes; 189 [11%]). Irrespective of ethnicity, SMR was elevated for all causes of death in people with severe mental illness (2·67, 95% CI 2·56–2·78; table 2). SMRs for cardiovascular and respiratory mortality were increased by 2–4 times for almost all ethnic groups and suicide by 5–10 times in all ethnic groups. SMRs for cancer mortality were also raised, albeit to a lower extent (1·45, 1·32–1·60), across the entire cohort, but south Asian people with severe mental illness had a reduced SMR for cancer mortality (0·49, 0·21–0·96; table 2). SMRs for deaths due to circulatory causes were elevated in black African people with severe mental illness (3·85, 2·71–5·31) relative to the white British group with severe mental illness (2·66, 2·38–2·96), albeit with overlapping 95% CIs.

ental illness had a reduced SMR for cancer mortality (0·49, 0·21–0·96; table 2). SMRs for deaths due to circulatory causes were elevated in black African people with severe mental illness (3·85, 2·71–5·31) relative to the white British group with severe mental illness (2·66, 2·38–2·96), albeit with overlapping 95% CIs. In Cox proportional hazards regression models using the white British group as reference, there was evidence of non-proportional hazards in the association of ethnicity with all-cause mortality (likelihood ratio test for interaction of ethnicity with time p=0·0055; χ2 test for non-zero slope in Schoenfeld residuals p=0·013; appendix pp 1–2). Proportional hazards assumptions were met when an interaction of time with ethnicity was fitted (table 3). Overall, relative to white British people with severe mental illness, mortality was reduced in black African, black Caribbean, and south Asian individuals with severe mental illness (likelihood ratio test for interaction with time since diagnosis p<0·0001); this difference reduced slightly for the black Caribbean group as time since diagnosis increased (table 3). Female sex and affective diagnoses (rather than non-affective diagnoses) were associated with a lower all-cause mortality, whereas comorbid alcohol or substance misuse and being single, divorced, widowed, or separated were associated with a 19% increased risk of all-cause mortality after adjustment for all other variables (table 4).

an exception for the recent treatment variables, which were unavailable before 2006 only because the Prescribed Drug Register was not available: the missingness mechanism was thus known and thought to be unrelated to the missing values themselves. We imputed missing data via multiple imputation using chained equations. We assessed the internal validity of the model using bootstrapping to assess its predictive accuracy.16 We used bootstrapping to create 100 samples drawn with replacement from the derivation dataset. To test the external validation of the model, we selected at random (using the sample function in R version 2.3.1) a subsample of geographical regions (the validation sample)17 on the basis of the residential geographical location of the individual at the time of diagnosis, comprising around one-fifth of the total sample, and removed it from the dataset used to fit the model (the derivation sample). We chose the number of regions for the validation sample to be large enough for a useful assessment of external validity to be made.18 The geographical regions and method for selection of the validation sample are described in the appendix (p 6).

affective diagnoses (rather than non-affective diagnoses) were associated with a lower all-cause mortality, whereas comorbid alcohol or substance misuse and being single, divorced, widowed, or separated were associated with a 19% increased risk of all-cause mortality after adjustment for all other variables (table 4). In competing risks regression models, women had a reduced risk of unnatural-cause mortality relative to men (table 5). Affective psychosis was associated with a reduced natural-cause mortality, but no difference by diagnosis was found for unnatural-cause mortality (table 5). Comorbid alcohol or substance misuse was associated with a doubling in unnatural-cause mortality (table 5). With the exception of the Irish group for natural-cause mortality, mortality was reduced for each of the ethnic minority groups (relative to white British people with severe mental illness) for both natural-cause and unnatural-cause mortality in adjusted models (table 5). In sensitivity analyses with emigrations out of the cohort as a competing risk, sub-HR estimates for ethnicity and all-cause mortality followed similar trends to those highlighted in table 3 (appendix p 3)—ie, the observed differences in mortality were not affected by emigration.

mortality in adjusted models (table 5). In sensitivity analyses with emigrations out of the cohort as a competing risk, sub-HR estimates for ethnicity and all-cause mortality followed similar trends to those highlighted in table 3 (appendix p 3)—ie, the observed differences in mortality were not affected by emigration. In a post-hoc analysis, we fitted an interaction between ethnicity and severe mental illness diagnoses (affective vs non-affective) in their association with all-cause mortality to assess the possibility that the association of ethnicity with all-cause mortality might be modified by diagnosis. We did not find any evidence in support of effect modification by ethnicity in these models (likelihood ratio test for interaction p=0·39; appendix p 4). Discussion Findings from our cohort study confirm that people with severe mental illness have an elevated mortality risk compared with the general population. Mortality from respiratory disorders and cardiovascular disease was elevated by up to 4 times and suicide by 5–10 times across almost all ethnic minority groups.

Discussion Findings from our cohort study confirm that people with severe mental illness have an elevated mortality risk compared with the general population. Mortality from respiratory disorders and cardiovascular disease was elevated by up to 4 times and suicide by 5–10 times across almost all ethnic minority groups. Among individuals with severe mental illness, black African and black Caribbean people had reduced mortality compared with white British people for all causes, natural causes, and unnatural causes of deaths. Similar trends, particularly for natural-cause mortality, were observed in the south Asian group. As time since diagnosis increased, the HRs for all-cause mortality between black Caribbean and white British individuals with severe mental illness became more similar, but the black Caribbean group continued to have a reduced mortality risk relative to the white British group by the end of the study. Other associations with mortality in our study are consistent with the wider literature; in particular, comorbid alcohol and substance misuse diagnoses were associated with a doubling in risk of unnatural-cause mortality,16 female sex was associated with reduced all-cause and unnatural-cause mortality relative to men,16, 17 affective diagnoses were associated with a reduced risk of natural-cause mortality relative to non-affective diagnoses,18 and being single, divorced, widowed, or separated was associated with an increased mortality risk.19

le sex was associated with reduced all-cause and unnatural-cause mortality relative to men,16, 17 affective diagnoses were associated with a reduced risk of natural-cause mortality relative to non-affective diagnoses,18 and being single, divorced, widowed, or separated was associated with an increased mortality risk.19 The use of a large cohort from an ethnically diverse location allowed us to assess differences in mortality outcomes in severe mental illness for each ethnic minority group. Previous studies16, 17 did not have adequate power to detect differences by ethnicity. By prospectively assessing mortality and tracing emigrations out of the cohort, we were able to assess the possibility of bias through emigration. The long follow-up of 8 years allowed a detailed assessment. Our sample was comprehensive: since we would have included everyone in contact with the mental health trust with severe mental illness, coverage within this area was likely to have been almost complete.6 Although we might have excluded people with psychosis who sought private health care, this number is likely to have been very small.6 It is of course possible that people with non-psychotic bipolar disorders were more likely than those with psychotic disorders to access private services; however, in the UK state-funded services tend to provide most health care and use of private services tends to be minimal;6 therefore, this potential source of bias was likely to have been low. The findings are generalisable insofar as the location (urban inner city) typifies areas where ethnic minority communities tend to reside.

the UK state-funded services tend to provide most health care and use of private services tends to be minimal;6 therefore, this potential source of bias was likely to have been low. The findings are generalisable insofar as the location (urban inner city) typifies areas where ethnic minority communities tend to reside. Our study had a few limitations. First, diagnoses of severe mental illness were not based on research diagnostic criteria. Racially biased diagnostic practices might have led to ethnic minority groups being more likely to receive a psychosis diagnosis that did not meet research diagnostic criteria,20 which might have led to a lower recorded mortality risk if such mis-diagnosis in ethnic minority groups meant that these individuals were likely to have had less severe mental illness. However, the direction of association observed in our study is consistent with findings from another study16 using research diagnostic criteria by clinicians blinded to the ethnicity of participants. Second, the south Asian group in our study might have masked important differences for Indian, Pakistani, and Bangladeshi individuals. Although we adjusted for area-level deprivation, indicators of individual-level socioeconomic position were not available. Additionally, future work could assess the role of mediators of premature mortality, particularly tobacco use, type 2 diabetes, obesity, and hypertension.

n, Pakistani, and Bangladeshi individuals. Although we adjusted for area-level deprivation, indicators of individual-level socioeconomic position were not available. Additionally, future work could assess the role of mediators of premature mortality, particularly tobacco use, type 2 diabetes, obesity, and hypertension. Presently, work is underway on this data source to develop natural language-processing applications to extract information on physical health from unstructured or free text,7 as well as to use database linkage with primary care records to enhance assessment of cardiovascular disease and other physical health indicators within the record.21 Other natural language-processing applications developed in the Clinical Record Interactive Search system in current research use include those ascertaining text on tobacco22 and cannabis use,23 medications for diabetes and other physical disorders,7 and more than 70 different mental health symptoms.24 These efforts have been supplemented by a range of new algorithms for ascertaining recorded body-mass index and mentions of comorbid physical disorders and use of several common illicit drugs. The role of these factors as potential mediators for premature mortality could thus be assessed in future work using this data source.

fforts have been supplemented by a range of new algorithms for ascertaining recorded body-mass index and mentions of comorbid physical disorders and use of several common illicit drugs. The role of these factors as potential mediators for premature mortality could thus be assessed in future work using this data source. Overall, differences in SMRs in our cohort reflect those noted in previous work.1, 4, 17, 25 SMRs for cardiovascular and respiratory mortality were elevated irrespective of ethnicity. SMRs for cancers are also consistent with the findings from a meta-analysis of mortality outcomes in schizophrenia.1 Among individuals with severe mental illness, the finding of a lower all-cause SMR in ethnic minority groups relative to non-minority reference groups has been reported in a study from the USA.4

hnicity. SMRs for cancers are also consistent with the findings from a meta-analysis of mortality outcomes in schizophrenia.1 Among individuals with severe mental illness, the finding of a lower all-cause SMR in ethnic minority groups relative to non-minority reference groups has been reported in a study from the USA.4 The finding of a much lower SMR for cancers in south Asian groups with severe mental illness is still consistent with wider findings from the literature1 and might also reflect a lower prevalence of cancers reported for this ethnic group compared with non-south Asians.26 Investigators of a US study of more than 1 million people with schizophrenia4 also reported lower SMR for cancer mortality in all the ethnic minority groups surveyed (black non-Hispanic [1·2, 95% CI 1·2–1·3], Hispanic [1·6, 1·4–1·7], and other non-Hispanic [1·4, 1·2–1·7]) than in white non-Hispanic individuals (2·0, 2·0–2·1). Our finding of a lower SMR for suicide in the black Caribbean group relative to the white British group with severe mental illness is broadly consistent with previous research from the UK.27, 28 Investigators of the AESOP-10 study16 used 10 year follow-up data from 557 people with first-episode psychosis and determined diagnoses at baseline by assessors blinded to the ethnicity of respondents. In this study,16 unnatural-cause mortality was lower in black and minority ethnic groups than in white British people, although 95% CIs were wide. Among people with schizophrenia, suicide risk has been reported to be lower in first-generation migrants than in native Dutch people in the Netherlands,29 as well as lower in black non-Hispanic people than in white non-Hispanic people in the USA.4

ethnic groups than in white British people, although 95% CIs were wide. Among people with schizophrenia, suicide risk has been reported to be lower in first-generation migrants than in native Dutch people in the Netherlands,29 as well as lower in black non-Hispanic people than in white non-Hispanic people in the USA.4 The apparently reduced mortality risk in ethnic minority groups with severe mental illness deserves further exploration. Social factors that increase mortality risk might be less prevalent in black African, black Caribbean, and south Asian groups with severe mental illness than in white British people with severe mental illness. For example, findings from previous studies have indicated the role of ethnic density (ie, a high proportion of people of the same ethnicity living in the same area) in reducing self-harm and suicide risks in ethnic minorities in the UK30, 31 and the Netherlands.32 Group density effects are also associated with reduced alcohol use in black Caribbean, black African, and Indian individuals living in areas of high ethnic density.33 Social support and buffering from discrimination and social isolation, as well as protective social norms, have been implicated in these effects.30, 33, 34 The catchment area for this study—similar to many other urban locations where ethnic minority communities live—is notable for being ethnically diverse and having one of the largest black Caribbean and black African communities in the UK.5, 7 Therefore, these socioenvironmental factors might have played a part in mediating mortality risk for our study cohort. By contrast, an increasing body of evidence shows an elevated cardiovascular risk (eg, type 2 diabetes) in these groups,2 and these complex interactions will need to be explored in future research.35

7 Therefore, these socioenvironmental factors might have played a part in mediating mortality risk for our study cohort. By contrast, an increasing body of evidence shows an elevated cardiovascular risk (eg, type 2 diabetes) in these groups,2 and these complex interactions will need to be explored in future research.35 Other factors might also have led to the lower mortality risk in black Caribbean, black African, and south Asian individuals with severe mental illness, relative to the white British group with severe mental illness. For example, if mortality rates were higher in these ethnic minority groups before receiving a diagnosis for a severe mental illness (and therefore their entry into the study cohort), this could have led to an artificially reduced risk of mortality relative to the white British group. Another possibility is that people who made first contact with services through either primary care or emergency departments might have had their physical health optimised, leading to improved mortality outcomes. However, findings from studies assessing ethnic minority pathways into care36, 37 have consistently indicated that ethnic minority groups with psychosis are less likely than white British individuals to make first contact through these routes—in particular, black individuals with psychosis are more likely to experience criminal justice routes into care—so this possibility seems less likely but could be explored in future work. Finally, other factors such as differences in physical health (eg, presence of type 2 diabetes38 and other cardiovascular disease indicators), differential prescribing of psychotropic medications by ethnicity (including antipsychotic medications and associated weight gain), and the presence of untreated primary health conditions (including cardiovascular disease)21 might have had a role in the differential mortality risk by ethnicity observed, and should be investigated in future work.

psychotropic medications by ethnicity (including antipsychotic medications and associated weight gain), and the presence of untreated primary health conditions (including cardiovascular disease)21 might have had a role in the differential mortality risk by ethnicity observed, and should be investigated in future work. The catchment area for the study reflects relatively recent migration trends spanning one to two generations. Therefore, a further possibility is that the trends in our study reflect migrant groups being selectively more resilient to ill-health effects. Lower suicide risk in some ethnic minority groups, with protective effects being lost in younger and second-generation migrants, has previously been noted,27, 29 leading some to implicate healthy migrant effects or the loss of acculturation health benefits over time and generation.27 Our findings extend these observations to natural-cause mortality in severe mental illness. We were unable to assess the effect of migrant or generational status, but this alongside the role of acculturation should be reviewed in future research.

or the loss of acculturation health benefits over time and generation.27 Our findings extend these observations to natural-cause mortality in severe mental illness. We were unable to assess the effect of migrant or generational status, but this alongside the role of acculturation should be reviewed in future research. Irrespective of ethnicity, people with severe mental illness have excess mortality, underlining an urgent need to address tractable causes within this group of people. Reduced mortality risk in black African, black Caribbean, and south Asian groups with severe mental illness, relative to a white British reference group with severe mental illness, might be due to several factors, including but not limited to differential socioenvironmental factors, differences in underlying physical health, and differences in the prescribing of psychotropic medications. These factors could be relevant to improving mortality outcomes in all people with severe mental illness. Supplementary Material Supplementary appendix

Irrespective of ethnicity, people with severe mental illness have excess mortality, underlining an urgent need to address tractable causes within this group of people. Reduced mortality risk in black African, black Caribbean, and south Asian groups with severe mental illness, relative to a white British reference group with severe mental illness, might be due to several factors, including but not limited to differential socioenvironmental factors, differences in underlying physical health, and differences in the prescribing of psychotropic medications. These factors could be relevant to improving mortality outcomes in all people with severe mental illness. Supplementary Material Supplementary appendix Acknowledgments We thank Michael Dewey for statistical advice. We are grateful to Hitesh Shetty for supporting data extractions for the study. JD-M and RD are Clinician Scientist Fellows funded by the UK Health Foundation working together with the UK Academy of Medical Sciences. CM is supported by a European Research Council Consolidator Award (ERC-CoG-2014—Proposal 648837, REACH). RS and C-KC are partly funded by the UK National Institute for Health Research Biomedical Research Centre at South London and Maudsley National Health Service Foundation Trust and King's College London.

al Sciences. CM is supported by a European Research Council Consolidator Award (ERC-CoG-2014—Proposal 648837, REACH). RS and C-KC are partly funded by the UK National Institute for Health Research Biomedical Research Centre at South London and Maudsley National Health Service Foundation Trust and King's College London. Contributors JD-M conceived the study and led the study design and analysis. C-KC assisted in analysis. All authors contributed to data interpretation, were involved in the drafting of the manuscript and revision for intellectual content, approved the final version prior to publication, and agree to be held accountable for all aspects of the work. JD-M is the guarantor for the data and analyses. Declaration of interests RS reports grants from Janssen and Roche and a PhD studentship funded by GlaxoSmithKline outside of the submitted work. All other authors declare no competing interests. Figure Study flow for regression models *Some individuals had missing information in more than one category. Table 1 Baseline characteristics and crude mortality rates n (%) All causes Natural causes Unnatural causes Number of deaths Crude mortality per 100 000 person years (95% CI) Number of deaths Crude mortality per 100 000 person years (95% CI) Number of deaths Crude mortality per 100 000 person years (95% CI)

*Some individuals had missing information in more than one category. Table 1 Baseline characteristics and crude mortality rates n (%) All causes Natural causes Unnatural causes Number of deaths Crude mortality per 100 000 person years (95% CI) Number of deaths Crude mortality per 100 000 person years (95% CI) Number of deaths Crude mortality per 100 000 person years (95% CI) Total cohort 18 201 1767 1439·7 (1374·1–1508·4) 1417 1154·6 (1096·0–1216·3) 192 156·4 (135·8–180·2) Diagnosis Affective 5041 (27·7%) 409 1338·9 (1215·3–1475·2) 327 1070·5 (960·5–1193·0) 51 167·0 (126·9–219·7) Non-affective 13 160 (72·3%) 1358 1473·1 (1396·8–1553·6) 1090 1182·4 (1114·3–1254·7) 141 153·0 (129·7–180·4) Ethnicity White British 9047 (49·7%) 1130 1989·8 (1877·1–2109·2) 913 1607·7 (1506·7–1715·4) 125 220·1 (184·7–262·3) Black African 2510 (13·8%) 106 645·9 (533·9–781·3) 82 499·7 (402·42–620·4) 18 109·7 (69·1–174·1) Black Caribbean 4840 (26·6%) 332 880·5 (790·7–980·5) 256 678·9 (600·7–767·4) 33 87·5 (62·2–123·1) South Asian 1256 (6·9%) 95 1178·2 (963·6–1440·6) 77 955·0 (763·8–1194·0) 11 136·4 (75·6–246·3) Irish 548 (3·0%) 104 2765·9 (2282·3–3352·1) 89 2367·0 (1923·0–2913·6) 5 133·0 (55·3–319·5) Sex Male 9610 (52·8%) 908 1340·3 (1255·9–1430·4) 707 1043·6 (969·4–1123·4) 132 194·8 (164·3–231·1) Female 8591 (47·2%) 859 1562·3 (1461·2–1670·3) 710 1291·3 (1199·7–1389·8) 60 109·1 (84·7–140·5) Area-level deprivation (IMD rank) Median (IQR) 7200 (4737–11 610)* ·· ·· ·· ·· ·· ·· IMD fifths First (least deprived) 3667 (20·2%) 377 1442·3 (1303·8–1595·5) 305 1116·8 (1043·0–1305·4) 33 126·2 (89·8–177·6) Second 3646 (20·0%) 342 1332·4 (1198·4–1481·4) 273 1063·6 (944·6–1197·5) 35 136·4 (97·9–189·9) Third 3629 (19·9%) 313 1250·5 (1119·3–1397·0) 247 986·8 (871·1–117·9) 37 147·8 (107·1–204·0) Fourth 3673 (20·2%) 371 1493·2 (1348·7–1653·2) 297 1195·4 (1066·9–1339·4) 39 156·9 (114·7–214·8) Fifth (most deprived) 3586 (19·7%) 364 1729·4 (1560·5–1916·5) 295 1401·5 (1250·4–1571·0) 48 220·0 (171·9–302·6) Marital status Married or cohabiting 2781 (15·2%) 267 1583·1 (1404·2–1784·9) 220 1304·4 (1143·0–1488·7) 22 130·4 (85·9–198·1) Single, divorced, widowed, or separated 15 420 (84·7%) 1500 1416·9 (1347·0–1490·4) 1197 1130·7 (1068·4–1196·6) 170 160·6 (138·2–186·6) Comorbid substance misuse diagnosis None 15 046 (82·7%) 1519 1505·8 (1431·9–1583·4) 1251 1240·1 (1173·3–1310·8) 128 126·9 (106·7–150·9) Present 3155 (17·3%) 248 1134·8 (1002·0–1285·2) 166 759·6 (652·4–884·4) 64 292·9 (229·2–374·2) IMD=Index of Multiple Deprivation.

4) 1197 1130·7 (1068·4–1196·6) 170 160·6 (138·2–186·6) Comorbid substance misuse diagnosis None 15 046 (82·7%) 1519 1505·8 (1431·9–1583·4) 1251 1240·1 (1173·3–1310·8) 128 126·9 (106·7–150·9) Present 3155 (17·3%) 248 1134·8 (1002·0–1285·2) 166 759·6 (652·4–884·4) 64 292·9 (229·2–374·2) IMD=Index of Multiple Deprivation. Data for 158 individuals who died from unknown causes not elsewhere classified (R00–99) have not been displayed. * Range for the IMD at the level of Lower Super Output Area for England is 1 (most deprived) to 32 844 (least deprived); therefore, the rank of 7200 is out of 32 844 small areas in England—ie, the study area is among the 22% (IQR 14–35) most deprived in the country. Table 2 SMRs for selected causes in people with severe mental illness

* Range for the IMD at the level of Lower Super Output Area for England is 1 (most deprived) to 32 844 (least deprived); therefore, the rank of 7200 is out of 32 844 small areas in England—ie, the study area is among the 22% (IQR 14–35) most deprived in the country. Table 2 SMRs for selected causes in people with severe mental illness Total sample White British Black Caribbean Black African South Asian Irish All-cause mortality Number of deaths 2390 1317 384 127 106 115 SMR (95% CI) 2·67 (2·56–2·78) 2·96 (2·80–3·12) 2·11 (1·90–2·33) 2·98 (2·48–3·54) 2·34 (1·91–2·83) 2·59 (2·14–3·11) p value <0·001 <0·001 <0·001 <0·001 <0·001 <0·001 Natural-cause mortality Circulatory system, including cardiovascular and cerebrovascular disease Number of deaths 647 331 121 37 28 33 SMR (95% CI) 2·65 (2·45–2·86) 2·66 (2·38–2·96) 2·46 (2·04–2·94) 3·85 (2·71–5·31) 2·33 (1·55–3·37) 2·58 (1·78–3·63) p value <0·001 <0·001 <0·001 <0·001 <0·001 0·021 Respiratory system, including COPD and pneumonia Number of deaths 370 237 29 11 23 24 SMR (95% CI) 3·38 (3·04–3·74) 4·19 (3·67–4·76) 1·32 (0·89–1·90) 2·87 (1·43–5·14) 4·38 (2·78–6·57) 4·06 (2·60–6·04) p value <0·001 <0·001 0·17 0·0043 <0·001 <0·001 Cancers Number of deaths 433 244 63 43 ≤10* 18 SMR (95% CI) 1·45 (1·32–1·60) 1·68 (1·47–1·90) 1·05 (0·80–1·34) 1·79 (1·19–2·59) 0·49 (0·21–0·96) 1·25 (0·74–1·97) p value <0·001 <0·001 0·76 0·0062 0·035 0·41 Unnatural-cause mortality Suicide, self-harm, and events of undetermined intent Number of deaths 138 71 21 14 ≤10* ≤10* SMR (95% CI) 7·65 (6·43–9·04) 10·41 (8·13–13·13) 4·87 (3·01–7·44) 6·71 (3·67–11·26) 7·41 (2·98–15·27) 6·67 (1·38–19·49) p value <0·001 <0·001 <0·001 <0·001 <0·001 0·024 Other external causes of mortality† Number of deaths 126 77 19 ≤10* ≤10* ≤10* SMR (95% CI) 4·01 (3·34–4·78) 5·79 (4·57–7·23) 2·68 (1·61–4·19) 1·42 (0·39–3·63) 1·58 (1·03–7·38) 2·78 (0·57–8·13) p value <0·001 <0·001 <0·001 0·62 0·047 0·19 SMRs standardised by age and sex to the population of England and Wales in 2011. Total number of observed deaths is greater than in regression models, as mortality in the sample was assessed irrespective of missing data for covariates and ethnicity. p values derived through Byaar approximation. SMR=standardised mortality ratio. COPD=chronic obstructive pulmonary disease.

England and Wales in 2011. Total number of observed deaths is greater than in regression models, as mortality in the sample was assessed irrespective of missing data for covariates and ethnicity. p values derived through Byaar approximation. SMR=standardised mortality ratio. COPD=chronic obstructive pulmonary disease. * Exact number not shown to protect patient confidentiality. † Included accidents, falls, and assaults. Table 3 Adjusted HRs for all-cause mortality, by ethnic group and time since diagnosis Total cases Number of deaths Time since diagnosis 0 to <3·21 years 3·21 to <6·71 years ≥6·71 years Crude Adjusted* Crude Adjusted* Crude Adjusted* HR (95% CI) p value HR (95% CI) p value HR (95% CI) p value HR (95% CI) p value HR (95% CI) p value HR (95% CI) p value White British 9047 1130 1 ·· 1 ·· 1 ·· 1 ·· 1 ·· 1 ·· Black Caribbean 4840 332 0·27 (0·21–0·36) <0·001 0·35 (0·26–0·46) <0·0001 0·46 (0·37–0·56) <0·001 0·58 (0·47–0·72) <0·0001 0·48 (0·40–0·58) <0·0001 0·67 (0·56–0·81) <0·0001 Black African 2510 106 0·36 (0·26–0·50) <0·001 0·65 (0·46–0·91) 0·013 0·37 (0·27–0·51) <0·0001 0·68 (0·49–0·94) 0·021 0·24 (0·16–0·35) <0·0001 0·43 (0·30–0·64) <0·0001 South Asian 1256 95 0·59 (0·41–0·86) 0·0060 0·69 (0·48–1·01) 0·055 0·58 (0·40–0·83) 0·003 0·69 (0·48–0·99) 0·042 0·59 (0·41–0·84) 0·0040 0·73 (0·51–1·04) 0·086 Irish 548 104 1·40 (0·97–2·00) 0·069 0·93 (0·65–1·34) 0·71 1·52 (1·09–2·12) 0·014 1·03 (0·74–1·44) 0·85 1·17 (0·82–1·66) 0·38 0·88 (0·62–1·25) 0·48 Models derived from Cox regression analyses; p values calculated with the Wald test. HR=hazard ratio.

0·69 (0·48–0·99) 0·042 0·59 (0·41–0·84) 0·0040 0·73 (0·51–1·04) 0·086 Irish 548 104 1·40 (0·97–2·00) 0·069 0·93 (0·65–1·34) 0·71 1·52 (1·09–2·12) 0·014 1·03 (0·74–1·44) 0·85 1·17 (0·82–1·66) 0·38 0·88 (0·62–1·25) 0·48 Models derived from Cox regression analyses; p values calculated with the Wald test. HR=hazard ratio. * Adjusted for age, sex, diagnosis, marital status, area-level deprivation, and comorbid alcohol or substance misuse. Table 4 Adjusted HRs for all-cause mortality, by covariates

0·69 (0·48–0·99) 0·042 0·59 (0·41–0·84) 0·0040 0·73 (0·51–1·04) 0·086 Irish 548 104 1·40 (0·97–2·00) 0·069 0·93 (0·65–1·34) 0·71 1·52 (1·09–2·12) 0·014 1·03 (0·74–1·44) 0·85 1·17 (0·82–1·66) 0·38 0·88 (0·62–1·25) 0·48 Models derived from Cox regression analyses; p values calculated with the Wald test. HR=hazard ratio. * Adjusted for age, sex, diagnosis, marital status, area-level deprivation, and comorbid alcohol or substance misuse. Table 4 Adjusted HRs for all-cause mortality, by covariates Total cases Number of deaths Crude estimates Adjusted estimates* HR (95% CI) p value (Wald test) p value (LRT) HR (95% CI) p value (Wald test) p value (LRT) Sex Male 9610 908 1 ·· 0·00020 1 ·· 0·0030 Female 8591 859 1·19 (1·09–1·31) <0·0001 ·· 0·86 (0·79–0·95) 0·0030 ·· Diagnosis Non-affective 13 160 1358 1 ·· 0·28 1 ·· 0·0033 Affective 5041 409 0·94 (0·84–1·05) 0·28 ·· 0·84 (0·75–0·95) 0·0040 ·· Marital status Married or cohabiting 2781 267 1 ·· 0·031 1 ·· 0·0087 Single, divorced, widowed, or separated 15 420 1500 0·86 (0·76–0·98) 0·028 ·· 1·19 (1·04–1·36) 0·010 ·· Comorbid alcohol or substance misuse None 15 046 1519 1 ·· <0·001 1 ·· 0·016 Present 3155 248 0·75 (0·65–0·85) <0·0001 ·· 1·19 (1·04–1·37) 0·015 ·· Area-level deprivation (IMD fifths) First (least deprived) 3667 377 1 ·· <0·0001 1 ·· 0·30 Second 3646 342 0·93 (0·80–1·07) 0·32 ·· 0·91 (0·78–1·05) 0·19 ·· Third 3629 313 0·87 (0·75–1·01) 0·075 ·· 0·86 (0·74–1·00) 0·054 ·· Fourth 3673 371 1·05 (0·91–1·22) 0·48 ·· 0·93 (0·81–1·08) 0·36 ·· Fifth (most deprived) 3586 364 1·26 (1·09–1·46) 0·0010 ·· 0·98 (0·85–1·14) 0·80 ·· Models derived from Cox regression analyses. HR=hazard ratio. LRT=likelihood ratio test. IMD=Index of Multiple Deprivation.

(0·75–1·01) 0·075 ·· 0·86 (0·74–1·00) 0·054 ·· Fourth 3673 371 1·05 (0·91–1·22) 0·48 ·· 0·93 (0·81–1·08) 0·36 ·· Fifth (most deprived) 3586 364 1·26 (1·09–1·46) 0·0010 ·· 0·98 (0·85–1·14) 0·80 ·· Models derived from Cox regression analyses. HR=hazard ratio. LRT=likelihood ratio test. IMD=Index of Multiple Deprivation. * Adjusted for age, ethnicity, and all other variables shown in the table. Table 5 Sub-HRs for natural-cause and unnatural-cause mortality

(0·75–1·01) 0·075 ·· 0·86 (0·74–1·00) 0·054 ·· Fourth 3673 371 1·05 (0·91–1·22) 0·48 ·· 0·93 (0·81–1·08) 0·36 ·· Fifth (most deprived) 3586 364 1·26 (1·09–1·46) 0·0010 ·· 0·98 (0·85–1·14) 0·80 ·· Models derived from Cox regression analyses. HR=hazard ratio. LRT=likelihood ratio test. IMD=Index of Multiple Deprivation. * Adjusted for age, ethnicity, and all other variables shown in the table. Table 5 Sub-HRs for natural-cause and unnatural-cause mortality Natural causes Unnatural causes Number of deaths Crude estimates Adjusted estimates* Number of deaths Crude estimates Adjusted estimates* Sub-HR (95% CI) p value (Wald test) p value (LRT) Sub-HR (95% CI) p value (Wald test) p value (LRT) Sub-HR (95% CI) p value (Wald test) p value (LRT) Sub-HR (95% CI) p value (Wald test) p value (LRT) Ethnicity White British 913 1 ·· <0·0001 1 ·· <0·0001 125 1 ·· 0·00010 1 ·· 0·00030 Black African 82 0·31 (0·25–0·39) <0·0001 ·· 0·49 (0·39–0·61) <0·0001 ·· 18 0·52 (0·32–0·85) 0·009 ·· 0·55 (0·33–0·92) 0·023 ·· Black Caribbean 256 0·42 (0·37–0·48) <0·0001 ·· 0·52 (0·45–0·59) <0·0001 ·· 33 0·43 (0·30–0·64) <0·0001 ·· 0·42 (0·28–0·63) <0·0001 ·· South Asian 77 0·59 (0·47–0·74) <0·0001 ·· 0·66 (0·53–0·83) <0·0001 ·· 11 0·64 (0·34–1·18) 0·15 ·· 0·68 (0·36–1·28) 0·23 ·· Irish 89 1·47 (1·18–1·83) 0·0010 ·· 1·08 (0·87–1·34) 0·48 ·· 5 0·60 (0·25–1·48) 0·27 ·· 0·59 (0·24–1·46) 0·26 ·· Sex Male 707 1 ·· <0·0001 1 ·· 0·13 132 1 ·· 0·00010 1 ·· 0·0020 Female 710 1·25 (1·12–1·38) <0·0001 ·· 0·92 (0·83–1·02) 0·13 ·· 60 0·54 (0·40–0·73) <0·0001 ·· 0·61 (0·44–0·83) 0·0020 ·· Diagnosis Non-affective 327 1 ·· 0·11 1 ·· 0·00040 51 1 ·· 0·75 1 ·· 0·98 Affective 1090 0·90 (0·80–1·02) 0·11 ·· 0·80 (0·70–0·90) <0·0001 ·· 141 1·05 (0·77–1·45) 0·75 ·· 1·00 (0·72–1·39) 0·98 ·· Marital status Married or cohabiting 220 1 ·· 0·050 1 ·· 0·036 22 1 ·· 0·26 1 ·· 0·64 Single, divorced, widowed, or separated 1197 0·87 (0·75–1·00) 0·050 ·· 1·17 (1·01–1·36) 0·036 ·· 170 1·29 (0·83–2·01) 0·26 ·· 1·12 (0·71–1·76) 0·64 ·· Comorbid alcohol or substance misuse None 1251 1 ·· <0·0001 1 ·· 0·12 128 1 ·· <0·0001 1 ·· <0·0001 Present 166 0·61 (0·51–0·71) <0·0001 ·· 0·88 (0·74–1·04) 0·12 ·· 64 2·36 (1·75–3·19) <0·0001 ·· 2·08 (1·53–2·83) <0·0001 ·· Area-level deprivation (IMD fifths) First (least deprived) 305 1 ·· 0·0010 1 ·· 0·16 33 1 ·· 0·13 1 ·· 0·24 Second 273 0·91 (0·77–1·07) 0·27 ·· 0·87 (0·75–1·02) 0·086 ·· 35 1·08 (0·67–1·74) 0·75 ·· 1·03 (0·64–1·66) 0·89 ·· Third 247 0·85 (0·72–1·00) 0·054 ·· 0·91 (0·78–1·07) 0·014 ·· 37 1·18 (0·74–1·88) 0·50 ·· 1·17 (0·73–1·87) 0·52 ·· Fourth 297 1·03 (0·88–1·21) 0·73 ·· 0·91 (0·78–1·07) 0·25 ·· 39 1·23 (0·77–1·96) 0·38 ·· 1·20 (0

273 0·91 (0·77–1·07) 0·27 ·· 0·87 (0·75–1·02) 0·086 ·· 35 1·08 (0·67–1·74) 0·75 ·· 1·03 (0·64–1·66) 0·89 ·· Third 247 0·85 (0·72–1·00) 0·054 ·· 0·91 (0·78–1·07) 0·014 ·· 37 1·18 (0·74–1·88) 0·50 ·· 1·17 (0·73–1·87) 0·52 ·· Fourth 297 1·03 (0·88–1·21) 0·73 ·· 0·91 (0·78–1·07) 0·25 ·· 39 1·23 (0·77–1·96) 0·38 ·· 1·20 (0 ·75–1·91) 0·44 ·· Fifth (most deprived) 295 1·19 (1·02–1·40) 0·029 ·· 0·91 (0·77–1·06) 0·23 ·· 48 1·70 (1·09–2·64) 0·019 ·· 1·60 (1·02–2·52) 0·04 ·· Models took into account competing risks. Data for 158 individuals who died from unknown causes not elsewhere classified (R00–99) have not been displayed. HR=hazard ratio. LRT=likelihood ratio test. IMD=Index of Multiple Deprivation. * Adjusted for age and all other variables shown in the table.

Introduction Although absolute risks of people with schizophrenia spectrum and bipolar disorder committing violent crime are typically around 5–10% within 5 years of diagnosis and most patients are not violent in their lifetimes,1, 2 violence perpetrated by individuals with these disorders is an important preventable cause of morbidity. Furthermore, it contributes to stigma and the large numbers of people with mental illness in prisons. One of the main approaches to reduce violence risk has been to use structured risk assessment tools, which range from checklists to complex decision trees, and to stratify individuals into high-risk and low-risk groups. These tools are used in mental health services, especially in forensic psychiatry, and are recommended in clinical guidelines.3, 4, 5 Such stratification can help target resources, tailor treatment and risk management, and inform decisions about assertive community treatment, hospital treatment, and other services.5

s. These tools are used in mental health services, especially in forensic psychiatry, and are recommended in clinical guidelines.3, 4, 5 Such stratification can help target resources, tailor treatment and risk management, and inform decisions about assertive community treatment, hospital treatment, and other services.5 Current risk assessment instruments, however, have been limited by low-to-moderate accuracy,6 poor reporting standards,7 and inconsistent definitions of what constitutes high risk.8 The tools have rarely been developed in individuals with psychosis.9 Additionally, many have considerable resource implications, with current approaches taking around 16 person-hours in one forensic psychiatric setting,10 and most instruments requiring several hours. By contrast, some areas of medicine, in particular cardiovascular medicine, have developed scalable risk prediction scores, such as the Framingham Risk Score and QRISK prediction algorithm, which can be used in primary and secondary care to inform discussions between clinicians and individuals about risk.11 A key factor in their widespread use is their ease and simplicity. The need for shorter violence risk assessments than at present, validated in appropriate patient groups, has been highlighted by an American Psychiatric Association taskforce.5 To address the need for a scalable and valid tool to assess violence risk in patients with schizophrenia spectrum and bipolar disorder, we describe the derivation of a score based on routinely collected factors and present findings from external validation.

emoved it from the dataset used to fit the model (the derivation sample). We chose the number of regions for the validation sample to be large enough for a useful assessment of external validity to be made.18 The geographical regions and method for selection of the validation sample are described in the appendix (p 6). In external validation, we summarised predictive accuracy using: first, the concordance index19 to assess discrimination (ability of the model to distinguish between those who do and do not commit a violent crime, with a value of 1 meaning perfect discrimination); second, the Brier score20 for calibration (model goodness of fit—whether or not the predicted risk is systematically off target, with 0 meaning perfect calibration; the Brier score measures the mean squared difference between the predicted probability and the actual outcome [violent crime or no violent crime]); third, the net reclassification index21 (how well a model rightly or wrongly reclassifies patients compared with alternative models); and fourth, sensitivity and specificity based on a 5% threshold of predicted probability. We compared the proportions of predicted and observed events at different levels of predicted probability using a calibration plot. On the basis of research that has found an incidence of violent crime in schizophrenia spectrum disorders at 1 year of 4%,1 we prespecified a 5% cutoff for low-to-high risk of violent offending. We used a higher cutoff than this 4% incidence as the previous data were based on a younger age cohort and only on patients with schizophrenia spectrum disorder. We used Stata version 12 and R version 3.2.1 for all analyses. We followed the TRIPOD statement (appendix p 7).

off for low-to-high risk of violent offending. We used a higher cutoff than this 4% incidence as the previous data were based on a younger age cohort and only on patients with schizophrenia spectrum disorder. We used Stata version 12 and R version 3.2.1 for all analyses. We followed the TRIPOD statement (appendix p 7). Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of the cohort of 75 158 patients with schizophrenia spectrum or bipolar disorder, we assigned 58 771 (78%) to the derivation sample and 16 387 (22%) to the validation sample (table 1). Overall, 869 (2·3%) of 37 221 men and 181 (0·5%) of 37 937 women committed violent crime over a 12 month period (figure 1). In the derivation sample, 830 (1%) individuals committed a violent offence within 12 months of their patient episode, 1702 (3%) died within 12 months, and 40 611 (69%) were outpatients at the time of the episode.Figure 1 Violent crime over a 12 month period in different populations by sex *Data taken from the general population sample in Fazel and colleagues.1 Table 1 Baseline characteristics of the derivation sample diagnosed with schizophrenia spectrum or bipolar disorder, with grouping of violence risk factors

Results Of the cohort of 75 158 patients with schizophrenia spectrum or bipolar disorder, we assigned 58 771 (78%) to the derivation sample and 16 387 (22%) to the validation sample (table 1). Overall, 869 (2·3%) of 37 221 men and 181 (0·5%) of 37 937 women committed violent crime over a 12 month period (figure 1). In the derivation sample, 830 (1%) individuals committed a violent offence within 12 months of their patient episode, 1702 (3%) died within 12 months, and 40 611 (69%) were outpatients at the time of the episode.Figure 1 Violent crime over a 12 month period in different populations by sex *Data taken from the general population sample in Fazel and colleagues.1 Table 1 Baseline characteristics of the derivation sample diagnosed with schizophrenia spectrum or bipolar disorder, with grouping of violence risk factors Patients (n=58 771) Group 1* Male sex 29 077 (49%) Age (years) 44 (13) Previous violent crime 9212 (16%) Previous drug use 7123 (12%) Previous alcohol use 8897 (15%) Previous self-harm 11 510 (20%) Educational level Lower secondary (<16 years of age) 17 814/50 752 (35%) Upper secondary (16–18 years of age) 26 449/50 752 (52%) Postsecondary (>18 years of age) 6489/50 752 (13%) Parental drug or alcohol use 5214/47 957 (11%) Parental violent crime 3203/47 957 (7%) Sibling violent crime 4028 (7%) Group 2† Diagnosis Schizophrenia spectrum disorder 36 755 (63%) Bipolar disorder 22 016 (37%) Recent treatment (within preceding 6 months) Mood stabiliser 10 390/34 039 (31%) Antipsychotic 18 401/34 039 (54%) Antidepressant 13 255/34 039 (39%) Dependence 1030/34 039 (3%) Inpatient at time of episode 18 160 (31%) Length of first inpatient stay >7 days 24 532 (42%) More than seven patient episodes 16 686 (28%) Group 3‡ Received benefits§ 37 210/57 876 (64%) Deprivation First decile (lowest) 2793/56 617 (5%) Fifth decile 4862/56 617 (9%) Tenth decile (highest) 10 769/56 617 (19%) Never married 34 506/57 459 (60%) Personal income First decile (lowest) 5444/57 876 (9%) Fifth decile 9169/57 876 (16%) Tenth decile (highest) 2009/57 876 (3%) Children in household 11 079 (19%) Parental psychiatric admission to hospital 13 225/47 957 (28%) Parental suicide 1417/47 957 (3%) Comorbid depression 11 934/36 755 (32%)¶ Recent death of family member (within preceding 6 months) 953/47 957 (2%) Data are n (%), mean (SD), or n/N (%).

decile (highest) 2009/57 876 (3%) Children in household 11 079 (19%) Parental psychiatric admission to hospital 13 225/47 957 (28%) Parental suicide 1417/47 957 (3%) Comorbid depression 11 934/36 755 (32%)¶ Recent death of family member (within preceding 6 months) 953/47 957 (2%) Data are n (%), mean (SD), or n/N (%). * Variables included in the model on the basis of previous evidence. † Variables considered for the model with strong evidence, but retained in the model if significant. ‡ Variables considered for the model with weaker evidence, but retained in the model if significant. § Welfare or disability benefits. ¶ Denominator is the number of patients with schizophrenia spectrum disorder.

* Variables included in the model on the basis of previous evidence. † Variables considered for the model with strong evidence, but retained in the model if significant. ‡ Variables considered for the model with weaker evidence, but retained in the model if significant. § Welfare or disability benefits. ¶ Denominator is the number of patients with schizophrenia spectrum disorder. We included 16 items in the final model (table 2; appendix p 8). The strongest predictors of violent offending within 12 months were previous violent crime conviction, male sex, and age. The decline in probability of violent offending was approximately linearly related to increasing age (appendix p 9). Personal income and benefit receipt were among the weaker predictors of violent offending. Shrinkage effects (a measure of the adjustment required for a model fitted to sample data so that it does not overestimate predictive performance) were negligible: the average estimate of the shrinkage heuristic from the bootstrapped samples was 99%. We arrived at the final model by including all group 1 variables and the group 2 and 3 variables that retained significance with multivariable analyses.Table 2 Associations between risk factors and violent crime in the derivation sample from the multiple regression model (after multiple imputation)

ped samples was 99%. We arrived at the final model by including all group 1 variables and the group 2 and 3 variables that retained significance with multivariable analyses.Table 2 Associations between risk factors and violent crime in the derivation sample from the multiple regression model (after multiple imputation) Adjusted odds ratio (95% CI) p value Male sex 2·32 (1·91–2·81) <0·0001 Age at hospital discharge (per 10 years) 0·63 (0·58–0·67) <0·0001 Previous violent crime* 5·03 (4·23–5·98) <0·0001 Previous drug use† 1·45 (1·23–1·72) <0·0001 Previous alcohol use† 1·75 (1·47–2·09) <0·0001 Previous self-harm† 1·23 (1·04–1·45) 0·02 Educational level .. 0·31 Lower secondary (<16 years of age) 1 (reference) .. Upper secondary (16–18 years of age) 0·88 (0·75–1·04) .. Postsecondary (>18 years of age) 0·93 (0·69–1·26) .. Parental drug or alcohol use† 1·11 (0·91–1·35) 0·30 Parental violent crime* 1·16 (0·92–1·46) 0·21 Sibling violent crime* 0·90 (0·71–1·13) 0·35 Recent treatment‡—antipsychotic 0·62 (0·51–0·77) <0·0001 Recent treatment‡—antidepressant 0·80 (0·65–0·99) 0·04 Recent treatment‡—dependence§ 1·78 (1·22–2·60) 0·003 Inpatient at time of episode 1·37 (1·18–1·59) <0·0001 Received benefits¶ 1·42 (1·17–1·72) 0·0003 Personal income .. 0·046 Fifth decile 0·84 (0·65–1·10) .. Tenth decile (highest) 0·88 (0·50–1·57) .. * Conviction for homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), illegal threats, or harassment.

nefits¶ 1·42 (1·17–1·72) 0·0003 Personal income .. 0·046 Fifth decile 0·84 (0·65–1·10) .. Tenth decile (highest) 0·88 (0·50–1·57) .. * Conviction for homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), illegal threats, or harassment. † Inpatient or outpatient International Classification of Diseases diagnosis in patient register. ‡ Dispensed within the last 6 months. § Drugs used in addictive disorders. ¶ Welfare or disability benefits.

nefits¶ 1·42 (1·17–1·72) 0·0003 Personal income .. 0·046 Fifth decile 0·84 (0·65–1·10) .. Tenth decile (highest) 0·88 (0·50–1·57) .. * Conviction for homicide, assault, robbery, arson, any sexual offence (rape, sexual coercion, child molestation, indecent exposure, or sexual harassment), illegal threats, or harassment. † Inpatient or outpatient International Classification of Diseases diagnosis in patient register. ‡ Dispensed within the last 6 months. § Drugs used in addictive disorders. ¶ Welfare or disability benefits. The model showed good overall discrimination on the basis of the results from both internal validation with use of bootstrapping (c-index 0·86 [95% CI 0·84–0·89]; Brier score 0·0132; net reclassification index 1·14) and external validation (c-index 0·89 [0·85–0·93]; Brier score 0·0120; net reclassification index 1·28). When we used the prespecified 5% risk cutoff for violent crime in 1 year, the sensitivity was 49% (95% CI 45–52) and the specificity was 94% (94–94) in internal validation. The sensitivity in external validation was slightly higher (62% [55–68]) than in internal validation, with the same specificity (94% [93–94]). The positive predictive value was 11% and the negative predictive value was more than 99%. The 2 × 2 tables used to derive sensitivity are shown in the appendix (p 10). Receiver operating characteristic curves are shown in figure 2. Calibration plots indicate adequate calibration of the predicted probabilities against observed proportions of violent offending (figure 3).Figure 2 Model discrimination in the (A) derivation and (B) external validation samples

hown in the appendix (p 10). Receiver operating characteristic curves are shown in figure 2. Calibration plots indicate adequate calibration of the predicted probabilities against observed proportions of violent offending (figure 3).Figure 2 Model discrimination in the (A) derivation and (B) external validation samples Figure 3 Predicted and observed risks of violent crime in the (A) derivation and (B) external validation samples Individuals are grouped by predicted probability, numbers are the number of individuals in each grouping, and error bars are 95% CIs for the proportion of events in each group. We applied the coefficients to develop a web calculator called the Oxford Mental Illness and Violence tool [OxMIV]), which is free to use. This tool provides both a risk classification (low or high) and a probability of violent offending within the next 12 months. A beta version of the online risk calculator for violent offending (based on the coefficients in the appendix [p 8]) can be found online. If missing values are present, this calculator reports the upper and lower range of estimates of risk.

(low or high) and a probability of violent offending within the next 12 months. A beta version of the online risk calculator for violent offending (based on the coefficients in the appendix [p 8]) can be found online. If missing values are present, this calculator reports the upper and lower range of estimates of risk. Discussion We have described the development of a clinical prediction score and web calculator (OxMIV) for risk of committing violent crime in individuals with diagnoses of schizophrenia spectrum or bipolar disorder with good measures of discrimination and calibration. We developed a 16-item model from prespecified criminal history, sociodemographic, and clinical risk factors. The prediction score is brief and simple to use, relies on information that can be routinely collected, and is able to stratify individuals into high-risk and low-risk groups. It is primarily intended as an adjunct to clinical decision making and to anchor decisions on violence risk using an evidence-based approach, which will need to be complemented by individualised and contextual factors.22

can be routinely collected, and is able to stratify individuals into high-risk and low-risk groups. It is primarily intended as an adjunct to clinical decision making and to anchor decisions on violence risk using an evidence-based approach, which will need to be complemented by individualised and contextual factors.22 Current risk assessment tools predicting violence risk in psychiatric populations are limited by being time-consuming, requiring training, having direct costs in most cases, and randomised controlled trial evidence23 that instruments based on structured clinical judgment do not improve patient outcomes. Despite these limitations, they are widely used in general and forensic psychiatry.24 To our knowledge, none of the current approaches use web-based calculators, are free to use, and incorporate treatment information. By contrast, OxMIV addresses these limitations, and as the derivation sample was based on 58 771 patients, it provides substantially improved precision for identified risk factors. The performance statistics of OxMIV are better than are those reported for other instruments currently used in mental health. For violent crime over 1 year, the model had a sensitivity of 62% and a specificity of 94% in the external validation sample with use of prespecified risk thresholds. The positive predictive value was 11% and the negative predictive value was more than 99%.

eported for other instruments currently used in mental health. For violent crime over 1 year, the model had a sensitivity of 62% and a specificity of 94% in the external validation sample with use of prespecified risk thresholds. The positive predictive value was 11% and the negative predictive value was more than 99%. Authors of a 2011 systematic review9 found that only two instruments had been validated in 861 patients with psychosis, with areas under the curves (AUCs) that ranged from 0·60 to 0·77, with little other performance information. By contrast, the overall c-index (equivalent to an AUC) of OxMIV was 0·89 in the external validation. A broad review6 of all violence instruments used in criminal justice reported a median AUC of 0·72 (IQR 0·68–0·78), with a sensitivity higher than that of OxMIV, at 92%, but a specificity lower, at 36%, with these differences in sensitivity and specificity being explained by different thresholds used in criminal justice populations and in the population of this study. Authors of a review25 of commonly used cardiovascular risk scores reported AUCs in the range of 0·70 to 0·75. Another limitation of current approaches is that they might increase stigma, particularly by overestimating risks. More precision than at present about future probabilities of violent crime could partly address this limitation, as well as more awareness of the epidemiological evidence for actual violence over a 12 month period (figure 1).

f current approaches is that they might increase stigma, particularly by overestimating risks. More precision than at present about future probabilities of violent crime could partly address this limitation, as well as more awareness of the epidemiological evidence for actual violence over a 12 month period (figure 1). One clinical implication of OxMIV is that it could be used to screen for low violence risk in general adult psychiatric services. This use is facilitated by the high negative predictive value (which was was in fact 99·5%)—in other words, of those identified as low risk, 199 of 200 did not in fact offend violently within 1 year. But the low positive predictive value of the tool means that it should not be used to predict violent crime as only around one in 10 of those identified as high risk will actually violently offend. Rather, it can identify patients at higher than average risk (figure 4). However, as OxMIV's sensitivity was 62% using the 5% threshold, it can be used to stratify patients into low-risk and high-risk groups because nearly two-thirds of all those who do violently offend will be picked up by this tool. If the consequences of identification of someone at high risk are not harmful, then this identification has the potential to reduce violence risk. At the same time, this identification cannot be used to detain individuals or extend their detention in the absence of other clinical factors and detailed assessment.Figure 4 Observed and predicted risk of violent crime in severe mental illness NPV=negative predictive value. PPV=positive predictive value.

One clinical implication of OxMIV is that it could be used to screen for low violence risk in general adult psychiatric services. This use is facilitated by the high negative predictive value (which was was in fact 99·5%)—in other words, of those identified as low risk, 199 of 200 did not in fact offend violently within 1 year. But the low positive predictive value of the tool means that it should not be used to predict violent crime as only around one in 10 of those identified as high risk will actually violently offend. Rather, it can identify patients at higher than average risk (figure 4). However, as OxMIV's sensitivity was 62% using the 5% threshold, it can be used to stratify patients into low-risk and high-risk groups because nearly two-thirds of all those who do violently offend will be picked up by this tool. If the consequences of identification of someone at high risk are not harmful, then this identification has the potential to reduce violence risk. At the same time, this identification cannot be used to detain individuals or extend their detention in the absence of other clinical factors and detailed assessment.Figure 4 Observed and predicted risk of violent crime in severe mental illness NPV=negative predictive value. PPV=positive predictive value. Use of a 5% threshold was based on 1 year postdiagnosis incidence data,1 although violence incidence in this sample was lower than that in the previous data as it included individuals at random points after diagnosis, meaning that individuals were older than in previous studies. Nevertheless, the 5% threshold might be considered low by some clinicians, but thresholds higher than 5% will be subject to more false negatives than those of 5% or lower. Broadly, use of this score provides a framework to make decisions between patients, carers, and health-care staff, and allow for anchoring of clinical risk in evidence, particularly as clinicians might overestimate risk.5 Furthermore, it could replace tools that are developed by local or regional general adult psychiatric services that have no supporting evidence and are often mandatory for clinicians to complete.

ealth-care staff, and allow for anchoring of clinical risk in evidence, particularly as clinicians might overestimate risk.5 Furthermore, it could replace tools that are developed by local or regional general adult psychiatric services that have no supporting evidence and are often mandatory for clinicians to complete. In terms of administration of this prediction score, including its communication to individuals and linkage to treatment, it is important to emphasise that it is only an adjunct to clinical decision making and that clinical judgment should supplement it with relevant individual factors. One strength is that any health-care professional can use OxMIV, including nursing, psychology, and medical staff, in primary and secondary care, but it should not be administered by non-clinical staff as it relies on diagnostic and treatment information and should only be part of a wider clinical assessment that considers other individual factors. The tool can be used at any point in a patient's pathway, apart from in forensic psychiatric patients and released prisoners,26 as baseline risks and the effects of risk factors will be different. Some of the items might not be routinely available, and we have provided for the possibility of scoring unknown for the socioeconomic and parental items. However, the risk factors included underscore the importance of taking a full psychiatric history to assist in prognosis.

e effects of risk factors will be different. Some of the items might not be routinely available, and we have provided for the possibility of scoring unknown for the socioeconomic and parental items. However, the risk factors included underscore the importance of taking a full psychiatric history to assist in prognosis. Strengths of our study include that it was based on a total cohort of all patients with diagnosed schizophrenia spectrum or bipolar disorder, with high-quality registers being linked to provide information about covariates and outcomes. Unlike previous risk assessment instruments,27 we have reported measures of discrimination and calibration in both derivation and external validation populations. Predictive accuracy was similar between the derivation and external validation samples, with the external validation sample being geographically separated from the derivation one. This finding would suggest that the model has the potential to be applied to different populations. Additionally, our shrinkage estimate was 99%, which is higher than the 85% that is recommended.16 Another methodological strength was use of imputation to replace missing data, which is new in the field of violence risk assessment. Finally, we have provided a free web calculator version for clinical staff.

populations. Additionally, our shrinkage estimate was 99%, which is higher than the 85% that is recommended.16 Another methodological strength was use of imputation to replace missing data, which is new in the field of violence risk assessment. Finally, we have provided a free web calculator version for clinical staff. A limitation of the study is validation in one country. Prevalence estimates for diagnoses of psychotic disorders vary minimally across European countries and the USA.28, 29, 30 In relation to the outcome of violent offending, police-recorded rates of serious violent offences, such as assault, robbery, and rape, have been shown to be fairly similar across high-income countries.31 Some miscalibration occurred for individuals at the highest risk at 1 year, but only a small number of individuals had a very high predicted risk, and predicted scores were not consistently higher or lower than observed ones. Nevertheless, we dealt with this miscalibration in the calculator by including a maximum risk of 20% so that any individuals with a predicted risk of 20% or higher all receive the same risk estimate (≥20%). We did not include migration data, which could lead to reduced time at risk for some individuals. However, the effects are likely to be minimal—in a related cohort of patients with schizophrenia spectrum disorders, 0·8% emigrated within 12 months of diagnosis.1 Being on medication for alcohol and drug misuse was associated with increasing risk as it was probably confounded by indication and acted as a proxy for severe comorbidity. Furthermore, our model does not include information about risk factors that could be collected in an interview, including specific symptoms,32 premorbid conduct problems, anger, victimisation, and comorbid personality disorder, which might further enhance the performance of the tool, but at the cost of making it more complex and time-consuming than without inclusion of this information. We did not consider comorbid personality disorders as their validity in Swedish registers is not known, but is likely to be low. As the tool mostly contains static factors, it should not be used to monitor within-individual changes in risk, but should be used as a cross-sectional score at a particular point in time. Some specific items, such as personal income and benefit receipt, might not be easily generalisable, but we have allowed for them to be scored as unknown and in the fitted model they were fairly weak predictors of violent offending.

s in risk, but should be used as a cross-sectional score at a particular point in time. Some specific items, such as personal income and benefit receipt, might not be easily generalisable, but we have allowed for them to be scored as unknown and in the fitted model they were fairly weak predictors of violent offending. The possibility to include missing data addresses one previously highlighted concern with existing structured instruments.5

s in risk, but should be used as a cross-sectional score at a particular point in time. Some specific items, such as personal income and benefit receipt, might not be easily generalisable, but we have allowed for them to be scored as unknown and in the fitted model they were fairly weak predictors of violent offending. The possibility to include missing data addresses one previously highlighted concern with existing structured instruments.5 An important issue is the implication of being labelled high risk and potential misuses, which could include restrictions on freedom (such as detention in hospital) and further stigma. The tool always needs to be used in conjunction with clinical decision making, and the ethics of deprivation of liberty versus risk to others should be carefully considered.33 Although the tool is freely available online, which allows for its widespread use in clinical services, this easy accessibility risks that it could be used for the wrong purposes and in the wrong contexts. Balancing of these issues will remain a challenge, and clear guidelines on the tool's intended population and how it should be used need to be established and regularly updated. Another limitation is use of violent crime as the primary outcome, which, although generalisable (as definitions are common across countries), with clear effects on public health, the absolute rates reported for violent crime are lower than for any violence, and the risk calculator provides a conservative estimate of violence risk. Additionally, the tool should not be used to assess risk of violence before hospital discharge in inpatients with psychiatric disorders, and separate models for subgroups of violent crime were not feasible. Future research should assess whether or not use of this prediction model improves outcomes for individuals with severe mental illness by reducing their risk of violent offending.

ce before hospital discharge in inpatients with psychiatric disorders, and separate models for subgroups of violent crime were not feasible. Future research should assess whether or not use of this prediction model improves outcomes for individuals with severe mental illness by reducing their risk of violent offending. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Wellcome Trust (095806) and Swedish Research Council. Contributors SF conceived the study and drafted the manuscript. SF, AW, TRF, and SM designed the methods. PL and HL obtained the data and designed the study. TRF analysed data. SF, AW, and TRF critically revised the manuscript with input from the other authors. Declaration of interests HL reports grants from Shire and has served as a speaker for Eli Lilly and Shire outside of the submitted work. PL has served as a speaker for Medice outside of the submitted work. SF has received a speaker's fee from Janssen outside of the submitted work, which was donated to charity. All other authors declare no competing interests.

Introduction Risk of relapse after the first episode of psychosis is high,1 constituting a substantial burden for health-care systems around the world,2, 3 and this relapse affects both individuals and society at large.4 In particular, relapse during the first few years after onset of the psychotic episode is an important determinant for long-term clinical and functional outcome.5 Hence, prevention of relapse is a crucial treatment target, which in turn underscores the importance of identification of modifiable risk factors that could influence relapse. Although the multifactorial nature of relapse is well known,6 two consistently identified modifiable risk factors influencing relapse are continued cannabis use following onset of psychosis7, 8, 9 and medication non-adherence,10, 11 both of which are unlikely to be the result of confounding or reverse causation.12 Despite the fact that the prevalence of post-onset cannabis use13 and medication non-adherence12 in patients with psychosis is high, understanding of the effects of this remains poor. There is poor understanding about how risk factors such as cannabis use might affect outcome in psychosis. Previous studies9, 14 have shown that the effect of cannabis use on risk of relapse was reduced when medication adherence was controlled for, suggesting that cannabis use could adversely affect psychosis outcome partly by influencing adherence to antipsychotic medication. This is consistent with independent evidence from a meta-analysis15 suggesting a significant effect of continued cannabis use on adherence to antipsychotic medication in patients with psychosis (p<0·0001), which was also confirmed by the five studies9, 12, 16, 17, 18 that investigated this issue subsequently. However, no study to date has systematically investigated to what extent the association between cannabis use and relapse of psychosis is mediated by non-adherence with prescribed psychotropic medication.

p<0·0001), which was also confirmed by the five studies9, 12, 16, 17, 18 that investigated this issue subsequently. However, no study to date has systematically investigated to what extent the association between cannabis use and relapse of psychosis is mediated by non-adherence with prescribed psychotropic medication. By elucidating the mechanistic pathway from cannabis use to psychosis relapse in first episode of psychosis in terms of potential mediational processes, we might be able to help identify alternative targets for intervention that could help mitigate the harm from cannabis use. Hence, in the present study, we aimed to explore whether some of the adverse effects of continued cannabis use on risk of relapse can be explained by its association with medication adherence; whether the association between continued cannabis use and risk of relapse is only partly, but not fully, mediated by medication adherence; and whether mediation effects are also present for other relapse-related outcomes, including number of relapses, length of relapse, time until relapse occurs, and intensity of care (eg, low-intensity outpatient care or high-intensity involuntary admission under section). Research in context Evidence before this study

By elucidating the mechanistic pathway from cannabis use to psychosis relapse in first episode of psychosis in terms of potential mediational processes, we might be able to help identify alternative targets for intervention that could help mitigate the harm from cannabis use. Hence, in the present study, we aimed to explore whether some of the adverse effects of continued cannabis use on risk of relapse can be explained by its association with medication adherence; whether the association between continued cannabis use and risk of relapse is only partly, but not fully, mediated by medication adherence; and whether mediation effects are also present for other relapse-related outcomes, including number of relapses, length of relapse, time until relapse occurs, and intensity of care (eg, low-intensity outpatient care or high-intensity involuntary admission under section). Research in context Evidence before this study We searched MEDLINE databases from inception to April 12, 2017, using a combination of search terms for describing diagnosis (psychosis: “psychosis”, “psychot*”, “schizophren*”, “schizoaff*”), exposure (cannabis use: “cannabi*”), and outcome of interest (medication adherence: “adheren*”, “complian*”), which retrieved 2092 articles, of which 20 were selected according to the following three criteria: (1) investigated the relationship between cannabis use and medication adherence; (2) the majority of the sample were taking antipsychotic medication; (3) participants were diagnosed with schizophrenia or any psychotic disorder using standardised criteria. We have previously summarised 15 of these studies (those published before April 27, 2015) as part of a meta-analysis, and these results showed that continued cannabis use increased the risk for non-adherence to antipsychotic medications. Results of the five additional studies that had been published since the original literature search were consistent with the results of our previous meta-analysis and confirmed that cannabis users were less likely to adhere to their prescribed medication than people who did not use cannabis. Of all relevant studies, only one investigated whether the effect of cannabis use on medication adherence mediated its effects on outcome in psychosis. They used data obtained from clinical records to report that poor medication adherence mediated the adverse effect of cannabis use on non-remission at 1 year in patients with psychosis. However, not all patients continue using cannabis following the onset of psychosis and a substantial proportion stop using the drug, a factor that was not accounted for by Colizzi and colleagues. Hence, whether non-adherence to antipsychotics truly mediates the effect of continued cannabis use following the onset of psychosis and the extent of this effect is unclear.

following the onset of psychosis and a substantial proportion stop using the drug, a factor that was not accounted for by Colizzi and colleagues. Hence, whether non-adherence to antipsychotics truly mediates the effect of continued cannabis use following the onset of psychosis and the extent of this effect is unclear. Added value of this study The present study extends current evidence on cannabis use being associated with increased risk of relapse in psychosis by investigating how it might be exerting this effect, focusing particularly on adherence to antipsychotic medication. The study benefits from data obtained in follow-up assessments of a large sample of patients with first-episode psychosis, which allowed a more detailed assessment of cannabis use profiles and pattern of medication adherence and the consideration of potential confounders. We explored consistency of effects by using different outcome measures of relapse, including risk or number of relapses, length of relapse, time until relapse, and severity of relapse. Implications of all the available evidence Collectively, the results of the present study and previous evidence indicate that relapse of psychosis associated with continued cannabis use might be partly mediated through non-adherence with prescribed medication. Hence, future investigations should test whether interventions aimed at improving medication adherence could partly help mitigate the adverse effects of cannabis use on outcome in psychosis.

se of psychosis associated with continued cannabis use might be partly mediated through non-adherence with prescribed medication. Hence, future investigations should test whether interventions aimed at improving medication adherence could partly help mitigate the adverse effects of cannabis use on outcome in psychosis. Methods Study design and participants All patients included in this prospective analysis were recruited from four different adult inpatient and outpatient units of the South London and Maudsley (SLAM) Mental Health National Health Service (NHS) Foundation Trust in Lambeth, Southwark, Lewisham, and Croydon as part of a follow-up study aiming to investigate the role of cannabis use within the first 2 years after onset of psychosis. Patients had a clinical diagnosis of first-episode non-organic (non-affective [ICD-10 codes F20–F29] or affective [F30–F33]) psychosis19 and were aged between 18 and 65 years when referred to local psychiatric services in south London, UK. We have previously reported on methods for assessment of patients and data acquisition.9, 12 This study was granted ethical approval by South London & Maudsley NHS Foundation Trust and the Institute of Psychiatry Local Research Ethics Committee. All patients included in the study gave written informed consent.

don, UK. We have previously reported on methods for assessment of patients and data acquisition.9, 12 This study was granted ethical approval by South London & Maudsley NHS Foundation Trust and the Institute of Psychiatry Local Research Ethics Committee. All patients included in the study gave written informed consent. Outcomes We obtained information regarding use of services, including number, duration, and legal status of inpatient admissions and referral to crisis intervention team or standard treatment by a community mental health team from electronic patient records, using the WHO Life Chart Schedule.20 Age of onset of psychosis was defined as the age on the date of referral to local psychiatric services for a first episode of psychosis. Our main outcome variable of interest was risk of relapse, which we defined as admission to a psychiatric inpatient unit owing to exacerbation of psychotic symptoms within 2 years following first presentation to psychiatric services. This outcome has been linked to both cannabis use and medication adherence in those with first episode of psychosis.9, 12 Other relapse-related outcome measures included the number of relapses; the length of relapse; the time to first relapse; the care intensity at follow-up (rating each patient's intensity of service use over the first 2 years following illness onset; appendix).

on adherence in those with first episode of psychosis.9, 12 Other relapse-related outcome measures included the number of relapses; the length of relapse; the time to first relapse; the care intensity at follow-up (rating each patient's intensity of service use over the first 2 years following illness onset; appendix). We assessed cannabis use as a predictor variable using a modified version of the Cannabis Experience Questionnaire (CEQmv),9 obtaining data on cannabis use over the first 2 years following onset of psychosis. In line with previous work,12 cannabis users were classified on the basis of their pattern of continuation of use after onset, categorising them into different categories (0=not a cannabis user [no use or use only once or twice after onset], 1=intermittent cannabis user [used cannabis more than twice but not every month within the 2-year period], or 2=continued cannabis user [used cannabis every month throughout all of the 24 follow-up months]). The cannabis use variable was coded as ordered categorical. We assessed medication adherence as a mediator variable within the first 2 years after onset by use of the Life Chart Schedule.20 Similar to previous reports,12 the variable was classified on the basis of information on prescription and ratings of adherence (3=non-adherence [67–100% of the time non-compliant]; 2=irregular adherence [34–66% of the time non-compliant]; 1=good adherence [0–33% of the time non-compliant], or 0=medication not prescribed within the 2 years following the onset of illness).

fied on the basis of information on prescription and ratings of adherence (3=non-adherence [67–100% of the time non-compliant]; 2=irregular adherence [34–66% of the time non-compliant]; 1=good adherence [0–33% of the time non-compliant], or 0=medication not prescribed within the 2 years following the onset of illness). Other factors that have previously been reported to be associated with relapse were also included in the model as covariates, including other illicit drug use,21, 22 ethnic origin,4, 9 and care intensity at psychosis onset as an index of illness severity when presenting with the first episode.9, 23 As done in previous studies,12 data from the CEQmv and WHO Life Chart Schedule20 were used to derive the following variables. Other drug use was defined as the use of illicit drugs other than cannabis within the first 2 years after onset. This variable was coded as a categorical variable (2=regular use [six times or more]; 1=experimental use [less than six times]; 0=no use). Care intensities at onset and follow-up were computed by rating each patient's intensity of service use at onset or follow-up, respectively.

is within the first 2 years after onset. This variable was coded as a categorical variable (2=regular use [six times or more]; 1=experimental use [less than six times]; 0=no use). Care intensities at onset and follow-up were computed by rating each patient's intensity of service use at onset or follow-up, respectively. Statistical analysis We created structural equation modelling analyses represented by path diagrams to measure the mediating effect of medication adherence on the association between cannabis use and relapse (appendix). We estimated standardised direct, indirect, and total effects using R and its package Lavaan.24 We estimated bias-corrected 95% CIs using 1000 bootstrap samples. The initial simple models estimated path coefficients for continued cannabis use as a predictor for medication adherence, continued cannabis use as a predictor for relapse and relapse-related outcomes, and medication adherence as a predictor for relapse and relapse-related outcomes. As part of the mediation analysis, a direct effect refers to the standardised path coefficient between continued cannabis use and risk of relapse (path C), and an indirect effect to the product of the standardised path coefficient between path A and path B (figure). The total effect of cannabis use on risk of relapse is the sum of direct and indirect effects. Mediation occurred if the indirect effect was significant. Structural equations for each endogenous variable in the pathway model were adjusted for the potential confounding effects of ethnic origin, other illicit drug use, and illness severity at onset as indexed by the level of care intensity at onset. We aimed to further explore an alternative reverse mediation model to compare with the proposed mediation model. In this reverse mediation model for risk of relapse and related outcomes, continued cannabis use was treated as the mediator variable and medication adherence as the independent variable. It is suggested that the predicted mediation model would be more convincing if the reverse model identifies only non-significant indirect paths.25Figure Path model—medication adherence as a mediator of the association between continued cannabis use and relapse

iable and medication adherence as the independent variable. It is suggested that the predicted mediation model would be more convincing if the reverse model identifies only non-significant indirect paths.25Figure Path model—medication adherence as a mediator of the association between continued cannabis use and relapse Conceptualised pathways between continuation of cannabis use and risk of relapse, with the total effects transmitting both directly (solid line—C), and indirectly (dashed lines—A and B) via medication adherence. Role of the funding source The views expressed are those of the authors and not necessarily those of the NHS, the National Institute of Health Research (NIHR), or the Department of Health. The funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. All authors have approved the final version of the paper.

funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. All authors have approved the final version of the paper. Results 397 patients who presented with their first episode of psychosis between April 12, 2002, and July 26, 2013, were approached for follow-up and had an assessment up until September, 2015. Of the 397 patients, 133 refused to take part in this study and 19 could not be included because of missing data. We followed up 245 patients with first-episode psychosis for 2 years from onset. 91 (37%) of 245 patients with first-episode psychosis had a relapse over the 2 years after onset of psychosis (table 1). Most patients reported regular (45%) or irregular (42%) adherence with the prescribed medication, whereas only a small subset of patients (14%) reported non-compliance. Although most patients were classified as not cannabis users following the onset of psychosis (146 [60%], including 98 [40%] who were never a regular user, and 48 [20%] who had been former regular users), the remaining patients were classified either as intermittent cannabis users (36 patients [15%]) or continued cannabis users (63 patients [26%]). Comparing those who relapsed to those who did not relapse revealed that the relapsing patient group was more likely to be classified as continued cannabis users (p=0·0018, 95% CI 0·09–0·43) and as non-adherent (p<0·0001, 0·18–0·58) or irregularly adherent (p=0·0001, 0·13–0·39) with the prescribed medication. With regard to the other demographic and clinical characteristics, the relapsing group was more likely to be of non-white ethnic origin (p=0·015, 0·034 to 0·30) and appeared to have used other illicit drugs more regularly following the onset of first-episode psychosis (p=0·099, −0·04 to 0·41).Table 1 Sample characteristics

regard to the other demographic and clinical characteristics, the relapsing group was more likely to be of non-white ethnic origin (p=0·015, 0·034 to 0·30) and appeared to have used other illicit drugs more regularly following the onset of first-episode psychosis (p=0·099, −0·04 to 0·41).Table 1 Sample characteristics Total (n=245) Non-relapsing (n=154) Relapsing (n=91) p value* 95% CI† Gender, male 147 (60%) 94 (64%) 53 (36%) 0·77 −0·16 to 0·64 Ethnic origin, non-white 161 (66%) 92 (57%) 69 (43%) 0·015 0·034 to 0·30 Age of onset 28·20 (8·11) 28·43 (8·00) 27·81 (8·33) 0·34 −0·89 to 2·53 Onset diagnosis, non-affective 202 (82%) 125 (62%) 77 (38%) 0·61 −0·22 to 0·11 Care intensity at onset‡ .. .. .. 0·015 .. Referral to community team only 40 (16%) 25 (63%) 15 (38%) .. .. Required contact with crisis team 18 (<1%) 5 (28%) 13 (72%) 0·030 0·05 to 0·64 Required hospital admission (non-compulsory) 74 (30%) 50 (68%) 24 (32%) 0·74 −0·26 to 0·15 Required hospital admission (compulsory) 113 (46%) 74 (66%) 39 (35%) 0·88 −0·22 to 0·16 Other drug use, regular use 26 (11%) 12 (46%) 14 (54%) 0·099 −0·04 to 0·41 Cannabis use‡ .. .. .. 0·0034 .. Never (regular) user 98 (40%) 69 (70%) 29 (30%) .. .. Former (regular) user 48 (20%) 35 (73%) 13 (27%) 0·90 −0·20 to 0·17 Intermittent user 36 (15%) 22 (61%) 14 (39%) 0·42 −0·11 to 0·30 Continued user 63 (26%) 28 (44%) 35 (56%) 0·0018 0·09 to 0·43 Medication adherence‡ .. .. .. <0·0001 .. Regular adherence 109 (45%) 86 (79%) 23 (21%) .. .. Irregular adherence 102 (42%) 54 (53%) 48 (47%) 0·0001 0·13 to 0·39 Non-adherence 34 (14%) 14 (41%) 20 (59%) <0·0001 0·18 to 0·58 Data are n (%) or mean (SD).

0·30 Continued user 63 (26%) 28 (44%) 35 (56%) 0·0018 0·09 to 0·43 Medication adherence‡ .. .. .. <0·0001 .. Regular adherence 109 (45%) 86 (79%) 23 (21%) .. .. Irregular adherence 102 (42%) 54 (53%) 48 (47%) 0·0001 0·13 to 0·39 Non-adherence 34 (14%) 14 (41%) 20 (59%) <0·0001 0·18 to 0·58 Data are n (%) or mean (SD). * χ2 test for independence to compare distributions and Mann-Whitney U (two-sided) test to compare means between the groups (non-relapsing vs relapsing). † 95% CI to estimate the difference between two proportions and two means. ‡ χ2 test for independence and 95% CI estimate of the difference between two proportions for care intensity at onset (reference group=referral to community team only), cannabis use (reference group=never [regular] user), and medication adherence (reference group=regular adherence). The simple path models identified the following associations between the variables of interest (table 2): a significant positive association between the level of cannabis use continuation and the risk of relapse, number of relapses, length of relapse, care intensity index at follow-up, and time until a relapse occurred. The proposed mediator, medication adherence, was linked to both the cannabis use variable and the relapse outcome, including risk of relapse, number of relapses, time until a relapse occurred, and care intensity index, but not length of relapse, suggesting that poor adherence was predictive of worse outcome in the 2 years following onset of psychosis.Table 2 Path estimates for cannabis use and medication adherence as predictors for relapse outcome

f relapse, number of relapses, time until a relapse occurred, and care intensity index, but not length of relapse, suggesting that poor adherence was predictive of worse outcome in the 2 years following onset of psychosis.Table 2 Path estimates for cannabis use and medication adherence as predictors for relapse outcome Cannabis (n=245) Medication adherence (n=245) β (SE) 95%CI p value R2 β (SE) 95%CI p value R2 Risk of relapse 0·35 (0·10) 0·16 to 0·53 0·0003 0·081 0·56 (0·12) 0·33 to 0·80 <0·0001 0·14 Number of relapses 0·28 (0·09) 0·10 to 0·46 0·0019 0·055 0·51 (0·11) 0·28 to 0·73 <0·0001 0·11 Length of relapse 0·53 (0·17) 0·20 to 0·85 0·0016 0·039 0·38 (0·21) −0·02 to 0·79 0·064 0·01 Time until relapse −1·11 (0·50) −2·08 to −0·14 0·024 0·020 −2·33 (0·60) −3·49 to −1·16) <0·0001 0·06 Care intensity index 0·32 (0·09) 0·15 to 0·49 0·0002 0·070 0·52 (0·11) 0·31 to 0·74 <0·0001 0·12 Significant association of cannabis use as a predictor for medication non-adherence (β=0·31, SE=0·11, 95% CI 0·093 to 0·53, p=0·0052, R2=0·046).

11 (0·50) −2·08 to −0·14 0·024 0·020 −2·33 (0·60) −3·49 to −1·16) <0·0001 0·06 Care intensity index 0·32 (0·09) 0·15 to 0·49 0·0002 0·070 0·52 (0·11) 0·31 to 0·74 <0·0001 0·12 Significant association of cannabis use as a predictor for medication non-adherence (β=0·31, SE=0·11, 95% CI 0·093 to 0·53, p=0·0052, R2=0·046). Adjusting all models for ethnic origin, other illicit drug use, and intensity of care at onset, the structural equation modelling analyses are reported in table 3. They revealed that the association between continued cannabis use and risk of relapse was mediated (26·4% as the proportion of total effect mediated) by medication adherence (direct effect: β=0·22, 95% CI 0·03–0·42, p=0·027; indirect effect: β=0·08, 0·004–0·16, p=0·040), suggesting a partial but not full mediation by medication adherence of the effect of cannabis use on risk of relapse. For risk of relapse, the model explained 25% of the variance, so 75% is not explained by this model. Explained variance is the extent to which a statistical model accounts for the variation in the dependent variable. The direct effect is the path between continued cannabis use and risk of relapse, and the indirect effect is the product of the path coefficients of the effect of cannabis use and medication adherence and the effect of medication adherence on risk of relapse. A similar effect was seen for care intensity index at follow-up, for which medication adherence mediated 19·7% of the effect of continued cannabis use, again implicating partial mediation based on significant indirect and direct effects. A larger proportion of the effect of continued cannabis use on the number of relapses of psychosis and time until relapse was mediated by medication adherence. No significant mediation effect was present for length of relapse. The adjusted models explained moderate amounts of variance for outcomes defined as risk of relapse (R2=0·23), number of relapses (R2=0·21), length of relapse (R2=0·07), time until relapse (R2=0·08), and care intensity index (R2=0·15).Table 3 Mediation of the effect of cannabis use on relapse outcome

length of relapse. The adjusted models explained moderate amounts of variance for outcomes defined as risk of relapse (R2=0·23), number of relapses (R2=0·21), length of relapse (R2=0·07), time until relapse (R2=0·08), and care intensity index (R2=0·15).Table 3 Mediation of the effect of cannabis use on relapse outcome β (SE) p value 95% CI* R2 Risk of relapse .. .. .. 0·25 Indirect effect 0·08 (0·04) 0·040 0·004 to 0·16 .. Direct effect 0·22 (0·10) 0·027 0·03 to 0·42 .. Proportion of total effect mediated (%) 26·4% .. .. .. Number of relapses .. .. .. 0·21 Indirect effect 0·07 (0·04) 0·040 0·003 to 0·14 .. Direct effect 0·13 (0·10) 0·20 −0·07 to 0·34 .. Proportion of total effect mediated (%) 35·5% .. .. .. Length of relapse .. .. .. 0·07 Indirect effect 0·03 (0·03) 0·35 −0·03 to 0·09 .. Direct effect 0·40 (0·18) 0·024 0·05 to 0·75 .. Proportion of total effect mediated (%) 6·4% .. .. .. Time until relapse .. .. .. 0·08 Indirect effect −0·26 (0·13) 0·051 −0·53 to 0·001 .. Direct effect −0·67 (0·52) 0·20 −1·68 to 0·35 .. Proportion of total effect mediated (%) 28·3% .. .. .. Care-intensity index .. .. .. 0·15 Indirect effect 0·06 (0·03) 0·035 0·004 to 0·11 .. Direct effect 0·24 (0·09) 0·008 0·06 to 0·42 .. Proportion of total effect mediated (%) 19·7% .. .. .. Pathways for all endogenous variables were adjusted for other illicit drug use, ethnic origin, and onset care intensity. * Bias corrected and accelerated 95% CI 1000 bootstrap samples.

β (SE) p value 95% CI* R2 Risk of relapse .. .. .. 0·25 Indirect effect 0·08 (0·04) 0·040 0·004 to 0·16 .. Direct effect 0·22 (0·10) 0·027 0·03 to 0·42 .. Proportion of total effect mediated (%) 26·4% .. .. .. Number of relapses .. .. .. 0·21 Indirect effect 0·07 (0·04) 0·040 0·003 to 0·14 .. Direct effect 0·13 (0·10) 0·20 −0·07 to 0·34 .. Proportion of total effect mediated (%) 35·5% .. .. .. Length of relapse .. .. .. 0·07 Indirect effect 0·03 (0·03) 0·35 −0·03 to 0·09 .. Direct effect 0·40 (0·18) 0·024 0·05 to 0·75 .. Proportion of total effect mediated (%) 6·4% .. .. .. Time until relapse .. .. .. 0·08 Indirect effect −0·26 (0·13) 0·051 −0·53 to 0·001 .. Direct effect −0·67 (0·52) 0·20 −1·68 to 0·35 .. Proportion of total effect mediated (%) 28·3% .. .. .. Care-intensity index .. .. .. 0·15 Indirect effect 0·06 (0·03) 0·035 0·004 to 0·11 .. Direct effect 0·24 (0·09) 0·008 0·06 to 0·42 .. Proportion of total effect mediated (%) 19·7% .. .. .. Pathways for all endogenous variables were adjusted for other illicit drug use, ethnic origin, and onset care intensity. * Bias corrected and accelerated 95% CI 1000 bootstrap samples. Testing of the alternative models with the cannabis use variable as the proposed mediator indicated that continued cannabis use did not mediate the association between medication non-adherence and the different relapse outcomes after controlling for covariates. We tested two different theoretical models, which were a mediation model that tested whether medication adherence mediated the effect of cannabis use on outcome and a reverse arrow model that tested whether cannabis use mediated the effect of medication adherence on outcome.25 Studies have used this approach to evaluate which of the proposed mediators (medication adherence or cannabis use) is more valid. In those models, medication adherence had a significant direct effect on risk of relapse (β=0·45, 95% CI 0·20–0·70, p=0·0004), number of relapses (β=0·42, 0·18–0·65, p=0·0005), care intensity index at follow-up (β=0·42, 0·19–0·65, p=0·0003), and time until a relapse occurred (β=–2·00, −3·18 to −0·81, p=0·0010), indicating that cannabis use did not fully confound the effects of medication adherence on outcome. There were no indirect effects for risk of relapse, number of relapses, time until a relapse occurred, and care intensity index at follow-up, which further suggested that cannabis use did not mediate the effects of medication adherence on these relapse outcomes. For length of relapse, there were no indirect or direct effects for medication adherence.

or risk of relapse, number of relapses, time until a relapse occurred, and care intensity index at follow-up, which further suggested that cannabis use did not mediate the effects of medication adherence on these relapse outcomes. For length of relapse, there were no indirect or direct effects for medication adherence. Discussion To the best of our knowledge, this is the first study that examines medication adherence as a mediator of the association between continued cannabis use following illness onset and relapse, as indexed by admission to hospital, in patients with first-episode psychosis. The adverse effects of continued cannabis use on risk of relapse were partly but not fully mediated by its association with non-adherence with prescribed antipsychotic medication. More specifically, medication non-adherence mediated the effect of continued cannabis use on risk of relapse (26%), number of relapses (36%), time until a relapse occurred (28%), and care intensity index at follow-up (20%). Medication non-adherence did not mediate the effect of continued cannabis use on length of relapse of psychosis. Our results not only indicate that those patients who continue to use cannabis following onset of their psychotic illness are also more likely to not take medications prescribed for their psychosis but also that this effect can partly explain why patients with first-episode psychosis who continue to use cannabis often suffer from a relapsing form of the illness.7, 26 We8, 9 and others7 have shown that cannabis use, especially continued use after onset of psychosis, is associated with relapse of psychosis resulting in admission to hospital and that this effect is more likely than not to be a causal association.12 This is consistent with other evidence implicating worse outcomes in patients with first-episode psychosis who continued to use cannabis when compared with those who stop using the substance.27 Here, we extend this previous work to show that the adverse effect of continued cannabis use on outcome in early psychosis is partly mediated by an effect on adherence with medication treatment.

patients with first-episode psychosis who continued to use cannabis when compared with those who stop using the substance.27 Here, we extend this previous work to show that the adverse effect of continued cannabis use on outcome in early psychosis is partly mediated by an effect on adherence with medication treatment. These findings are consistent with studies that identified an association between cannabis use and medication adherence,8, 9, 15, 28 as well as between medication adherence and increased risk of relapse of psychosis.10, 11 We have reported8 that failure of treatment with antipsychotic medication as indexed by the number of unique antipsychotic prescriptions could partly mediate the adverse effect of cannabis use on subsequent risk of relapse in first-episode psychosis. Although a change of antipsychotic medication could reflect a clinical judgment of failed treatment, several separate considerations either alone or in combination could lead to such a judgment, including one of treatment resistance, poor tolerability, or non-adherence to a specific antipsychotic. Until now, it has not been known which of these factors might explain how cannabis use could increase the risk of relapse. Results from the present study clearly point toward a mediating influence of poor medication adherence. Whether treatment resistance or poor tolerability also mediate some of the effects of cannabis use on relapse of psychosis is yet to be tested. Furthermore, other factors, such as depressive symptoms29 or cognitive function30 that were not systematically investigated in this study could also have influenced the association between cannabis use and risk of relapse.

ity also mediate some of the effects of cannabis use on relapse of psychosis is yet to be tested. Furthermore, other factors, such as depressive symptoms29 or cognitive function30 that were not systematically investigated in this study could also have influenced the association between cannabis use and risk of relapse. Overall, our results suggest that although efforts should no doubt continue to develop more effective interventions to help patients with psychosis to reduce their cannabis use—eg, similar to those cannabis-focused treatment programmes that are currently under assessment,31 another potential approach to mitigating the harm from cannabis use might lie in ensuring better adherence of patients to their prescribed medication. It is worth noting that despite the identified mediation effect, a considerable proportion of the variance in the risk of relapse and related outcomes remains unexplained, varying between 7% and 25% depending on the specific outcome. Future studies including much larger samples are needed to consider other risk factors of interest as well as more complex model pathways to address the issue of unexplained variance in relapse outcome. In this context, it should be pointed out that the identified associations could also be bidirectional. It is worth noting that as the present study was an observational study, temporal ambiguity between the mediator and predictor variable as well as unmeasured confounders could have biased our results. Nevertheless, to partly address this limitation of absence of experimental data, we compared the proposed mediation model with an alternative path model with reversed arrows (ie, by including cannabis use as a mediating factor instead of medication adherence) but the results were not supportive of alternative path models that included cannabis use as a mediator of the associations between medication adherence and relapse outcome (for path estimates see appendix). Although other limitations of this study might relate to the nature of the retrospective assessment of cannabis use and medication adherence, and the inclusion of a selective subset of inner city patients with first-episode psychosis who were at least 18 years old, those issues are unlikely to have affected the results (appendix). We did not consider those who started using cannabis after the onset of psychosis but had no history of premorbid regular use as a separate group, since only three participants belonged to this category.

episode psychosis who were at least 18 years old, those issues are unlikely to have affected the results (appendix). We did not consider those who started using cannabis after the onset of psychosis but had no history of premorbid regular use as a separate group, since only three participants belonged to this category. How continued cannabis use might have resulted in poor adherence to medications in psychosis patients is unclear. Although it is possible that increased severity of psychosis,7 and consequently, impaired insight or memory32 as a result of continued cannabis use might explain poor adherence, this possibility was not investigated in the present study and warrants investigation in the future. Our results suggest that up to a third of the adverse effect of cannabis use on outcome in first-episode psychosis could be mediated through its effect on medication adherence, suggesting that interventions aimed at improving medication adherence might partly help mitigate the adverse effects of cannabis use on outcome in psychosis. Supplementary Material Supplementary appendix Acknowledgments SB is funded through an NIHR Clinician Scientist award (NIHR-CS-011-001). The fee for open access was paid by King's College London.

Our results suggest that up to a third of the adverse effect of cannabis use on outcome in first-episode psychosis could be mediated through its effect on medication adherence, suggesting that interventions aimed at improving medication adherence might partly help mitigate the adverse effects of cannabis use on outcome in psychosis. Supplementary Material Supplementary appendix Acknowledgments SB is funded through an NIHR Clinician Scientist award (NIHR-CS-011-001). The fee for open access was paid by King's College London. Contributors SB and TS had full access to all the data in the study and take full responsibility for the integrity of the data and accuracy of the data analysis. SB designed the study and supervised the data collection and analysis. TS analysed the data and wrote the first draft of the manuscript together with SB. All other authors provided data, reviewed the results, and contributed to the final draft of the manuscript. Declaration of interests RM reports honoraria and speakers fees from Lundbeck, Janssen, Sunovian, and Otsuka, outside the submitted work. All other authors declare no competing interests.

Introduction Sleep problems are a common occurrence in patients with mental health disorders. The traditional view is that disrupted sleep is a symptom, consequence, or non-specific epiphenomenon of the disorders; the clinical result is that the treatment of sleep problems is given a low priority. An alternative perspective is that disturbed sleep is a contributory causal factor in the occurrence of many mental health disorders.1 An escalating cycle then emerges between the distress of the mental health symptoms, effect on daytime functioning, and struggles in gaining restorative sleep. From this alternative perspective, the treatment of sleep problems attains a higher clinical importance. We are particularly interested in the putative causal association between disturbed sleep and psychotic experiences.2, 3 The interventionist–causal model approach to establishment of a causal association is to manipulate the hypothesised mechanistic variable and observe the effect on the outcome of interest;4 if a causal association exists then the outcome variable should alter. The effects of the manipulation can then be substantiated further by use of mediation analysis.5, 6 In the present study, we aimed to improve sleep in individuals with insomnia to determine the effect on psychotic experiences. This approach therefore informs both theoretical understanding and clinical practice.

alter. The effects of the manipulation can then be substantiated further by use of mediation analysis.5, 6 In the present study, we aimed to improve sleep in individuals with insomnia to determine the effect on psychotic experiences. This approach therefore informs both theoretical understanding and clinical practice. The most common form of sleep disruption is insomnia, comprising sustained difficulties in initiating or staying asleep, or both, which cause problems during the day. The association of insomnia with psychotic experiences in the general population has been established.3 There are multiple, independent, psychotic experiences. Each psychotic experience exists on a spectrum of severity in the general population with differing heritability and differing strength of association with insomnia.7 Paranoia and hallucinations have the strongest links with insomnia.2, 7, 8 However, the effect of altering the amount of sleep disruption—eg, by targeted sleep treatment—on these psychotic experiences remains to be established. Clinical guidelines recommend the use of cognitive behavioural therapy (CBT) as the first-line treatment for insomnia.9 Digital forms of CBT for insomnia that require no therapist to be present have been shown to be efficacious as well.10, 11, 12 In patients with current delusions and hallucinations, results of our pilot randomised controlled trial13 have shown that insomnia can be substantially reduced with CBT, but the trial was underpowered to establish with sufficient precision the consequences for psychotic experiences. Therefore, we undertook a clinical trial that was large enough to definitively test the causal association between insomnia and self-reported psychotic experiences.

ia can be substantially reduced with CBT, but the trial was underpowered to establish with sufficient precision the consequences for psychotic experiences. Therefore, we undertook a clinical trial that was large enough to definitively test the causal association between insomnia and self-reported psychotic experiences. Research in context Evidence before this study If insomnia is a contributory cause of psychotic experiences, then the key test is whether improving sleep will lead to a reduction in psychotic experiences. We therefore searched for randomised controlled studies that set out to reduce insomnia and examine the effects on psychotic experiences. On June 23, 2017, we searched the entire archive (ie, using no date restrictions) of PubMed for: (Sleep OR Insomnia) AND (Delus* OR Hallucinat* OR Psychosis OR Psychotic OR Schizophren*) AND (CBT OR hypnotic OR medication) AND (Random* OR RCT). 130 papers were identified and only two were randomised controlled trials that tested the effects of sleep treatment on psychotic experiences, with the larger of the trials being our own with 50 patients with schizophrenia or related disorders. These trials were underpowered to determine with any precision the potential link between insomnia and psychotic experiences. Added value of this study

If insomnia is a contributory cause of psychotic experiences, then the key test is whether improving sleep will lead to a reduction in psychotic experiences. We therefore searched for randomised controlled studies that set out to reduce insomnia and examine the effects on psychotic experiences. On June 23, 2017, we searched the entire archive (ie, using no date restrictions) of PubMed for: (Sleep OR Insomnia) AND (Delus* OR Hallucinat* OR Psychosis OR Psychotic OR Schizophren*) AND (CBT OR hypnotic OR medication) AND (Random* OR RCT). 130 papers were identified and only two were randomised controlled trials that tested the effects of sleep treatment on psychotic experiences, with the larger of the trials being our own with 50 patients with schizophrenia or related disorders. These trials were underpowered to determine with any precision the potential link between insomnia and psychotic experiences. Added value of this study We undertook what might be the largest randomised controlled trial to date of a psychological treatment. It is the first study adequately powered to determine the effects of treating sleep dysfunction on psychotic experiences. It shows very clearly that treatment of insomnia in students leads to a reduction in psychotic experiences. A mediation analysis supports this interpretation. Furthermore, the trial is consistent with a small number of other randomised controlled trials that indicate multiple other benefits for mental health when treating sleep problems. Implications of all the available evidence

We undertook what might be the largest randomised controlled trial to date of a psychological treatment. It is the first study adequately powered to determine the effects of treating sleep dysfunction on psychotic experiences. It shows very clearly that treatment of insomnia in students leads to a reduction in psychotic experiences. A mediation analysis supports this interpretation. Furthermore, the trial is consistent with a small number of other randomised controlled trials that indicate multiple other benefits for mental health when treating sleep problems. Implications of all the available evidence Sleep disruption might have a contributory causal role in the occurrence of psychotic experiences and a wide range of other mental health problems. Adequately powered tests in other populations would be helpful, but research indicating that the treatment of disrupted sleep requires a higher priority in mental health provision is accumulating.

ave a contributory causal role in the occurrence of psychotic experiences and a wide range of other mental health problems. Adequately powered tests in other populations would be helpful, but research indicating that the treatment of disrupted sleep requires a higher priority in mental health provision is accumulating. To test thousands of individuals, we did an online study using digital CBT for insomnia treatment. A participant pool (university students) was selected that would be easily reachable (since we would have access to large email lists) and at an age when mental health disorders emerge. We have previously shown in a student population that sleep problems are associated with elevated levels of paranoia, hallucinations, anxiety, depression, and manic symptoms.8 Our main aim required a comparison between the effects of a reduction in insomnia (in a group receiving recommended treatment) and continued sleep disruption (in a group receiving usual care, which is likely to mean receiving little or no treatment). Clear change in sleep in one group relative to another was required to test the mechanistic hypothesis. We were not investigating the intervention elements that might lead to change. Our primary aim was to find out whether CBT for insomnia, compared with a usual practice control group, reduced insomnia and reduced paranoia and hallucinations by the end of treatment, and whether the changes in insomnia mediated the changes in psychotic experiences. We also aimed to determine the potential effects of sleep improvement on a wider range of mental health outcomes in this general population group. Our secondary aims were to investigate whether digital CBT for insomnia, compared with usual practice, reduced depression, anxiety, nightmares, and mania; improved psychological wellbeing; and led to the occurrence of fewer mental health disorders.

wider range of mental health outcomes in this general population group. Our secondary aims were to investigate whether digital CBT for insomnia, compared with usual practice, reduced depression, anxiety, nightmares, and mania; improved psychological wellbeing; and led to the occurrence of fewer mental health disorders. Methods Study design and participants We did this single-blind, randomised controlled trial (OASIS; Oxford Access for Students Improving Sleep) of digital CBT versus treatment as usual (usual practice). Screening, informed consent, assessments, allocation to condition, and the delivery of the intervention were carried out online using an automated system, a specially configured instance of True Colours, which is a system for the scheduled collection of outcome measures.14 Participants in the control group were given access to the sleep intervention after their final assessment. The study received overall ethical approval from the University of Oxford Medical Sciences Inter-Divisional Ethics Committee and then local approvals at the other participating universities. The OASIS trial protocol has been published.15

ol group were given access to the sleep intervention after their final assessment. The study received overall ethical approval from the University of Oxford Medical Sciences Inter-Divisional Ethics Committee and then local approvals at the other participating universities. The OASIS trial protocol has been published.15 Participants were eligible if they were attending university; had a positive screen for insomnia, as indicated by a score of 16 or lower on the Sleep Condition Indicator (SCI);16 and were 18 years or older. We had no exclusion criteria. 26 UK universities took part (appendix 1), ensuring a range in geographical locations and academic ability. The principal method of recruitment was sending a circular email within universities that contained a link to the web-based screening. When a circular email was not possible, recruitment was via advertisment on websites and displaying posters, or both. Recruitment began on March 5, 2015, and ended on Feb 17, 2016. We collected the final data on July 28, 2016. Randomisation and masking As recommended for large clinical trials,17 we used simple randomisation (1:1) with an automated online system, ensuring that the research team was unable to affect randomisation. Participants completed all the assessments independently online and therefore their responses could not be affected by the trial team.

sking As recommended for large clinical trials,17 we used simple randomisation (1:1) with an automated online system, ensuring that the research team was unable to affect randomisation. Participants completed all the assessments independently online and therefore their responses could not be affected by the trial team. Procedures Assessments took place at weeks 0 (baseline), 3, 10 (end of therapy), and 22. The week 3 assessment comprised only the primary outcome measures. The week 3 assessment was carried out to assess in the mediation analyses the temporal order of changes. Participants received an email prompt to complete the assessments online. The order of the assessments was consistent across timepoints. If participants did not complete the assessment, then they received up to two email reminders 2 days apart.

ried out to assess in the mediation analyses the temporal order of changes. Participants received an email prompt to complete the assessments online. The order of the assessments was consistent across timepoints. If participants did not complete the assessment, then they received up to two email reminders 2 days apart. The CBT for insomnia intervention is called Sleepio.11, 18 It is provided in six sessions, lasting an average of 20 min each. Sessions are unlocked weekly, although participants can move at a slower pace. The full programme is accessible via any web browser and all participants start the programme online. Certain tools (eg, sleep diaries and relaxation audios) could also be accessed using the web browser of a smartphone. All of the sessions, sleep diaries, relaxation audios, and the scheduling tool could be accessed with an iPhone. Completion of an initial assessment drives the algorithms that personalise the programme. For example, the assessment leads to a tailored choice of treatment goal, with progress then reviewed at each subsequent session. The treatment includes behavioural, cognitive, and educational components. The behavioural techniques include sleep restriction (ie, reducing the sleep window to enhance sleep consolidation), stimulus control (eg, getting out of bed after 15–20 min of wakefulness), and relaxation (eg, tensing and relaxing muscles when in bed). The cognitive techniques include paradoxical intention (eg, trying to stay awake), belief restructuring (eg, targeting unrealistic expectations about sleep), mindfulness (eg, acknowledging thoughts and feelings without dwelling on them), imagery (eg, generating positive mental images), and putting the day to rest (eg, setting time aside to reflect on the day). The educational component covers information about the processes of sleep and sleep hygiene. The programme is interactive, and content is presented by an animated therapist. Participants make a time for the session and are prompted via email or text message via a short message service if they do not attend. Participants complete daily sleep diaries throughout the intervention, which are used by the programme to tailor the advice. Sleep restriction is introduced in the third session of the course. The animated therapist proposes a new sleep window, which is calculated from the sleep diary data, and engages with the participant to help them select the timing of the window (eg, earlier versus later in the night).

he programme to tailor the advice. Sleep restriction is introduced in the third session of the course. The animated therapist proposes a new sleep window, which is calculated from the sleep diary data, and engages with the participant to help them select the timing of the window (eg, earlier versus later in the night). A more lenient sleep window is used for those individuals reporting substantial physical problems, other mental health problems, or moderate-to-severe sleepiness. The sleep window is regularly reviewed at each session after it has been introduced. If the sleep diary data indicate a sleep efficiency of 90% or higher, the animated therapist advises that 15 min is added to the sleep window. Throughout the course of therapy, participants had access to a moderated online community and an online library of information about sleep. Participants could also view their online case file, which included four sections: a progress review, a reminder of strategies, an agreed sleep schedule, and a list of further reading. Usual practice (treatment as usual) referred to the current care that the participants were receiving. The amount of treatment input was likely to be minimal, with prescription of medication for a small proportion. We did not attempt to alter the current care that participants received.

, and a list of further reading. Usual practice (treatment as usual) referred to the current care that the participants were receiving. The amount of treatment input was likely to be minimal, with prescription of medication for a small proportion. We did not attempt to alter the current care that participants received. Outcomes The primary outcomes were insomnia, paranoia, and hallucinatory experiences. The primary measure for insomnia was the SCI-8.16 This score is an 8-item measure, validated against DSM-5 criteria, assessing sleep and its impact on daytime functioning over the past week. Scores can range from 0 to 32 with higher scores indicating better sleep. A clinical cutoff of less than 17 correctly identifies 89% of individuals with probable insomnia disorder. We used a version of the SCI that included one additional question, as a secondary outcome, regarding early morning waking. The internal consistency (Cronbach's α at baseline) of the scale in the present study was 0·63. Paranoia was assessed with the Green et al Paranoid Thought Scales (GPTS), part B.19 This scale assesses persecutory ideation, and the timeframe used was the past fortnight. The scale comprises 16 items, each rated on a 1 (not at all) to 5 (totally) scale. High scores indicate higher levels of paranoia. The internal consistency of the scale in the present study was 0·94.

hought Scales (GPTS), part B.19 This scale assesses persecutory ideation, and the timeframe used was the past fortnight. The scale comprises 16 items, each rated on a 1 (not at all) to 5 (totally) scale. High scores indicate higher levels of paranoia. The internal consistency of the scale in the present study was 0·94. The measure for hallucinations was the Specific Psychotic Experiences Questionnaire—Hallucinations subscale.20 The scale comprises nine items rated on a 0 (not at all) to 5 (more than once per day) scale. The timeframe was the past fortnight. Higher scores indicate greater occurrences of hallucinatory experiences. The internal consistency of the scale in the present study was 0·93. The secondary outcome measures for sleep were the Insomnia Severity Index (ISI;21 Cronbach's α in the present study=0·67) and the Disturbing Dreams and Nightmare Severity Index22 (Cronbach's α in the present study=0·91). The secondary outcome measure for psychotic experiences was the 16-item version of the Prodromal Questionnaire23 (Cronbach's α in the present study=0·79). A score of 6 or more has 87% specificity and 87% sensitivity to correctly classify ultra-high risk for psychosis mental states in a help-seeking sample.

study=0·91). The secondary outcome measure for psychotic experiences was the 16-item version of the Prodromal Questionnaire23 (Cronbach's α in the present study=0·79). A score of 6 or more has 87% specificity and 87% sensitivity to correctly classify ultra-high risk for psychosis mental states in a help-seeking sample. The measures to assess affective symptoms were the Patient Health Questionnaire 9-item version24 (PHQ-9; Cronbach's α in the present study=0·85), the Generalised Anxiety Disorder 7-item version25 (GAD-7; Cronbach's α in the present study=0·89), and the Altman Mania Scale26 (Cronbach's α in the present study=0·64). Psychological wellbeing was assessed with the Warwick–Edinburgh Mental Wellbeing Scale (WEMWBS;27 Cronbach's α in the present study=0·89), and the Work and Social Adjustment Scale (WSAS;28 Cronbach's α in the present study=0·83). To assess the development of mental health disorders, we used established cutoffs on the Prodromal Questionnaire,23 Altman Mania Scale,26 PHQ,24 and GAD-7.25 Participants were also asked at each assessment timepoint whether they were in contact with mental health services, had received a mental health diagnosis, took medication for a mental health problem, or were currently receiving any other psychological therapy. If the trial team were informed of the occurrence of a serious adverse event for a trial participant, then this was recorded. Serious adverse events were defined as deaths, suicide attempts, serious violent incidents, admissions to secure units, and formal complaints about the online intervention.

To assess the development of mental health disorders, we used established cutoffs on the Prodromal Questionnaire,23 Altman Mania Scale,26 PHQ,24 and GAD-7.25 Participants were also asked at each assessment timepoint whether they were in contact with mental health services, had received a mental health diagnosis, took medication for a mental health problem, or were currently receiving any other psychological therapy. If the trial team were informed of the occurrence of a serious adverse event for a trial participant, then this was recorded. Serious adverse events were defined as deaths, suicide attempts, serious violent incidents, admissions to secure units, and formal complaints about the online intervention. Statistical analysis We calculated the sample size on the basis of change in paranoia (one of the two psychotic symptoms studied), since following this intervention change would be expected to be lower in psychotic experiences than in insomnia. Based on the SDs observed from a previous study29 for the GPTS (SD 10·4), a total sample size of 2614 participants (ie, 1307 per group) would provide 90% power to detect a small effect size in paranoia, with a standardised mean difference of 0·15, while accounting for a high amount of expected attrition (40%). In a study amendment, the sample size was increased because of a higher than initially expected dropout rate.

participants (ie, 1307 per group) would provide 90% power to detect a small effect size in paranoia, with a standardised mean difference of 0·15, while accounting for a high amount of expected attrition (40%). In a study amendment, the sample size was increased because of a higher than initially expected dropout rate. An outline of the analysis strategy was provided in the published trial protocol15 and a full statistical analysis plan was agreed before the trial analysis (appendix 2). All the analyses were validated by a second statistician. Analyses were by intention to treat and were carried out at the end of the last follow-up assessment (with no interim analyses). We analysed each continuous outcome with a linear mixed effects regression model to account for the repeated measures over time, and we analysed binary outcomes with a logistic mixed effects model. Mixed effects models are the recommended statistical technique for analysing clinical trials when outcomes are collected at repeated timepoints,30 and in this trial included outcome data at weeks 3, 10, and 22 available for all participants who had been randomly assigned. The method has the advantage of implicitly accounting for data missing at random. The estimated (adjusted) treatment differences from these analyses are therefore reported. The linear mixed effects models included the outcome as the response variable, timepoint, randomised group, and baseline score as fixed effects and random effects were estimated for students nested within universities. A student is located within one university, and so to estimate the random intercepts, we accounted for random variation between universities and between students within the same university. We modelled an interaction between time and randomised group as a fixed effect to allow estimation of treatment effect at all three timepoints. Sex and course level were included as covariates in the model. We used an unstructured variance–covariance matrix to model the within-subject error correlation structure. Results are presented as mean adjusted differences in scores between the randomised groups, with 95% CIs and associated two-sided p values. We confirmed the normality assumption of the residuals for each outcome. No deviations from normality were apparent and therefore maximum likelihood estimates were reported.

cture. Results are presented as mean adjusted differences in scores between the randomised groups, with 95% CIs and associated two-sided p values. We confirmed the normality assumption of the residuals for each outcome. No deviations from normality were apparent and therefore maximum likelihood estimates were reported. We did sensitivity analyses (pattern mixture models, inclusion of baseline covariates predictive of missing data, and imputation) for the three main outcomes, examining the robustness of the results to different assumptions regarding missing data. We calculated standardised effect sizes with Cohen's d, dividing the treatment effect by the shared SD at baseline. We used similar logistic mixed effects models for the secondary binary outcomes.

tion) for the three main outcomes, examining the robustness of the results to different assumptions regarding missing data. We calculated standardised effect sizes with Cohen's d, dividing the treatment effect by the shared SD at baseline. We used similar logistic mixed effects models for the secondary binary outcomes. To test the mediation hypotheses, we determined the extent of mediation of the week 3 and week 10 insomnia scores on the week 10 paranoia and hallucination outcomes. The approach used was similar to the method of Baron and Kenny,5, 31 but made use of linear mixed effects models at each step. The approach involved four steps and three separate model fits. In two separate linear mixed effects models, the intervention was shown to be correlated with the outcome and then with the mediator. We then fitted the data to a third model with the outcome as the response and both the intervention and mediator as covariates. The parameters were extracted as per Baron and Kenny31 to obtain the total, direct, and indirect effects, and finally the percentage mediation was determined. In all models, we included baseline amounts of both the outcome and mediator as covariates. This method is similar to the mediation analysis in the study by Freeman and colleagues,32 but made use of linear mixed effects models to account for repeated measurements, rather than through structural equation modelling. We used STATA version 14.1 for the statistical analysis. The trial is registered with the ISRCTN registry, number ISRCTN61272251.

To test the mediation hypotheses, we determined the extent of mediation of the week 3 and week 10 insomnia scores on the week 10 paranoia and hallucination outcomes. The approach used was similar to the method of Baron and Kenny,5, 31 but made use of linear mixed effects models at each step. The approach involved four steps and three separate model fits. In two separate linear mixed effects models, the intervention was shown to be correlated with the outcome and then with the mediator. We then fitted the data to a third model with the outcome as the response and both the intervention and mediator as covariates. The parameters were extracted as per Baron and Kenny31 to obtain the total, direct, and indirect effects, and finally the percentage mediation was determined. In all models, we included baseline amounts of both the outcome and mediator as covariates. This method is similar to the mediation analysis in the study by Freeman and colleagues,32 but made use of linear mixed effects models to account for repeated measurements, rather than through structural equation modelling. We used STATA version 14.1 for the statistical analysis. The trial is registered with the ISRCTN registry, number ISRCTN61272251. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

The trial is registered with the ISRCTN registry, number ISRCTN61272251. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between March 5, 2015, and Feb 17, 2016, we randomly assigned 3755 participants to receive digital CBT for insomnia (n=1891) or usual practice (n=1864; figure). The sample was predominately female, studying for their first university degree, and two-thirds were of white British ethnicity (table 1). Around a fifth of the participants were in contact with mental health services (table 1). The two randomised groups were well matched at baseline (table 1).Figure Trial profile SCI=Sleep Condition Indicator. *Some participants excluded for two or more reasons. †Not all participants accessed all treatment sessions. ‡Had at least one measurement at week 3, 10, or 22. Table 1 Baseline characteristics

Results Between March 5, 2015, and Feb 17, 2016, we randomly assigned 3755 participants to receive digital CBT for insomnia (n=1891) or usual practice (n=1864; figure). The sample was predominately female, studying for their first university degree, and two-thirds were of white British ethnicity (table 1). Around a fifth of the participants were in contact with mental health services (table 1). The two randomised groups were well matched at baseline (table 1).Figure Trial profile SCI=Sleep Condition Indicator. *Some participants excluded for two or more reasons. †Not all participants accessed all treatment sessions. ‡Had at least one measurement at week 3, 10, or 22. Table 1 Baseline characteristics Control group (n=1864) Treatment group (n=1891) Mean age (years) 24·6 (7·6) 24·8 (7·7) Mean total UCAS points 753·0 (517·3) 720·8 (456·3) Sex Male 530 (28%) 513 (27%) Female 1315 (71%) 1361 (72%) Other 19 (1%) 17 (1%) Course level Undergraduate 1352 (73%) 1389 (73%) Postgraduate 480 (26%) 473 (25%) Other 32 (2%) 29 (2%) Ethnic origin White British 1212 (65%) 1265 (67%) Irish 32 (2%) 27 (1%) Other 284 (15%) 261 (14%) Mixed White and Caribbean 13 (1%) 11 (1%) White and African 9 (<1%) 13 (1%) White and Asian 31 (2%) 27 (1%) Other 36 (2%) 29 (2%) Asian Indian 26 (1%) 43 (2%) Pakistani 23 (1%) 22 (1%) Bangladeshi 9 (<1%) 7 (<1%) Chinese 95 (5%) 73 (4%) Other 25 (1%) 32 (2%) Black African 26 (1%) 23 (1%) Caribbean 10 (1%) 17 (1%) Other 2 (<1%) 3 (<1%) Arab 12 (1%) 14 (1%) Other 19 (1%) 24 (1%) Mean insomnia (SCI-8) score 10·1 (4·3) 9·9 (4·3) Mean paranoia (GPTS) score 24·8 (11·6) 25·4 (11·9) Mean hallucinations (SPEQ) score 5·3 (6·9) 5·3 (6·4) Mean insomnia (SCI-9) score 12·1 (4·9) 11·9 (4·8) Mean insomnia (ISI) score 15·3 (4·0) 15·4 (3·9) Mean nightmares (DDNSI) score 8·1 (8·2) 7·7 (7·8) Mean prodromal psychosis (PQ-16) score 4·9 (3·4) 4·8 (3·3) Mean depression (PHQ-9) score 12·7 (5·9) 12·9 (5·8) Mean anxiety (GAD-7) score 9·0 (5·6) 9·4 (5·6) Mean mania (Altman) score 3·5 (3·0) 3·5 (3·0) Mean functioning (WSAS) score 17·7 (7·6) 17·6 (7·6) Mean wellbeing (WEMWBS) score 37·9 (8·8) 37·8 (8·5) Ultra-high risk of psychosis (PQ-16) cutoff (≥6) Above 706 (38%) 711 (38%) Below 1158 (62%) 1180 (62%) Depressive disorder (PHQ-9) cutoff (≥10) Above 1238 (66%) 1286 (68%) Below 626 (34%) 605 (32%) Anxiety disorder (GAD-7) cutoff (≥10) Above 781 (42%) 880 (47%) Below 1083 (58%) 1011 (53%) Mania disorder (Altman) score cutoff (≥6) Above 422 (23%) 413 (22%) Below 1442 (77%) 1478 (78%) Contact with mental health services Yes 328 (18%) 346 (18%) No 1536 (82%) 1545 (82%) Any psychiatric diagnosis Yes 590 (32%) 641 (34%) No 1274 (68%) 1250 (66%) Previous diagnosis of a sleep disorder Yes 93 (5%) 96 (5%) No 1771 (95%) 1795 (95%) Any psychiatric medication Yes 433 (23%) 460 (24%) No 1431 (77%) 1431 (76%) Specific medication for a sleep disorder Yes 51 (3%) 55 (3%) No 1813 (97%) 1836 (97%)

Any psychiatric diagnosis Yes 590 (32%) 641 (34%) No 1274 (68%) 1250 (66%) Previous diagnosis of a sleep disorder Yes 93 (5%) 96 (5%) No 1771 (95%) 1795 (95%) Any psychiatric medication Yes 433 (23%) 460 (24%) No 1431 (77%) 1431 (76%) Specific medication for a sleep disorder Yes 51 (3%) 55 (3%) No 1813 (97%) 1836 (97%) Psychological therapy Yes 146 (8%) 135 (7%) No 1718 (92%) 1756 (93%) Data are mean (SD) or n (%). UCAS=Universities and Colleges Admissions Service. SCI-8=Sleep Condition Indicator 8-item version. GPTS=Green et al Paranoid Thought Scales. SPEQ=Specific Psychotic Experiences Questionnaire. SCI-9=SCI 9-item version. ISI=Insomnia Severity Index. DDNSI=Disturbing Dream and Nightmare Severity Index. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. WSAS=Work and Social Adjustment Scale. WEMWBS=Warwick–Edinburgh Mental Wellbeing Scale.

. ISI=Insomnia Severity Index. DDNSI=Disturbing Dream and Nightmare Severity Index. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. WSAS=Work and Social Adjustment Scale. WEMWBS=Warwick–Edinburgh Mental Wellbeing Scale. Appendix 1 has the full details of the sample, missing data patterns, sensitivity analyses, and all other analyses. The dropout from the study assessments was high (50%) during the course of the study, and was greater in the treatment group than in the control group, with this figure and pattern almost identical to the most comparable previous study.12 The baseline scores for the three primary outcomes (insomnia, paranoia, and hallucinations) were not associated with later missingness (appendix 1). Compared with participants who remained in the study, participants who dropped out from both groups were younger in age and more likely to be male (appendix 1). For the secondary measures, ISI, PHQ-9, Altman Mania Scale, and WSAS scores were slightly higher, and the WEMWBS score was slightly lower, in the missing groups than the non-missing groups (appendix 1). Treatment uptake was relatively low. In the intervention group, 1302 participants (69%) logged on for at least one treatment session, 953 (50%) accessed at least two sessions, 672 (36%) accessed at least three sessions, 497 (26%) accessed at least four sessions, 390 (21%) accessed at least five sessions, and 331 (18%) accessed six sessions (figure).

low. In the intervention group, 1302 participants (69%) logged on for at least one treatment session, 953 (50%) accessed at least two sessions, 672 (36%) accessed at least three sessions, 497 (26%) accessed at least four sessions, 390 (21%) accessed at least five sessions, and 331 (18%) accessed six sessions (figure). Regarding the primary measures, the sleep treatment was associated with significant reductions, at all timepoints, in insomnia, paranoia, and hallucinations compared with the control group (all p<0·0001; table 2). The reduction in insomnia after treatment was large, while the reduction in psychotic experiences was small (table 2). After treatment, 454 (62%) of 733 individuals in the treatment group and 326 (29%) of 1142 individuals in the control group scored outside the clinical cutoff for insomnia used for trial entry. The treatment differences were robust to the three different types of sensitivity analyses for missing data (appendix 1). A conservative imputation (given the general improvement in scores in both groups) of missing data was used, whereby the last available measurement for a participant was imputed for all further missing measurements of that participant. All three primary outcome differences remained significant with last observation carried forward imputations (appendix 1). Treatment differences also remained consistent with the primary analysis when we repeated the main analyses covarying for baseline variables that predicted missingness for each outcome. We also used pattern mixture models. Treatment differences would still be significant assuming the missing individuals in the treatment group had outcome scores 2 points worse for insomnia and paranoia, and 1 point worse for hallucinations (predictably the hallucination scale scores were much lower than for the other two outcome variables; appendix 1).Table 2 Primary outcome results

still be significant assuming the missing individuals in the treatment group had outcome scores 2 points worse for insomnia and paranoia, and 1 point worse for hallucinations (predictably the hallucination scale scores were much lower than for the other two outcome variables; appendix 1).Table 2 Primary outcome results Insomnia (SCI-8) Paranoia (GPTS) Hallucinations (SPEQ) Unadjusted mean Adjusted difference* (95% CI), d† p value* Unadjusted mean Adjusted difference* (95% CI), d† p value* Unadjusted mean Adjusted difference* (95% CI), d† p value* Control Treatment Control Treatment Control Treatment Week 3 12·34 (5·85) 14·96 (5·80) 2·62 (2·19 to 3·06), 0·61 <0·0001 24·63 (11·82) 22·61 (9·89) −1·81 (−2·49 to −1·13), 0·15 <0·0001 5·06 (6·89) 4·06 (5·84) −0·79 (−1·15 to −0·42), 0·12 <0·0001 Week 10 13·31 (6·45) 18·08 (6·66) 4·78 (4·29 to 5·26), 1·11 <0·0001 23·84 (12·16) 21·06 (9·08) −2·22 (−2·98 to −1·45), 0·19 <0·0001 4·89 (7·24) 3·12 (5·12) −1·58 (−1·98 to −1·18), 0·24 <0·0001 Week 22 14·43 (6·71) 19·27 (7·13) 4·81 (4·29 to 5·33), 1·12 <0·0001 23·84 (12·68) 20·75 (9·19) −2·78 (−3·60 to −1·96), 0·24 <0·0001 4·71 (7·43) 2·87 (5·45) −1·56 (−1·99 to −1·14), 0·23 <0·0001 Data are mean (SD). At week 3, 1398 participants were in the control group and 1044 participants were in the treatment group. At week 10, 1142 participants were in the control group and 733 participants were in the treatment group. At week 22, 971 participants were in the control group and 603 participants were in the treatment group. SCI-8=Sleep Condition Indicator 8-item version. GPTS=Green et al Paranoid Thought Scales. SPEQ=Specific Psychotic Experiences Questionnaire.

ere in the control group and 733 participants were in the treatment group. At week 22, 971 participants were in the control group and 603 participants were in the treatment group. SCI-8=Sleep Condition Indicator 8-item version. GPTS=Green et al Paranoid Thought Scales. SPEQ=Specific Psychotic Experiences Questionnaire. * Linear mixed effects model adjusted for gender, student status, week, and interaction of week with randomisation, and including a random effect for student within university. Covariance matrix of within subject measurements was unstructured. † d is standardised effect size (Cohen's d).

ere in the control group and 733 participants were in the treatment group. At week 22, 971 participants were in the control group and 603 participants were in the treatment group. SCI-8=Sleep Condition Indicator 8-item version. GPTS=Green et al Paranoid Thought Scales. SPEQ=Specific Psychotic Experiences Questionnaire. * Linear mixed effects model adjusted for gender, student status, week, and interaction of week with randomisation, and including a random effect for student within university. Covariance matrix of within subject measurements was unstructured. † d is standardised effect size (Cohen's d). For the mediation analyses, change in sleep over 3 weeks explained 30% of the intervention effect on paranoia at 10 weeks, with change in sleep over 10 weeks accounting for 58% of the treatment effect on paranoia (table 3). Change in sleep over 3 weeks explained 21% of the intervention effect on hallucinations at 10 weeks, with change in sleep over 10 weeks accounting for 39% of the intervention effect on hallucinations. Hence early changes in sleep explain approximately half of the total sleep-mediated changes in psychotic experiences by the end of treatment. In comparison, parallel analyses in the opposite direction indicated that changes in psychotic experiences explained a much smaller percentage of variation in improvements in sleep. Specifically, when paranoia and hallucination outcomes at 3 weeks were set as the mediators and the sleep outcome at 10 weeks as the main outcome, paranoia symptoms mediated just 3·8% of change in sleep and hallucinations mediated 3·4% of change in sleep. These outcomes lend further support to the causal pathway hypothesis proposed in this study.Table 3 Mediation analysis* results

ks were set as the mediators and the sleep outcome at 10 weeks as the main outcome, paranoia symptoms mediated just 3·8% of change in sleep and hallucinations mediated 3·4% of change in sleep. These outcomes lend further support to the causal pathway hypothesis proposed in this study.Table 3 Mediation analysis* results Total effect Direct effect Indirect effect Percentage mediated Effect size (95% CI) p value Effect size (95% CI) p value Effect size (95% CI) p value Paranoia (GPTS) outcome at week 10 Insomnia at week 3 (SCI-8) −2·27 (−3·03 to −1·51) <0·0001 −1·85 (−2·66 to 1·04) <0·0001 −0·67 (−0·86 to −0·48) <0·0001 29·5% Insomnia at week 10 (SCI-8) −2·27 (−3·03 to −1·51) <0·0001 −0·97 (−1·80 to −0·14) <0·0001 −1·31 (−1·60 to −1·02) <0·0001 57·8% Hallucinations (SPEQ) outcome at week 10 Insomnia at week 3 (SCI-8) −1·60 (−2·00 to −1·20) <0·0001 −1·36 (−1·79 to −0·94) <0·0001 −0·33 (−0·43 to −0·23) <0·0001 20·7% Insomnia at week 10 (SCI-8) −1·60 (−2·00 to −1·20) <0·0001 −0·90 (−1·34 to −0·46) <0·0001 −0·62 (−0·78 to −0·46) <0·0001 38·6% Total n=1718. GPTS=Green et al Paranoid Thought Scales. SCI-8=Sleep Condition Indicator 8-item version. SPEQ=Specific Psychotic Experiences Questionnaire. * Outcome and mediators modelled by means of linear mixed effects models and the total, direct, and indirect effects determined using the Baron and Kenny31 approach. The effect size is the adjusted treatment difference (ie, non-standardised treatment difference).

Total effect Direct effect Indirect effect Percentage mediated Effect size (95% CI) p value Effect size (95% CI) p value Effect size (95% CI) p value Paranoia (GPTS) outcome at week 10 Insomnia at week 3 (SCI-8) −2·27 (−3·03 to −1·51) <0·0001 −1·85 (−2·66 to 1·04) <0·0001 −0·67 (−0·86 to −0·48) <0·0001 29·5% Insomnia at week 10 (SCI-8) −2·27 (−3·03 to −1·51) <0·0001 −0·97 (−1·80 to −0·14) <0·0001 −1·31 (−1·60 to −1·02) <0·0001 57·8% Hallucinations (SPEQ) outcome at week 10 Insomnia at week 3 (SCI-8) −1·60 (−2·00 to −1·20) <0·0001 −1·36 (−1·79 to −0·94) <0·0001 −0·33 (−0·43 to −0·23) <0·0001 20·7% Insomnia at week 10 (SCI-8) −1·60 (−2·00 to −1·20) <0·0001 −0·90 (−1·34 to −0·46) <0·0001 −0·62 (−0·78 to −0·46) <0·0001 38·6% Total n=1718. GPTS=Green et al Paranoid Thought Scales. SCI-8=Sleep Condition Indicator 8-item version. SPEQ=Specific Psychotic Experiences Questionnaire. * Outcome and mediators modelled by means of linear mixed effects models and the total, direct, and indirect effects determined using the Baron and Kenny31 approach. The effect size is the adjusted treatment difference (ie, non-standardised treatment difference). The large improvement in insomnia is confirmed with the ISI assessment (table 4). The sleep treatment also led to improvements in depression, and improvements in anxiety, prodromal symptoms, nightmares, psychological wellbeing, and functioning, and all these improvements were maintained over time (Table 4, Table 5). Those participants randomised to the sleep treatment were also less likely to meet criteria over the course of the trial for a depressive episode, anxiety disorder, or ultra-high risk of psychosis (table 5). However, contact with mental health services did not differ between groups (table 5). Furthermore, the sleep treatment led to a small, sustained increase in symptoms of mania (table 5). With the sleep treatment, a greater risk also existed of meeting criteria for a manic episode (table 5). No adverse events were reported to the trial team.Table 4 Secondary outcome results

iffer between groups (table 5). Furthermore, the sleep treatment led to a small, sustained increase in symptoms of mania (table 5). With the sleep treatment, a greater risk also existed of meeting criteria for a manic episode (table 5). No adverse events were reported to the trial team.Table 4 Secondary outcome results Unadjusted mean Adjusted difference*(95% CI), d† p value Control‡ Treatment§ Insomnia (ISI) Week 10 12·95 (5·27) 9·23 (5·18) −3·72 (−4·16 to −3·29), 0·94 <0·0001 Week 22‡ 12·17 (5·29) 8·62 (5·48) −3·40 (−3·87 to −2·93), 0·86 <0·0001 Nightmares (DDNSI) Week 10 7·35 (7·85) 5·47 (6·91) −1·63 (−2·16 to −1·10), 0·20 <0·0001 Week 22‡§ 7·32 (7·93) 5·09 (6·66) −1·84 (−2·41 to −1·26), 0·23 <0·0001 Prodromal psychosis (PQ-16) Week 10 4·35 (3·71) 3·37 (3·29) −0·81 (−1·03 to −0·60), 0·24 <0·0001 Week 22 4·05 (3·83) 3·14 (3·24) −0·74 (−0·98 to −0·51), 0·22 <0·0001 Depression (PHQ-9) Week 10 11·27 (6·72) 8·44 (6·16) −2·83 (−3·30 to −2·35), 0·48 <0·0001 Week 22 10·34 (6·79) 8·00 (6·54) −2·44 (−2·95 to −1·94), 0·42 <0·0001 Anxiety (GAD-7) Week 10 8·35 (6·06) 6·53 (5·40) −1·86 (−2·29 to −1·43), 0·33 <0·0001 Week 22 7·67 (6·10) 6·14 (5·41) −1·56 (−2·01 to −1·10), 0·28 <0·0001 Mania (Altman) Week 10 2·97 (3·03) 3·77 (3·33) 0·93 (0·67 to 1·19), −0·31 <0·0001 Week 22 2·92 (3·06) 3·57 (3·41) 0·75 (0·46 to 1·03), −0·25 <0·0001 Functioning (WSAS) Week 10 15·92 (8·89) 11·43 (8·37) −4·36 (−5·03 to −3·69), 0·58 <0·0001 Week 22 14·92 (9·17) 10·25 (8·30) −4·33 (−5·05 to −3·62), 0·57 <0·0001 Wellbeing (WEMWBS) Week 10 38·73 (9·78) 40·92 (9·63) 2·47 (1·72 to 3·22), 0·29 <0·0001 Week 22 39·63 (10·19) 42·12 (10·36) 2·78 (1·97 to 3·60), 0·32 <0·0001 Data are mean (SD). ISI=Insomnia Severity Index. DDNSI=Disturbing Dream and Nightmare Severity Index. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. WSAS=Work and Social Adjustment Scale. WEMWBS=Warwick–Edinburgh Mental Wellbeing Scale.

. ISI=Insomnia Severity Index. DDNSI=Disturbing Dream and Nightmare Severity Index. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. WSAS=Work and Social Adjustment Scale. WEMWBS=Warwick–Edinburgh Mental Wellbeing Scale. * Linear mixed effects model adjusted for gender, student status, week, and interaction of week with randomisation, and including a random effect for student within university. Covariance matrix of within subject measurements was unstructured. † d is standardised effect size (Cohen's d). ‡ 1142 participants in the control group at week 10 and 971 participants at week 22, except for 970 participants for insomnia (ISI) and 963 participants for nightmares (DDNSI) at week 22. § 733 participants in the treatment group at week 10 and 603 participants at week 22, except for 599 participants for nightmares (DDNSI) at week 22. Table 5 Secondary dichotomous outcome results

‡ 1142 participants in the control group at week 10 and 971 participants at week 22, except for 970 participants for insomnia (ISI) and 963 participants for nightmares (DDNSI) at week 22. § 733 participants in the treatment group at week 10 and 603 participants at week 22, except for 599 participants for nightmares (DDNSI) at week 22. Table 5 Secondary dichotomous outcome results Adjusted odds ratio*(95% CI) p value Ultra-high risk of psychosis (PQ-16) Week 10 0·26 (0·15 to 0·46) <0·0001 Week 22 0·33 (0·18 to 0·59) 0·00026 Mania (Altman) Week 10 2·01 (1·48 to 2·73) <0·0001 Week 22 1·89 (1·34 to 2·66) 0·00027 Depressive disorder (PHQ-9) Week 10 0·21 (0·14 to 0·32) <0·0001 Week 22 0·32 (0·21 to 0·48) <0·0001 Anxiety disorder (GAD-7) Week 10 0·32 (0·21 to 0·48) <0·0001 Week 22 0·42 (0·27 to 0·64) <0·0001 Contacted mental health services Week 10 1·19 (0·70 to 2·04) 0·52 Week 22 0·98 (0·55 to 1·75) 0·94 Mental health diagnosis Week 10 1·33 (0·75 to 2·37) 0·33 Week 22 1·43 (0·78 to 2·63) 0·25 Psychiatric medication Week 10 0·77 (0·47 to 1·26) 0·30 Week 22 0·96 (0·58 to 1·59) 0·86 Psychological therapy Week 10 1·27 (0·48 to 3·35) 0·63 Week 22 0·41 (0·11 to 1·58) 0·20 1142 participants in the control group at week 10 and 971 participants at week 22. 733 participants in the treatment group at week 10 and 603 participants at week 22. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version.

0·20 1142 participants in the control group at week 10 and 971 participants at week 22. 733 participants in the treatment group at week 10 and 603 participants at week 22. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. * Logistic mixed effects model adjusted for gender, student status, week, and interaction of week with randomisation, and including a random effect for student within university. Covariance matrix of within subject measurements was unstructured.

0·20 1142 participants in the control group at week 10 and 971 participants at week 22. 733 participants in the treatment group at week 10 and 603 participants at week 22. PQ-16=Prodromal Questionnaire 16-item version. PHQ-9=Patient Health Questionnaire 9-item version. GAD-7=Generalised Anxiety Disorder 7-item version. * Logistic mixed effects model adjusted for gender, student status, week, and interaction of week with randomisation, and including a random effect for student within university. Covariance matrix of within subject measurements was unstructured. Discussion We aimed to investigate the effects on mental health of the reduction of sleep difficulties. The first necessary stage was for the intervention to reduce insomnia, which was achieved. A large effect size reduction was found with the digital CBT intervention in a large student population. But we designed the trial to establish the consequent effects on psychotic experiences. To our knowledge, the OASIS trial is the largest randomised controlled trial of a psychological intervention for a mental health problem. Students randomly assigned to the sleep intervention showed small, sustained reductions in paranoia and hallucinations, suggesting that disrupted sleep has a contributory causal role in the occurrence of these psychotic experiences in a specific population of young adults. The mediation analyses supported this interpretation—eg, improvements in sleep accounted for almost 60% of the change in paranoia after treatment. Insomnia might not be the largest cause of psychotic experiences but it is not an epiphenomenon. Hence, this study adds to our understanding of the causes of psychotic experiences and might indicate a promising route into the early treatment of some psychotic problems.

ost 60% of the change in paranoia after treatment. Insomnia might not be the largest cause of psychotic experiences but it is not an epiphenomenon. Hence, this study adds to our understanding of the causes of psychotic experiences and might indicate a promising route into the early treatment of some psychotic problems. The focus on a sleep intervention in a young adult population is important. Young people with incipient disorders might be very reluctant to seek help for psychiatric problems. Trouble sleeping is a common complaint with little stigma. Hence, it provides a much more acceptable focus for a first step in a care pathway. The digital sleep treatment gave added benefits. Depression in particular, but also anxiety, psychological wellbeing, nightmares, and perceived functioning all improved. The effects on anxiety and depression are consistent with the results of a meta-analysis.33 Participants who received the sleep treatment in the trial were less likely to report symptoms at a level that met criteria for ultra-high risk of psychosis, depression, or anxiety disorder. At baseline, the frequency of positive screens for psychosis risk with the Prodromal Questionnaire was high, similar to the risk found with this questionnaire for adolescents referred to treatment services;34 this high rate will reflect the well established associations of sleep difficulties with psychotic experiences2, 3, 7, 8 (ie, that participants have been selected for insomnia and therefore will score higher on psychosis measures), and also the limitations of brief self-report questionnaires for assessing psychosis risk. However, in the trial, sleep treatment did not affect contact with mental health services. Most participants were not in contact with these services so a longer follow-up period might be needed to truly test such effects. Furthermore, manic symptoms associated with the sleep treatment increased. This outcome might be due to an actual increase in problematic manic symptoms or it might simply reflect the overall increase in psychological wellbeing in the sample since the questionnaire domains concern cheerfulness, self-confidence, reduced need for sleep, increase in amount of activity, and talkativeness. The Altman scale has been found to correlate poorly with self-ratings of elation.35

s or it might simply reflect the overall increase in psychological wellbeing in the sample since the questionnaire domains concern cheerfulness, self-confidence, reduced need for sleep, increase in amount of activity, and talkativeness. The Altman scale has been found to correlate poorly with self-ratings of elation.35 Are the study results generalisable beyond a student population? We consider that the results are likely to apply to the wider adult population. We used a treatment developed for adults, which was not modified for students. The large treatment reduction in insomnia for the students is very similar to that found in trials with general adult populations,10, 11, 36 while previous studies36, 37 with community samples have shown self-help sleep treatment to reduce anxiety and depression. Nonetheless, only a direct comparison in a trial can definitively determine the generalisablilty of our findings. Although not the primary objective of the study, the trial does indicate that the provision of internet-delivered CBT for insomnia to university students is likely to lead to reductions overall in insomnia, and smaller reductions in a number of other mental health symptoms, with benefits for positive psychological wellbeing too. Tailoring of the intervention specifically for this population could enhance engagement and outcome effects. Support to complete the intervention might well be helpful too.

ns overall in insomnia, and smaller reductions in a number of other mental health symptoms, with benefits for positive psychological wellbeing too. Tailoring of the intervention specifically for this population could enhance engagement and outcome effects. Support to complete the intervention might well be helpful too. Several limitations exist in the study. First, the study relied on self-report questionnaires, albeit validated in their development against clinical interviews. Similar change captured in rater-assessed measures would have strengthened confidence in the study results. Second, the samples tested were predominately in the non-clinical range of psychotic experiences, restricting the conclusions to the less severe end of the psychosis spectrum. Third, the participants were self-selecting in responding to the invitation, which will have affected the representativeness of the sample. However, access to the study was via an Internet webpage, which is a simpler process than obtaining treatment from clinical services. The whole study could be completed in the privacy of the home, which means that far fewer barriers existed to participation than conventional patient trials. Fourth, the extent to which the results will generalise to the rest of the population is not known. Even within the student population, we do not know the representativeness of the participants. Fifth, bias in the outcome results will have been introduced because of the high dropout rate, especially in the treatment group, which is similar to comparable online studies.12 The results did remain robust against conservative assumptions in the sensitivity analyses about those participants who dropped out, but it is notable that treatment effects were greater for those participants who completed the sleep treatment. Finally, the causal argument rests on the plausible assumption that the sleep treatment first changes the occurrence of insomnia, since that was the topic of the intervention, but the mediation analyses in this trial based on week 10 outcomes cannot fully capture the temporal order of changes or rule out reverse causation. We were able to show a significant amount of mediation based on the week 3 insomnia score as a mediator, while evidence for reverse causation was weak, which does follow the predicted temporal causal pathway.

ial based on week 10 outcomes cannot fully capture the temporal order of changes or rule out reverse causation. We were able to show a significant amount of mediation based on the week 3 insomnia score as a mediator, while evidence for reverse causation was weak, which does follow the predicted temporal causal pathway. In reality, it is difficult in a clinical trial to capture potential temporal changes between mediator and outcomes, since improvements in paranoia and hallucinations are likely to closely parallel the improvement in sleep.

ial based on week 10 outcomes cannot fully capture the temporal order of changes or rule out reverse causation. We were able to show a significant amount of mediation based on the week 3 insomnia score as a mediator, while evidence for reverse causation was weak, which does follow the predicted temporal causal pathway. In reality, it is difficult in a clinical trial to capture potential temporal changes between mediator and outcomes, since improvements in paranoia and hallucinations are likely to closely parallel the improvement in sleep. This work can be taken forward in several possible ways. Determination of the mechanisms linking insomnia to psychotic experiences will shed further light on the causes of psychosis and potentially enable treatment improvement.3, 13 Of great clinical interest will be the evaluation of the effects of improving sleep for patients attending clinical services with ultra-high risk of psychosis, or established clinical psychotic experiences, or at the early stages of relapse. Our experience is that patients with psychosis value their sleep difficulties being appropriately addressed, that this enhances engagement with other treatments, and that better sleep can contribute to a reduction in psychotic experiences. Furthermore, a gap exists in mental health services regarding intervention for early, relatively non-specific presentations, and proper sleep treatment might provide a sensible first response. Overall, this trial indicates the importance of sleep difficulties for mental health in the general population and the need for a reconsideration in clinical services of the priority given to improving sleep.

rly, relatively non-specific presentations, and proper sleep treatment might provide a sensible first response. Overall, this trial indicates the importance of sleep difficulties for mental health in the general population and the need for a reconsideration in clinical services of the priority given to improving sleep. Supplementary Material Supplementary appendix 1 Supplementary appendix 2 Acknowledgments The study was done by the University of Oxford Sleep and Circadian Neuroscience Institute, with a grant from the Wellcome Trust (098461/Z/12/Z). The sleep treatment programme was provided to all the trial participants at no cost by Sleepio/Big Health Ltd. RGF and GMG are the principal investigators for the grant, and DFr and PJH are grant holders. BS is directly funded by the grant. The research in Nottingham (CHo, EBD, CG) was supported by the National Institute of Health Research (NIHR) MindTech Healthcare Technology Co-operative. DFr is supported by an NIHR Research Professorship. This study was funded by the Wellcome Trust and supported by the NIHR Oxford Health Biomedical Research Centre. The views expressed are those of the authors and not necessarily those of the National Health Service, NIHR, or Department of Health.

chnology Co-operative. DFr is supported by an NIHR Research Professorship. This study was funded by the Wellcome Trust and supported by the NIHR Oxford Health Biomedical Research Centre. The views expressed are those of the authors and not necessarily those of the National Health Service, NIHR, or Department of Health. Contributors DFr was the chief investigator, conceived the study, led the design, and drafted the paper. BS was the trial coordinator and contributed to the design. GMG conceived the study and contributed to the design. L-MY contributed to the design and was responsible for the main statistical outcome analyses. AN carried out the trial analyses. PJH conceived the study and contributed to the design. RE oversaw the statistical mediation analysis. AIL was responsible for the digital therapy programme. VW was responsible for the computer programming that carried out the screening, assessments, and links to the digital intervention. RGF contributed to the design of the trial. CHi was responsible for the development and coordination of the underlying True Colours platform. CAE conceived the study, contributed to the design, and was responsible for the digital therapy programme. Other authors led the research at their university site. All authors commented upon and approved the final manuscript.

was responsible for the development and coordination of the underlying True Colours platform. CAE conceived the study, contributed to the design, and was responsible for the digital therapy programme. Other authors led the research at their university site. All authors commented upon and approved the final manuscript. Declaration of interests DFr reports grants from Wellcome Trust, non-financial support (provision of the sleep treatment for the OASIS trial) from Sleepio, and grants from the UK National Institute of Health Research (NIHR), during the conduct of the study, and personal fees from Oxford Virtual Reality, a University of Oxford spin-out company, outside the submitted work. BS reports grants from Wellcome Trust, during the conduct of the study, and personal fees from Big Health Ltd, outside the submitted work. GMG reports personal fees from Servier, Lundbeck and Otsuka, Takeda, Merck, Medscape, Pfizer, P1vital, Compass Pathways, Shire, Angelini Spa, and Allergan, outside the submitted work. PJH reports grants from the Wellcome Trust and Medical Research Council, during the conduct of the study. The position of AIL at the University of Oxford is funded by Big Health Ltd. AIL reports non-financial support and other from Big Health Ltd, during the conduct of the study, and personal fees, non-financial support, and other from Big Health Ltd, outside the submitted work. CHi reports grants from Roche and Eli Lilly, outside the submitted work. AG reports grants from NIHR HTA Grant Number 13/15/04, outside the submitted work. PK reports research grant funding from NIHR and the Economic and Social Research Council. SJ reports research grant funding from NIHR outside the submitted work. PK is Vice President of the British Psychological Society. AMG reports grants from NIHR Maudsley Biomedical Research Centre, outside the submitted work, and is writing a book on sleep to be published by Bloomsbury Sigma. AMG has provided guidance and educational content for the Pediatric Sleep Council website; a freely available educational website providing tips to help with sleep in babies and young children. The website is partially supported by Johnson and Johnson, but they do not have any influence over content and do not advertise on it. AMG contributes to the British Broadcasting Corporation Focus magazine.

l website; a freely available educational website providing tips to help with sleep in babies and young children. The website is partially supported by Johnson and Johnson, but they do not have any influence over content and do not advertise on it. AMG contributes to the British Broadcasting Corporation Focus magazine. CAE is a co-founder and Chief Medical Officer of the digital cognitive behavioural therapy for an insomnia programme (Sleepio/Big Health Ltd) and reports other from Big Health Ltd, during the conduct of the study, and other from Big Health Ltd and personal fees from Warnford Wellness, outside the submitted work. CAE has a US patent issued: serial number 14/172,347, Interactive System for Sleep Improvement; docket number HAME-001. All other authors declare no competing interests.

Introduction Mental health disorders are very common and encompass great personal and societal costs, but far too few people receive the best treatments. For example, in the UK, one adult in six meets criteria for a common mental health disorder but most of these individuals do not receive treatment.1 People who do receive treatment are more likely to be given psychotropic medication than a psychological intervention. Yet, for many common mental health conditions, particularly anxiety disorders, evidence-based psychological treatments are both the best treatment option2 and the preference of patients.3 However, increasing the provision of psychological treatment is difficult. Therapists need to be trained and then adhere competently to evidence-based treatment protocols. Moreover, the best psychological treatments are not simply so-called talking therapies but take the form of direct active learning and coaching in real-world situations, whereas most therapists typically have little time for sessions outside the clinic. This situation results in high service pressures and unmet needs of patients.

he best psychological treatments are not simply so-called talking therapies but take the form of direct active learning and coaching in real-world situations, whereas most therapists typically have little time for sessions outside the clinic. This situation results in high service pressures and unmet needs of patients. Immersive virtual reality (VR) has the potential to substantially increase access to the best psychological interventions. First, treatments can be automated and provided in VR, so a therapist does not need to be present. Automated treatment delivered using VR consumer hardware could become a low-cost way of providing effective interventions at scale. Second, VR can deliver the most powerful element of direct therapeutic intervention—ie, direct coaching in everyday situations that trouble patients. This element is all too frequently missing in clinics. By putting on a headset, patients can be taken immediately into various situations, graded in difficulty, that they find cause psychological distress. Third, in VR, patients are willing to go into situations that trouble them and try alternative ways of responding, because they know it is a simulation. However, the learning transfers to the real world.4 Therapeutic intervention is therefore faster. Moreover, the engaging and entertaining nature of VR could result in higher treatment uptake than with conventional treatment. VR has been used successfully over the past 25 years for assessment, understanding, and treatment of mental health disorders.5 The increased accessibility and affordability of VR mean that this technique is now ready to move from specialist laboratories into clinics.

r treatment uptake than with conventional treatment. VR has been used successfully over the past 25 years for assessment, understanding, and treatment of mental health disorders.5 The increased accessibility and affordability of VR mean that this technique is now ready to move from specialist laboratories into clinics. Research in context Evidence before this study

r treatment uptake than with conventional treatment. VR has been used successfully over the past 25 years for assessment, understanding, and treatment of mental health disorders.5 The increased accessibility and affordability of VR mean that this technique is now ready to move from specialist laboratories into clinics. Research in context Evidence before this study At the beginning of 2017, we searched PubMed with no restrictions on date for all published work in the English language on use of virtual reality (VR) to assess, understand, and treat mental health problems. The general search terms were (“virtual reality” OR “immersive virtual reality”) AND (“assessment”, “treatment”, “research”, “study”, “experiment”, OR “understanding”) AND ([disorder-specific terms]). Our search retrieved 285 empirical papers. Over the past 25 years, VR has been used to aid therapist-delivered psychological interventions, principally exposure therapy for anxiety disorders. Meta-analyses indicate that treatment effect sizes using VR for anxiety disorders, including fear of heights, are large (Cohen's d=1·1). However, VR-assisted therapy has only been delivered in a few places, because of the specialist equipment that is needed and the reliance on a skilled therapist to provide the psychological expertise. We found no reports of automated VR being used to treat mental health problems. We updated our search of PubMed on May 14, 2018, with no restrictions by language, with the terms (“VR” OR “virtual reality”) AND (“treatment”) AND (“mental health”) AND (“automated”, “self-guided”, OR “automatic”). We did not retrieve any further research.

ts of automated VR being used to treat mental health problems. We updated our search of PubMed on May 14, 2018, with no restrictions by language, with the terms (“VR” OR “virtual reality”) AND (“treatment”) AND (“mental health”) AND (“automated”, “self-guided”, OR “automatic”). We did not retrieve any further research. Added value of this study Our study comprised 100 individuals who self-referred for therapy. We showed that psychological treatment for fear of heights can be automated using immersive VR. Automation was delivered via a virtual coach, rather than a therapist, which guided participants through a series of graded exercises beside virtual heights that could facilitate cognitive change. Participants interacted with the virtual coach to tailor the treatment. The clinical effectiveness in our study was at least as good as face-to-face therapy. Implications of all the available evidence Automated treatment delivered using VR consumer hardware could become a low-cost way of providing effective interventions at scale. Clinical testing is needed to ascertain whether automation of psychological treatment works for other mental disorders.

Our study comprised 100 individuals who self-referred for therapy. We showed that psychological treatment for fear of heights can be automated using immersive VR. Automation was delivered via a virtual coach, rather than a therapist, which guided participants through a series of graded exercises beside virtual heights that could facilitate cognitive change. Participants interacted with the virtual coach to tailor the treatment. The clinical effectiveness in our study was at least as good as face-to-face therapy. Implications of all the available evidence Automated treatment delivered using VR consumer hardware could become a low-cost way of providing effective interventions at scale. Clinical testing is needed to ascertain whether automation of psychological treatment works for other mental disorders. Previous research using VR for treatment of mental health disorders has relied on having a therapist present, typically to deliver exposure therapy. Effect sizes for VR-assisted therapy are usually large (Cohen's d=1·1).6 The first application of VR to mental health difficulties was for treating fear of heights,7 which is the most common phobia. One in five people report having had a strong unreasonable fear of heights during their lifetime, and one in 20 people reach diagnostic criteria for acrophobia.8 One of the largest trials for this disorder randomly allocated 33 people with a fear of heights to three 1-h sessions of either exposure in vivo or VR exposure—in both cases delivered by a therapist.9 Both treatments were equally effective and the benefits remained 6 months later. The therapeutic effects of VR exposure for fear of heights persist for at least a year.10

y allocated 33 people with a fear of heights to three 1-h sessions of either exposure in vivo or VR exposure—in both cases delivered by a therapist.9 Both treatments were equally effective and the benefits remained 6 months later. The therapeutic effects of VR exposure for fear of heights persist for at least a year.10 We aimed to automate the provision of a psychological intervention for fear of heights by programming a virtual coach to act as the therapist in VR. Our therapeutic approach is cognitive, focusing on the evaluation of threat predictions while dropping defensive behaviours, to develop memories of safety that counteract fear associations. The treatment is delivered using inexpensive consumer VR equipment. We postulated that, compared with usual care (in effect, no treatment), a large treatment benefit (reduction in fear of heights) would be seen with the VR intervention immediately after treatment (primary hypothesis) and that this benefit would persist after treatment had been completed (secondary hypothesis). Methods Study design and participants We did a single-blind, parallel group, randomised controlled, superiority trial comparing automated VR with usual care. The control condition was usual care so that we did not interfere with any other type of help that the individual might have been receiving. In effect, this was equivalent to receiving no treatment for fear of heights. At the end of the trial, control participants were offered the opportunity to have the VR treatment.

The control condition was usual care so that we did not interfere with any other type of help that the individual might have been receiving. In effect, this was equivalent to receiving no treatment for fear of heights. At the end of the trial, control participants were offered the opportunity to have the VR treatment. We recruited people with a fear of heights by advertisements on local radio (Oxfordshire, UK). Inclusion criteria were that a participant had to be older than 18 years and have a fear of heights corresponding to a Heights Interpretation Questionnaire (HIQ)11 score of more than 29. We used the same cutoff as was used in the treatment trial by Arroll and colleagues,12 which indicates at least a moderate fear of heights. We excluded people if they were currently receiving a psychological treatment for fear of heights, had photosensitive epilepsy, or had either no stereoscopic vision or balance problems. The study received ethics approval from the University of Oxford Medical Sciences Inter-Divisional ethics committee. All participants gave written informed consent for trial participation. No changes to methods were made after the trial began.

We recruited people with a fear of heights by advertisements on local radio (Oxfordshire, UK). Inclusion criteria were that a participant had to be older than 18 years and have a fear of heights corresponding to a Heights Interpretation Questionnaire (HIQ)11 score of more than 29. We used the same cutoff as was used in the treatment trial by Arroll and colleagues,12 which indicates at least a moderate fear of heights. We excluded people if they were currently receiving a psychological treatment for fear of heights, had photosensitive epilepsy, or had either no stereoscopic vision or balance problems. The study received ethics approval from the University of Oxford Medical Sciences Inter-Divisional ethics committee. All participants gave written informed consent for trial participation. No changes to methods were made after the trial began. Randomisation and masking Randomisation was done by an automated online system developed by the University of Oxford Primary Care Clinical Trials Unit. Participants were randomly allocated in a 1:1 ratio, stratified by fear of height severity (moderate [HIQ score 30–55] or severe [HIQ score 56–80]) and using randomised permuted block sizes of four and eight. The research assistant who gathered follow-up data was unaware of the random allocation, and masking was not broken.

cipants were randomly allocated in a 1:1 ratio, stratified by fear of height severity (moderate [HIQ score 30–55] or severe [HIQ score 56–80]) and using randomised permuted block sizes of four and eight. The research assistant who gathered follow-up data was unaware of the random allocation, and masking was not broken. Procedures After a potential participant contacted the study team, we asked them to complete the HIQ online. If they scored above the threshold (HIQ score >29), we checked inclusion and exclusion criteria over the telephone. Participants needed to score above the HIQ threshold again when they attended for the baseline assessment (and were excluded before randomisation if they did not). All assessments were done at the study centre (Oxford VR, Oxford, UK). At baseline, we obtained basic demographic data and assessed participants with the DSM-513 to see if they met criteria for acrophobia. We also asked participants if they were receiving any treatment for fear of heights.

ndomisation if they did not). All assessments were done at the study centre (Oxford VR, Oxford, UK). At baseline, we obtained basic demographic data and assessed participants with the DSM-513 to see if they met criteria for acrophobia. We also asked participants if they were receiving any treatment for fear of heights. The VR intervention consisted of a software application called Now I Can Do Heights, which was intended by Oxford VR to help users with acrophobia overcome their fear of heights. The application is intended for use by adults older than 18 years and was designed to be used without a supporting therapist, although it can be used alongside a therapist or in a clinical setting if needed. The application is a CE-marked class I active medical device (device code Z301 [standalone software]), in conformity with the essential requirements and provisions of EC directive 93/42/EEC (medical devices). The software was developed using Unity3D (version 5.6.0f3 [64-bit]; Unity Technologies, San Francisco, CA, USA) and delivered using a gaming personal computer (Chillblast Fusion Strix Gaming PC, Intel Core i7-7700K processor, 16 GB DDR4 3000 MHz memory, ASUS GeForce GTX 1080 8GB graphics card, 500 GB M.2 solid state drive/3 TB hard disc drive; Chillblast, Poole, UK) and the HTC Vive (HTC Corporation, New Taipei City, Taiwan)—a consumer VR head-mounted display that has associated hand controllers and headset tracking. Participants also wore headphones with a microphone (Creative Labs Sound Blaster Tactic3d Rage Wireless; Creative Labs, Jurong East, Singapore).

hillblast, Poole, UK) and the HTC Vive (HTC Corporation, New Taipei City, Taiwan)—a consumer VR head-mounted display that has associated hand controllers and headset tracking. Participants also wore headphones with a microphone (Creative Labs Sound Blaster Tactic3d Rage Wireless; Creative Labs, Jurong East, Singapore). The treatment was designed to be administered in roughly six 30-min VR sessions over a period of 2 weeks. We allowed a degree of flexibility in treatment length because individuals differ in the speed with which they progress. Participants underwent the treatment in the office of Oxford VR, with a graduate psychologist in the room to help the participant put on the headset and for basic safety reasons. Programme sessions ran automatically. At the beginning of every session, the participant's virtual body was calibrated by the software system reading data from head and hand VR trackers. Participants who were assigned the VR intervention undertook the treatment while standing up, and they could walk in the virtual environment.

e sessions ran automatically. At the beginning of every session, the participant's virtual body was calibrated by the software system reading data from head and hand VR trackers. Participants who were assigned the VR intervention undertook the treatment while standing up, and they could walk in the virtual environment. At the beginning of the first session, the participant had an initial assessment with the virtual coach in a virtual office. The virtual coach was animated using motion capture from an actor and voiced by the same person. The coach gave the participant basic information about fear of heights and its treatment from a cognitive perspective (eg, “The reason we're afraid of heights is because we think something bad is going to happen. And that makes us feel anxious. Then we end up avoiding heights because they feel so scary. But I'll show you how to look at those thoughts in a new way.”), then asked the participant a series of questions about their key fear about heights (eg, fear of falling, fear of the building collapsing, fear of throwing oneself off) and obtained ratings of belief conviction (rated on a scale from 0 [“I don't believe it will happen”] to 10 [“I'm absolutely certain it will happen”]). The basic mechanism of treatment was for individuals to find out how accurate their fears were. Hence, throughout the programme, participants were encouraged by the virtual coach to find out how safe they were and to put their expectations to the test (eg, “Remember: we're exploring here. We're testing out our expectations. We're discovering what happens when we venture into a situation we'd normally try to avoid.”). The virtual coach also explained how this learning depended on dropping safety-seeking behaviours (eg, “Many people try to deal with their fear of heights by using defences. They put up barriers between themselves and what they fear. The most obvious one is simply avoiding heights. But there are lots of subtler defences: closing your eyes when you're up high or not looking down, repeating a comforting phrase to yourself, taking off your shoes, holding on to something. We need to lower these defences. They can make us feel better in the short term. But they prevent us truly engaging with the situations that make us anxious—and stop us learning just how much we can achieve without them.”).

down, repeating a comforting phrase to yourself, taking off your shoes, holding on to something. We need to lower these defences. They can make us feel better in the short term. But they prevent us truly engaging with the situations that make us anxious—and stop us learning just how much we can achieve without them.”). The treatment was not designed as exposure therapy (ie, participants were not asked to remain in situations until anxiety reduced) but as repeated behavioural experiment tests (ie, to learn that they were safer than they had thought). Throughout the sessions, participants' responses to questions from the virtual coach were given either by means of voice recognition technology (via a microphone attached to the headphones) or by using a virtual watch. Belief ratings were repeated within VR at the end of every treatment session.

than they had thought). Throughout the sessions, participants' responses to questions from the virtual coach were given either by means of voice recognition technology (via a microphone attached to the headphones) or by using a virtual watch. Belief ratings were repeated within VR at the end of every treatment session. After the initial assessment stage (during the first session), the treatment began while still within the first session. The virtual coach took the participant to the atrium of a large ten-storey office complex. The environment featured many height cues (eg, balls in the air, people moving about). The participant then chose on which of the first five floors to begin the activities (ie, participants could not access higher than the fifth floor until later sessions). Guided throughout by the virtual coach, tasks were undertaken on every floor to enable the participant to find out whether his or her fears were accurate. Tasks were randomised across all the remaining floors for every participant, but were weighted towards the easiest tasks first (eg, a barrier gradually lowering) whereas harder tasks came later (eg, going out on a platform from the floor into the atrium space). Many tasks were designed to be engaging for participants (eg, rescuing a cat from a tree, playing a xylophone near the edge of the floor, throwing balls over the edge of the floor). The virtual coach described the tasks to the participant, provided empathic encouragement, repeated key learning points, and sought feedback on whether the participant felt safer than before. The participant could decide to repeat tasks or move up to the next floor. At the end of the session, the participant was brought back down to the atrium ground level and the coach would ask for belief ratings for the height threat and encourage the participant to try real heights between sessions. The sessions were saved so that the participant could begin the next session where they had left off. Pictures from the VR treatment can be seen in the appendix.

o the atrium ground level and the coach would ask for belief ratings for the height threat and encourage the participant to try real heights between sessions. The sessions were saved so that the participant could begin the next session where they had left off. Pictures from the VR treatment can be seen in the appendix. Outcomes The primary outcome measure was the HIQ,11 which is a 16-item self-report questionnaire. The scale strongly predicts distress, anxiety, and avoidance of real heights and has high internal consistency and convergent validity with other fear of heights measures.11 We asked participants to rate anxious fears (eg, that they will fall, faint, or hurt themselves) when imagining two different height situations (ie, being on a ladder against a two-storey house and on the balcony of a 15th floor building). Scores can range from 16 to 80, with higher scores indicating a greater severity of fear of heights. Internal reliability of the scale at baseline was very high (Cronbach's α=0·91).

imagining two different height situations (ie, being on a ladder against a two-storey house and on the balcony of a 15th floor building). Scores can range from 16 to 80, with higher scores indicating a greater severity of fear of heights. Internal reliability of the scale at baseline was very high (Cronbach's α=0·91). Secondary outcome measures were the acrophobia questionnaire (AQ)14 and the Improving Access to Psychological Therapies (IAPT) phobia scale–avoidance.15 The AQ is a 40-item self-report questionnaire. For 20 different height situations (eg, diving off the low board at a swimming pool, riding a Ferris wheel, walking on a footbridge over a motorway), we asked participants to rate their level of anxiety (on a scale from 0 to 6) and level of avoidance (scale from 0 to 2). We measured the AQ total score and two AQ subscale scores (anxiety and avoidance). Anxiety scores can range between 0 and 120 and avoidance scores between 0 and 40, with total AQ scores ranging between 0 and 160.14 Higher scores indicate a greater severity of fear of heights. The internal reliability of the scale at baseline was very high (Cronbach's α=0·90). The IAPT phobia scale–avoidance is a single-item scale taken from the routine outcome measures administered by the UK National Health Service's IAPT programme, which treats common emotional disorders.15 We asked participants to rate their avoidance of heights on a scale from 0 (would not avoid it) to 8 (always avoid it). Higher scores indicated greater avoidance.

tem scale taken from the routine outcome measures administered by the UK National Health Service's IAPT programme, which treats common emotional disorders.15 We asked participants to rate their avoidance of heights on a scale from 0 (would not avoid it) to 8 (always avoid it). Higher scores indicated greater avoidance. We also recorded the occurrence of any known serious adverse events in participants, which we defined as death, suicide attempts, serious violent incidents, admissions to psychiatric hospital, and formal complaints about the intervention. We used the Simulator Sickness Questionnaire (SSQ)16 to assess levels of discomfort provoked by a VR session. The SSQ is the most commonly used assessment of VR simulator sickness, but it was not developed in the context of the treatment of anxiety, and the 16 items in the scale overlap completely with symptoms of anxiety—eg, increased salivation, sweating, nausea, vertigo, burping, headache, blurred vision, and feeling dizzy. We administered the SSQ four times: before and after the first and last VR treatment sessions. Participants rated the items for how much they were experienced at that time, on a 4-point scale (from 0 [none] to 3 [very strong]). We used a simple raw total score of all items,17, 18 because we wanted to see absolute levels and potential increases in discomfort. The total score, therefore, can range between 0 and 48, with higher scores indicating greater discomfort.

experienced at that time, on a 4-point scale (from 0 [none] to 3 [very strong]). We used a simple raw total score of all items,17, 18 because we wanted to see absolute levels and potential increases in discomfort. The total score, therefore, can range between 0 and 48, with higher scores indicating greater discomfort. Statistical analysis A full statistical analysis plan was agreed before the trial analysis (appendix). We did analyses with Stata version 15.1. Analyses were done by intention to treat at the end of the trial and were validated by a second statistician. Because effect sizes for VR-assisted therapy for anxiety disorders are typically large (Cohen's d=1·1),6 our target sample size was 100 individuals, which would enable the trial to detect a standardised treatment effect of 0·65 (medium) with 90% power, and of 0·57 (medium) with 80% power, at a significance level of 0·05. These treatment effects amount to differences between treatments on the HIQ of 7·6 and 6·6, respectively, based on an SD of 11·7 (taken from the mean of the baseline SDs of the two groups reported by Arroll and colleagues).12

(medium) with 90% power, and of 0·57 (medium) with 80% power, at a significance level of 0·05. These treatment effects amount to differences between treatments on the HIQ of 7·6 and 6·6, respectively, based on an SD of 11·7 (taken from the mean of the baseline SDs of the two groups reported by Arroll and colleagues).12 We analysed continuous outcomes using a linear mixed effects model, which included a random effect for participant to account for repeated measures at 2 weeks and 4 weeks. We included as fixed effects treatment group, assessment timepoint (as a categorical variable), baseline score for the outcome scale, and the interaction between treatment group and assessment timepoint (to allow estimation of a treatment effect at each of the two timepoints). In the secondary outcome analyses, we also included baseline HIQ scores as a fixed effect because stratification was based on this variable. We did not need techniques for missing data. We present results as mean differences in scores between treatment groups, with 95% CIs and associated two-sided p values. We calculated effect sizes using Cohen's d, dividing the treatment effect by the shared SD at baseline. We also ascertained the number needed to treat (NNT) to reduce the fear of heights by at least 25%, 50%, and 75%, and to below the study entry criterion. This trial is registered with the ISRCTN registry, number ISRCTN11898283.

We analysed continuous outcomes using a linear mixed effects model, which included a random effect for participant to account for repeated measures at 2 weeks and 4 weeks. We included as fixed effects treatment group, assessment timepoint (as a categorical variable), baseline score for the outcome scale, and the interaction between treatment group and assessment timepoint (to allow estimation of a treatment effect at each of the two timepoints). In the secondary outcome analyses, we also included baseline HIQ scores as a fixed effect because stratification was based on this variable. We did not need techniques for missing data. We present results as mean differences in scores between treatment groups, with 95% CIs and associated two-sided p values. We calculated effect sizes using Cohen's d, dividing the treatment effect by the shared SD at baseline. We also ascertained the number needed to treat (NNT) to reduce the fear of heights by at least 25%, 50%, and 75%, and to below the study entry criterion. This trial is registered with the ISRCTN registry, number ISRCTN11898283. Role of the funding source Oxford VR own the automated treatment and helped design the trial. The National Institute for Health Research Oxford Health Biomedical Research Centre funded the randomisation programme and statistical analysis. Some of the authors are employed by the funders and contributed to the report. The decision to submit the trial results for publication was agreed during trial registration, before the trial began. The corresponding author had full access to all data in the study and had final responsibility for the decision to submit for publication.

ors are employed by the funders and contributed to the report. The decision to submit the trial results for publication was agreed during trial registration, before the trial began. The corresponding author had full access to all data in the study and had final responsibility for the decision to submit for publication. Results Between October, 2017, and February, 2018, 189 people self-referred to the study, of whom 89 were excluded after screening and assessment (figure 1). Thus, 100 participants underwent randomisation, which began on Nov 25, 2017, and ended on Feb 27, 2018. 49 participants were allocated the VR intervention and 51 were assigned to the control group. Final data were gathered on April 6, 2018.Figure 1 Trial profile HIQ=Heights Interpretation Questionnaire. VR=virtual reality. Participants had been experiencing a fear of heights for a mean of 30 years (SD 14·47). No participant was receiving any other form of help for their fear of heights. 90 participants met diagnostic criteria for acrophobia. Treatment groups were balanced across the variables (table 1), except for gender, with a higher proportion of women assigned to the control group. Scores on the HIQ were associated positively with scores on the AQ (r=0·68; p<0·0001) and with scores on the IAPT phobia scale (r=0·54; p<0·0001).Table 1 Baseline characteristics

ment groups were balanced across the variables (table 1), except for gender, with a higher proportion of women assigned to the control group. Scores on the HIQ were associated positively with scores on the AQ (r=0·68; p<0·0001) and with scores on the IAPT phobia scale (r=0·54; p<0·0001).Table 1 Baseline characteristics VR treatment group (n=49) Control group (n=51) Age (years) 45 (30–53) 46 (38–53) Men 29 19 Women 20 32 Ethnic origin White 47 45 Black African 0 1 Black Caribbean 1 0 Other 1 5 Employment status Full-time employed 30 33 Part-time employed 7 8 Unemployed 2 3 Retired 6 6 Student 4 1 Duration of fear of heights (years) 32·0 (13·8) 28·4 (15·0) Diagnosis of acrophobia 42 48 Data are number of participants, median (IQR), or mean (SD). Uptake of the VR treatment was very high. 47 (96%) of 49 people attended at least one VR session. The mean number of treatment sessions attended by these 47 participants was 4·66 (SD 1·27). The mean length of a treatment session was 26·8 min (SD 2·7). The mean total VR treatment time for these individuals was 124·43 min (SD 34·23). 44 (90%) people had a full course of the VR intervention (ie, they completed all the treatment exercises), which comprised three (n=4), four (n=18), five (n=8), six (n=12), seven (n=1), or eight (n=1) treatment sessions. Three (6%) people did not complete the intervention, with two people finding the VR sessions too difficult (attending three and four sessions) and one person unable to attend further appointment times (attending one session).

ee (n=4), four (n=18), five (n=8), six (n=12), seven (n=1), or eight (n=1) treatment sessions. Three (6%) people did not complete the intervention, with two people finding the VR sessions too difficult (attending three and four sessions) and one person unable to attend further appointment times (attending one session). All participants were followed up at all timepoints (baseline, 2 weeks, and 4 weeks) and no outcome data were missing. Table 2 summarises the mean scores of the primary and secondary outcome measures at every assessment point for each treatment group and the adjusted differences between groups. Participants allocated to the VR treatment group had very large reductions in scores on the three fear-of-heights assessments from baseline to 4 weeks, whereas scores for the control group remained stable (figure 2).Figure 2 Scores on the HIQ at every timepoint for each randomised group The minimum score on the HIQ is 16. Bars represent the mean, error bars the 95% CI. HIQ=Heights Interpretation Questionnaire. VR=virtual reality. Table 2 Outcome measure scores at every timepoint and differences between groups

All participants were followed up at all timepoints (baseline, 2 weeks, and 4 weeks) and no outcome data were missing. Table 2 summarises the mean scores of the primary and secondary outcome measures at every assessment point for each treatment group and the adjusted differences between groups. Participants allocated to the VR treatment group had very large reductions in scores on the three fear-of-heights assessments from baseline to 4 weeks, whereas scores for the control group remained stable (figure 2).Figure 2 Scores on the HIQ at every timepoint for each randomised group The minimum score on the HIQ is 16. Bars represent the mean, error bars the 95% CI. HIQ=Heights Interpretation Questionnaire. VR=virtual reality. Table 2 Outcome measure scores at every timepoint and differences between groups VR treatment group (n=49) Control group (n=51) Adjusted group difference (95% CI)* Effect size (Cohen's d) p value HIQ total 0 weeks 52·5 (12·7) 54·2 (11·6) .. .. .. 2 weeks 28·1 (10·2) 53·0 (11·8) −24·0 (−27·7 to −20·3) 2·0 <0·0001 4 weeks 27·5 (11·1) 52·6 (12·8) −24·3 (−27·9 to −20·6) 2·0 <0·0001 AQ total 0 weeks 71·9 (26·6) 70·3 (22·6) .. .. .. 2 weeks 29·1 (19·8) 73·9 (23·6) −45·2 (−52·1 to −38·2) 1·8 <0·0001 4 weeks 25·1 (19·3) 69·9 (22·1) −45·1 (−52·1 to −38·2) 1·8 <0·0001 AQ anxiety subscale 0 weeks 59·0 (20·8) 58·0 (18·8) .. .. .. 2 weeks 24·0 (16·3) 61·1 (19·5) −37·3 (−43·0 to −31·6) 1·9 <0·0001 4 weeks 20·4 (15·7) 57·3 (18·1) −37·2 (−42·9 to −31·5) 1·9 <0·0001 AQ avoidance subscale 0 weeks 12·9 (6·6) 12·3 (5·1) .. .. .. 2 weeks 5·1 (4·2) 12·9 (5·5) −8·0 (−9·6 to −6·4) 1·4 <0·0001 4 weeks 4·7 (4·4) 12·6 (5·7) −8·0 (−9·6 to −6·4) 1·4 <0·0001 IAPT phobia scale 0 weeks 4·6 (1·9) 4·5 (2·0) .. .. .. 2 weeks 2·1 (1·6) 4·5 (1·6) −2·3 (−2·9 to −1·8) 1·2 <0·0001 4 weeks 1·7 (1·5) 4·6 (1·7) −2·9 (−3·4 to −2·3) 1·5 <0·0001 Data are mean (SD), unless otherwise indicated. AQ=acrophobia questionnaire. HIQ=Heights Interpretation Questionnaire. IAPT=Improving Access to Psychological Therapies. VR=virtual reality.

0) .. .. .. 2 weeks 2·1 (1·6) 4·5 (1·6) −2·3 (−2·9 to −1·8) 1·2 <0·0001 4 weeks 1·7 (1·5) 4·6 (1·7) −2·9 (−3·4 to −2·3) 1·5 <0·0001 Data are mean (SD), unless otherwise indicated. AQ=acrophobia questionnaire. HIQ=Heights Interpretation Questionnaire. IAPT=Improving Access to Psychological Therapies. VR=virtual reality. * Adjusted for treatment group, assessment timepoint, baseline score for the outcome scale, and the interaction between treatment group and assessment timepoint (HIQ, AQ, and IAPT); further adjustment was made for baseline HIQ score (AQ and IAPT). The difference was assessed by linear mixed effects models.

0) .. .. .. 2 weeks 2·1 (1·6) 4·5 (1·6) −2·3 (−2·9 to −1·8) 1·2 <0·0001 4 weeks 1·7 (1·5) 4·6 (1·7) −2·9 (−3·4 to −2·3) 1·5 <0·0001 Data are mean (SD), unless otherwise indicated. AQ=acrophobia questionnaire. HIQ=Heights Interpretation Questionnaire. IAPT=Improving Access to Psychological Therapies. VR=virtual reality. * Adjusted for treatment group, assessment timepoint, baseline score for the outcome scale, and the interaction between treatment group and assessment timepoint (HIQ, AQ, and IAPT); further adjustment was made for baseline HIQ score (AQ and IAPT). The difference was assessed by linear mixed effects models. For the primary outcome analysis, the mean change in the HIQ total score at 2 weeks was −24·5 (SD 13·1) in participants allocated the VR intervention and −1·2 (7·3) in the control group. The adjusted difference in the treatment effect was −24·0 (95% CI −27·7 to −20·3; d=2·0; p<0·0001), in favour of the VR intervention (table 2). The mean change in the HIQ total score at 4 weeks was −25·1 (SD 13·9) in the VR treatment group and −1·5 (7·8) in the control group (adjusted difference in treatment effect −24·3, 95% CI −27·9 to −20·6; d=2·0; p<0·0001; table 2). 49 (100%) participants in the VR group showed a reduction in fear of heights on the HIQ, with the mean reduction being 68·0% (SD 26·6). By the follow-up timepoint, 34 (69%) of 49 people in the VR treatment group fell below the trial's fear of heights entry criterion on the HIQ, compared with none of 51 people in the control group (risk difference 0·61, 95% CI 0·48–0·75; NNT 1·6). In the VR group, 25 (51%) participants showed a reduction of 75% or greater (risk difference 0·45, 95% CI 0·31–0·59; NNT 2·2), 38 (78%) had a reduction of 50% or greater (0·78, 0·66–0·89; NNT 1·3), and 44 (90%) had a reduction of 25% or greater (0·86, 0·76–0·96; NNT 1·2). No participants in the control group showed a reduction of 75% or greater or 50% or greater, and nine (18%) had a reduction of 25% or greater. The mean reduction in the fear-of-heights score on the HIQ in the control group was 3·3% (SD 23·0). Outcomes were consistent with participant-described benefits of treatment (panel).Panel Participants' comments about VR treatment “What I'm noticing is that in day-to-day life I'm much less averse to edges, and steps, and heights, and I'm noticing in myself that when I'm doing the VR and outside I'm able to say ‘Hello’ to the edge instead of bracing against it and backing up. When I'm doing the VR I'm, as best as I'm able to, being open and curious around me as much as I can and noticing how the anxiety feels in my body, and then noticing that it goes really quickly now. So, when I've always got anxious about an edge I could feel the adrenaline in my legs, that fight/flight thing; that's not happening as much now.

m, as best as I'm able to, being open and curious around me as much as I can and noticing how the anxiety feels in my body, and then noticing that it goes really quickly now. So, when I've always got anxious about an edge I could feel the adrenaline in my legs, that fight/flight thing; that's not happening as much now. I'm still getting a bit of a reaction to it, both in VR and outside as well, but it's much more brief, and I can then feel my thighs soften up as I'm not bracing up against that edge. I feel as if I'm making enormous progress, and feel very happy with what I've gained.” “Everything I thought it was going to be, it wasn't. I anticipated it was just going to be like a game, it was going to be something that wasn't going to arouse my senses. I found myself even after the third floor, fourth floor, going up, feeling nervous, anxious about what's about to happen next. It definitely pushed the limits in terms of what I thought I would be able to achieve, and then got me to go past that. Now that it's done, after my fourth session, I have to say I feel better for it. I've already been experimenting in the weeks to see what it would be like in a real-life environment. And what I would like to say is that it's absolutely brilliant, honestly, I do think it's made a huge difference. I do think my nervousness about heights is definitely a lot better.”

I have to say I feel better for it. I've already been experimenting in the weeks to see what it would be like in a real-life environment. And what I would like to say is that it's absolutely brilliant, honestly, I do think it's made a huge difference. I do think my nervousness about heights is definitely a lot better.” “I've just finished my sessions, I did four in total. Last week, after my third session, I went up to the Westgate [a shopping centre]; the difference in my mental capacity to deal with heights was amazing. Previously I wouldn't go anywhere near the edges, I was almost hanging right off, looking vertically down. The sessions I've had here have given me a lot to think about, and certainly with regards to my fear of heights it feels like it's helped a lot. So, very worthwhile doing.” “I'm 60 years old and I've had a fear of heights, an extreme fear of heights, all my life. I came to the centre, I've had three sessions of VR and I've already surpassed everything that I imagined I could. I didn't actually think I'd be able to get to what is known as level 2 but I've achieved that. I am absolutely confident that I will be going to several other levels over the next week or two. For me, whilst it's not easy and I can't say that my fear of heights has gone, it has certainly improved and my confidence has improved so, so far so good!” VR=virtual reality.

“I'm 60 years old and I've had a fear of heights, an extreme fear of heights, all my life. I came to the centre, I've had three sessions of VR and I've already surpassed everything that I imagined I could. I didn't actually think I'd be able to get to what is known as level 2 but I've achieved that. I am absolutely confident that I will be going to several other levels over the next week or two. For me, whilst it's not easy and I can't say that my fear of heights has gone, it has certainly improved and my confidence has improved so, so far so good!” VR=virtual reality. We did preplanned subgroup analyses by gender (appendix). Gender had an independent effect on several outcomes but no significant interaction terms were noted in the models between gender and treatment group (ie, gender did not moderate treatment effects). In both treatment groups, men seemed to do slightly better than women. In response to the imbalance observed in gender across the two treatment groups, we also did sensitivity tests adjusting the outcome analyses for gender (appendix). Adjusting the treatment differences for gender did not change the conclusions, with large treatment effects still shown on all measures at all timepoints (data not shown).

he imbalance observed in gender across the two treatment groups, we also did sensitivity tests adjusting the outcome analyses for gender (appendix). Adjusting the treatment differences for gender did not change the conclusions, with large treatment effects still shown on all measures at all timepoints (data not shown). No adverse events were reported by any participant in either treatment group. Levels of discomfort (as assessed by the SSQ) were very low before entering VR for the first time (n=47; mean total score 1·60 [SD 1·88]) and discomfort increased only slightly after being in VR (n=47; mean total score 3·81 [SD 3·80]; mean increase 2·21, 95% CI 1·24–3·18; p<0·0001). Before the last session of VR, levels of discomfort were also very low (n=37; mean total score 1·21 [SD 2·27]) and, again, increased only slightly after being in VR (n=37; mean total score 2·57 [SD 3·98]; mean increase 1·35, 95% CI 0·55–2·16; p=0·002). Discussion The findings of our large randomised controlled trial show that an automated psychological intervention delivered by immersive VR is highly effective for reduction of fear of heights. All participants were followed up at every timepoint, meaning that the treatment effect estimates were unbiased by missing data. Treatment uptake was very high, indicating that the VR intervention was well received. Levels of discomfort after a VR session were very low, particularly since trial participants were facing their feared situation.

llowed up at every timepoint, meaning that the treatment effect estimates were unbiased by missing data. Treatment uptake was very high, indicating that the VR intervention was well received. Levels of discomfort after a VR session were very low, particularly since trial participants were facing their feared situation. Findings of a meta-analysis indicated a mean effect size reduction in phobias with therapist-assisted exposure treatment (using real heights) of d=1·1;19 our automated VR treatment produced effect sizes that greatly exceeded this value (d=2·0). Therefore, the treatment effects produced were at least as good as—and most likely better—than the best psychological intervention delivered face-to-face with a therapist. The initial cost of software development was reasonably high, with a team comprising psychologists, programmers, script writers, and an actor working intensively over approximately 6 months, but subsequent costs for this treatment are very low, with no need for a therapist to be present and inexpensive consumer VR hardware used. As VR becomes more widespread in households, such a treatment could be used at home in the future. Our view is that automated immersive VR has the potential to increase access to the best psychological interventions radically.

ry low, with no need for a therapist to be present and inexpensive consumer VR hardware used. As VR becomes more widespread in households, such a treatment could be used at home in the future. Our view is that automated immersive VR has the potential to increase access to the best psychological interventions radically. Our trial has several limitations. First, we do not know how representative our participants are of the wider population of people with acrophobia. Second, we relied on established acrophobia questionnaires and did not test behaviour at real heights. Third, we did not test long-term outcomes of the VR treatment, because previous studies of VR-assisted therapy have shown that reductions in anxiety can last several years.20, 21 Fourth, since most people do not receive treatment for a fear of heights, the clinical question addressed in the trial was the pragmatic one of how effective in total was the VR treatment against the absence of other treatment. Therefore, we can only conclude that the automated VR treatment produces large reductions in fear of heights, but we cannot pinpoint which elements caused the reduction. For example, we do not know whether similarly clinically effective automation of treatment can occur without use of a virtual coach. Fifth, the VR treatment was brief, and it is possible that further benefits might occur with a longer duration of treatment. Increased effectiveness might also occur with integration of the VR treatment into behavioural experiments in the real world. We are piloting provision of the VR treatment over half a day, and the outcomes are similarly encouraging. Finally, the extent to which learning from our work will transfer to other mental health conditions, particularly those seen in secondary care services, is uncertain. Automated treatment of anxiety disorders using VR might be a more tractable problem than for other disorders. We believe that transferability can only be determined by investment in high-quality VR treatments that are then tested in clinical trials. This endeavour is very important in mental health research in view of the potential benefits that might result from greatly increasing access to evidence-based treatment for mental health disorders.

ansferability can only be determined by investment in high-quality VR treatments that are then tested in clinical trials. This endeavour is very important in mental health research in view of the potential benefits that might result from greatly increasing access to evidence-based treatment for mental health disorders. Supplementary Material Supplementary appendix Acknowledgments This research was funded by Oxford VR, a University of Oxford spin-out company, and by the National Institute of Health Research (NIHR) Oxford Health Biomedical Research Centre. DF is supported by an NIHR Research Professorship. The views expressed are those of the authors and not necessarily those of the National Health Service, the NIHR, or the Department of Health. Contributors DF led the design of the treatment and trial and drafted the trial report. JF, MS, BS, SK, and PH contributed to the design of the treatment. JF and PH contributed to the design of the trial. MS, BS, and SK led the virtual reality programming. PH, EA, MD, and PB gathered data. AN did trial analyses. All authors commented on the paper. Declaration of interests DF, JF, MS, and BS are among the co-founders of Oxford VR. DF, JF, PH, EA, and MD are employed by Oxford VR. MS and BS are co-founders of Virtual Bodyworks. SK is employed by Virtual Bodyworks. PB and AN declare no competing interests.

Research in context Evidence before this study We searched PubMed, BIOSIS Previews, CINAHL, the Cochrane Central Register of Controlled Trials, Embase, ERIC (Education Resources Information Center), MEDLINE, PsycINFO, OpenGrey, Web of Science Core Collection, ProQuest Dissertations and Theses (UK and Ireland), ProQuest Dissertations and Theses (Abstracts and International), and the WHO International Trials Registry Platform (including ClinicalTrials.gov) from database inception up to April 7, 2017, with no restrictions by language, for published and unpublished double-blind randomised controlled trials comparing amphetamines (including lisdexamfetamine), atomoxetine, bupropion, clonidine, guanfacine, methylphenidate, and modafinil with each other or placebo. We used the search terms: “adhd” OR “hkd” OR “addh” OR “hyperkine*” OR “attention deficit*” OR “hyper-activ*” OR “hyperactiv*” OR “overactive” OR “inattentive” OR “impulsiv*”, combined with a list of ADHD medications (appendix pp 3–15). We also hand-searched the websites of the US Food and Drug Administration, the European Medicines Agency, and relevant drug manufacturers, and references of previous systematic reviews and guidelines, for additional information. Further, we contacted study authors and drug manufacturers to gather unpublished information or data. Over the past few decades, a substantial increase has been noted across many countries in prescription of medications for attention-deficit hyperactivity disorder (ADHD). However, the benefits and safety of these medications remain a matter for debate. Published meta-analyses of head-to-head trials and network meta-analyses provide inconsistent findings on the comparative benefits and harms of ADHD medications.

cription of medications for attention-deficit hyperactivity disorder (ADHD). However, the benefits and safety of these medications remain a matter for debate. Published meta-analyses of head-to-head trials and network meta-analyses provide inconsistent findings on the comparative benefits and harms of ADHD medications. Added value of this study Our study, based on advanced methodology for network meta-analyses, represents the most comprehensive synthesis to date on the comparative efficacy and tolerability of medications for ADHD across age groups. Unlike previous network meta-analyses of ADHD treatments, we have included unpublished data, which were gathered systematically from study authors, the websites of regulatory agencies, and drug manufacturers, using a common set of inclusion criteria for trials in children, adolescents, and adults. We focused on a series of clinically relevant outcomes—namely, efficacy on ADHD core symptoms, global clinical functioning, tolerability, effects on weight and blood pressure, and acceptability. We also investigated important effect-modifiers (eg, dose and comorbidities). We retained only a few studies with outcomes beyond 12 weeks. All medications we included in our study (except modafinil in adults) were more efficacious than placebo for the acute treatment of ADHD. Medications for ADHD were less efficacious and less well tolerated in adults than in children and adolescents. However, included drugs were not equivalent, and their profile in terms of efficacy, tolerability, and acceptability varied across age groups.

ere more efficacious than placebo for the acute treatment of ADHD. Medications for ADHD were less efficacious and less well tolerated in adults than in children and adolescents. However, included drugs were not equivalent, and their profile in terms of efficacy, tolerability, and acceptability varied across age groups. Implications of all the available evidence Evidence from our network meta-analysis supports methylphenidate (in children and adolescents) and amphetamines (in adults) as the preferred first pharmacological choice for short-term pharmacological treatment of ADHD. This network meta-analysis should inform future guidelines and daily clinical decision-making on the choice of medications for ADHD across age ranges, along with available evidence on cost-effectiveness and considering patients' preferences. The paucity of trials with randomised outcomes beyond 12 weeks highlights the need to fund studies to assess long-term effects of these drugs. Furthermore, future research should include individual patient data in network meta-analyses of ADHD medications, which will allow a more reliable estimation of predictors of individual response.

als with randomised outcomes beyond 12 weeks highlights the need to fund studies to assess long-term effects of these drugs. Furthermore, future research should include individual patient data in network meta-analyses of ADHD medications, which will allow a more reliable estimation of predictors of individual response. Introduction Attention-deficit hyperactivity disorder (ADHD) is characterised by age-inappropriate and impairing levels of inattention, hyperactivity, or impulsivity, or a combination.1 It is estimated to affect around 5% of school-age children (aged ≤18 years)2 and 2·5% of adults worldwide.3 Annual incremental costs for ADHD have been estimated at US$143–266 billion in the USA4 and are substantial in other countries.5, 6 Available pharmacological treatments for ADHD include psychostimulants (eg, methylphenidate and amphetamines) and non-psychostimulant medications (eg, atomoxetine and α2-agonists). In the past few decades, prescriptions for ADHD drugs have increased significantly both in the USA7 and other countries.8 However, even though recommended in clinical guidelines,9, 10, 11, 12, 13, 14 the efficacy and safety of ADHD medications remains controversial.15, 16, 17 Furthermore, current guidelines are inconsistent in their treatment recommendations.9, 10, 11, 12, 13, 14 Although some guidelines rank methylphenidate over amphetamines (eg, in children),9 others recommend psychostimulants as first-line treatment without any distinction between methylphenidate and amphetamines being made.10, 11 Additionally, the non-psychostimulant atomoxetine is variously recommended by available guidelines as third-line,9 second-line,10, 11 and potentially first-line treatment.12 The methods used for sequencing these recommendations are not always specified and most commonly—including the 2018 UK National Institute for Health and Care Excellence (NICE) guidelines9—incorporate national drug licencing regulatory approval and cost-effectiveness with expert opinion in conjunction with the few head-to-head comparisons that are available.

ecommendations are not always specified and most commonly—including the 2018 UK National Institute for Health and Care Excellence (NICE) guidelines9—incorporate national drug licencing regulatory approval and cost-effectiveness with expert opinion in conjunction with the few head-to-head comparisons that are available. Network meta-analyses facilitate estimation of the comparative efficacy and tolerability of two or more interventions, even when they have not been investigated head-to-head in randomised controlled trials.18 Thus, compared with standard pairwise meta-analyses, network meta-analyses have been found to increase the precision of the estimates.18 Previous network meta-analyses in ADHD have focused on either children and adolescents19, 20, 21, 22, 23, 24 or adults only,25, 26, 27, 28 have typically compared only a few drugs,24, 25, 27, 29 or have addressed exclusively the safety of treatments.26 To fill this gap, we did a systematic review and network meta-analysis of double-blind randomised controlled trials in children, adolescents, and adults with ADHD, using data from published reports and unpublished data gathered systematically from drug manufacturers or study authors. We aimed specifically to compare ADHD medications in terms of efficacy on core ADHD symptoms, clinical global functioning, tolerability, acceptability, and other clinically important outcomes—eg, blood pressure and weight changes.

ts and unpublished data gathered systematically from drug manufacturers or study authors. We aimed specifically to compare ADHD medications in terms of efficacy on core ADHD symptoms, clinical global functioning, tolerability, acceptability, and other clinically important outcomes—eg, blood pressure and weight changes. Methods Search strategy and selection criteria We searched PubMed, BIOSIS Previews, CINAHL, the Cochrane Central Register of Controlled Trials, EMBASE, ERIC, MEDLINE, PsycINFO, OpenGrey, Web of Science Core Collection, ProQuest Dissertations and Theses (UK and Ireland), ProQuest Dissertations and Theses (abstracts and international), and the WHO International Trials Registry Platform, including ClinicalTrials.gov, from the date of database inception to April 7, 2017, with no language restrictions. We used the search terms “adhd” OR “hkd” OR “addh” OR “hyperkine*” OR “attention deficit*” OR “hyper-activ*” OR “hyperactiv*” OR “overactive” OR “inattentive” OR “impulsiv*” combined with a list of ADHD medications (appendix pp 3–15). The US Food and Drug Administration (FDA), European Medicines Agency (EMA), and relevant drug manufacturers' websites, and references of previous systematic reviews and guidelines, were hand-searched for additional information. We also contacted study authors and drug manufacturers to gather unpublished information and data (appendix p 15).

dministration (FDA), European Medicines Agency (EMA), and relevant drug manufacturers' websites, and references of previous systematic reviews and guidelines, were hand-searched for additional information. We also contacted study authors and drug manufacturers to gather unpublished information and data (appendix p 15). We included double-blind randomised controlled trials (parallel group, crossover, or cluster), of at least 1 week's duration, that enrolled children (aged ≥5 years and <12 years), adolescents (aged ≥12 years and <18 years), or adults (≥18 years) with a primary diagnosis of ADHD according to DSM-III, DSM III-R, DSM-IV(TR), DSM-5, ICD-9, or ICD-10. We did not restrict our search by ADHD subtype or presentation, gender, intelligence quotient (IQ), socioeconomic status, or comorbidities (except for those needing concomitant pharmacotherapy). We included studies if they assessed any of the following medications, as oral monotherapy, compared with each other or with placebo: amphetamines (including lisdexamfetamine), atomoxetine, bupropion, clonidine, guanfacine, methyl-phenidate (including dexmethylphenidate), and modafinil. We excluded studies with enrichment designs (eg, trials selecting drug responders only after a run-in phase), because these types of trial can potentially inflate efficacy and tolerability estimates. Full inclusion and exclusion criteria are in the appendix (pp 16, 17). Our study protocol was registered with PROSPERO (number CRD42014008976) and published.30 We followed the PRISMA extension for network meta-analyses.31

We included double-blind randomised controlled trials (parallel group, crossover, or cluster), of at least 1 week's duration, that enrolled children (aged ≥5 years and <12 years), adolescents (aged ≥12 years and <18 years), or adults (≥18 years) with a primary diagnosis of ADHD according to DSM-III, DSM III-R, DSM-IV(TR), DSM-5, ICD-9, or ICD-10. We did not restrict our search by ADHD subtype or presentation, gender, intelligence quotient (IQ), socioeconomic status, or comorbidities (except for those needing concomitant pharmacotherapy). We included studies if they assessed any of the following medications, as oral monotherapy, compared with each other or with placebo: amphetamines (including lisdexamfetamine), atomoxetine, bupropion, clonidine, guanfacine, methyl-phenidate (including dexmethylphenidate), and modafinil. We excluded studies with enrichment designs (eg, trials selecting drug responders only after a run-in phase), because these types of trial can potentially inflate efficacy and tolerability estimates. Full inclusion and exclusion criteria are in the appendix (pp 16, 17). Our study protocol was registered with PROSPERO (number CRD42014008976) and published.30 We followed the PRISMA extension for network meta-analyses.31 Procedures Data were extracted by at least two independent investigators. We assessed risk of bias with the Cochrane risk of bias tool.32 We estimated the certainty of evidence with the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach for network meta-analyses (appendix pp 18, 19).33

Our study protocol was registered with PROSPERO (number CRD42014008976) and published.30 We followed the PRISMA extension for network meta-analyses.31 Procedures Data were extracted by at least two independent investigators. We assessed risk of bias with the Cochrane risk of bias tool.32 We estimated the certainty of evidence with the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach for network meta-analyses (appendix pp 18, 19).33 Outcomes For our primary analyses we considered efficacy, which we measured as the change in severity of ADHD core symptoms based on clinicians' ratings for children, adolescents, and adults. The appendix (pp 273, 274) contains a list of rating scales considered for inclusion. For children and adolescents, we also considered teachers' ratings as a primary efficacy outcome because they provide a complementary view to clinicians' ratings, and information from multiple raters increases the validity of ADHD diagnosis.34 We also considered tolerability in children, adolescents, and adults—ie, the proportion of participants who left the study because of any side-effect.

imary efficacy outcome because they provide a complementary view to clinicians' ratings, and information from multiple raters increases the validity of ADHD diagnosis.34 We also considered tolerability in children, adolescents, and adults—ie, the proportion of participants who left the study because of any side-effect. Secondary outcomes included the change in severity of ADHD core symptoms based on parents' ratings for children and adolescents and self-reports for adults, clinical global functioning measured by the Clinical Global Impression–Improvement (CGI-I, clinicians' ratings), acceptability (ie, the proportion of participants who left the study for any reason), and change in weight and blood pressure. We assessed those outcomes available at the times closest to 12 weeks (primary endpoint), 26 weeks, and 52 weeks.

red by the Clinical Global Impression–Improvement (CGI-I, clinicians' ratings), acceptability (ie, the proportion of participants who left the study for any reason), and change in weight and blood pressure. We assessed those outcomes available at the times closest to 12 weeks (primary endpoint), 26 weeks, and 52 weeks. Statistical analysis We did all analyses separately for studies in children and adolescents and for studies in adults. First, we did pairwise meta-analyses (active drug vs placebo, or active drug vs another active drug) for all outcomes and comparisons at every available timepoint, using a random-effects model.35 We calculated standardised mean differences (SMDs), Hedges's adjusted g, and odds ratios (ORs), with relative 95% CIs, for continuous and dichotomous outcomes. We assessed statistical heterogeneity within each pairwise comparison by calculating the I2 statistic and its 95% CI.36 Second, we did network meta-analyses within a frequentist framework assuming equal heterogeneity parameter τ across all comparisons and accounting for correlations induced by multiarm studies.37, 38 We based the assessment of statistical heterogeneity in the entire network on the magnitude of the common τ2 estimated from the network meta-analysis models.39 We compared the magnitude of the heterogeneity variance with the empirical distribution.40, 41 We used the loop-specific approach42 and the design-by-treatment model43 to evaluate incoherence locally and globally, respectively. We established a hierarchy of competing interventions using surface under the cumulative ranking curve (SUCRA) and mean ranks.44

erogeneity variance with the empirical distribution.40, 41 We used the loop-specific approach42 and the design-by-treatment model43 to evaluate incoherence locally and globally, respectively. We established a hierarchy of competing interventions using surface under the cumulative ranking curve (SUCRA) and mean ranks.44 We planned a set of subgroup and sensitivity analyses to assess the effect of clinical and study design effect-modifiers—eg, duration of study, gender, age (children vs adolescents), psychiatric comorbidities, IQ, crossover design, medication status, industry sponsorship, inequalities in doses, risk of bias, and data imputation.30 We restricted the primary analysis to studies using medications within the therapeutic range (as per FDA recommendations, where applicable). Additionally, we investigated effects at different dose regimens in two sets of sensitivity analyses. First, we excluded studies that did not use the FDA-licensed dose (appendix pp 277–79). Second, we included studies in which the dose ranges used were recommended in national or international guidelines or formularies but differed from FDA recom-mendations. Finally, to investigate possible differences between lisdexamfetamine and other amphetamines, we did a post-hoc analysis separating this compound, because lisdexamfetamine is metabolised differently from other amphetamines, which could affect its efficacy and tolerability.45

ries but differed from FDA recom-mendations. Finally, to investigate possible differences between lisdexamfetamine and other amphetamines, we did a post-hoc analysis separating this compound, because lisdexamfetamine is metabolised differently from other amphetamines, which could affect its efficacy and tolerability.45 We did all analyses with STATA version 14. Additional details are reported in the appendix (pp 20–24, 277–82). Changes to the original protocol are listed in the appendix (p 25). Role of the funding source The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. SCo, NA, CDG, and AC had full access to all data in the study, and AC was responsible for the final decision to submit for publication.

We did all analyses with STATA version 14. Additional details are reported in the appendix (pp 20–24, 277–82). Changes to the original protocol are listed in the appendix (p 25). Role of the funding source The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report. SCo, NA, CDG, and AC had full access to all data in the study, and AC was responsible for the final decision to submit for publication. Results The literature search, study selection, and data extraction were done between Jan 11, 2014, and Sept 9, 2017, and data analysis was done from Sept 10, 2017, to Feb 24, 2018. The study selection process is shown in figure 1; a list of excluded studies, with reasons for exclusion, and a list of retained studies is provided in the appendix (pp 26–272). 133 studies were retained for the network meta-analysis, 81 in children and adolescents, 51 in adults, and one including children, adolescents, and adults. In total, 14 346 children and adolescents and 10 296 adults were included. For 83% of studies, additional data and information not reported in the full-text paper were used. The appendix (pp 283–381) reports the main characteristics of included studies. The risk of bias was rated overall low in 23·5% of studies in children and adolescents, unclear in 65·4%, and high in 11·1%. The risk of bias was overall low in 27·5% of studies in adults, unclear in 56·8%, and high in 15·7% (appendix pp 382–458).Figure 1 Selection of studies for inclusion

eristics of included studies. The risk of bias was rated overall low in 23·5% of studies in children and adolescents, unclear in 65·4%, and high in 11·1%. The risk of bias was overall low in 27·5% of studies in adults, unclear in 56·8%, and high in 15·7% (appendix pp 382–458).Figure 1 Selection of studies for inclusion *The main reasons for exclusion included open-label or single-blind studies, studies including patients with comorbid disorders, and combination therapy trials. We only searched for completed trials, which removed ongoing studies, particularly from clinicaltrials.gov. Figure 2 shows the network plots for the primary outcomes closest to 12 weeks. Network plots for secondary outcomes are reported in the appendix (pp 624–29). Results of the pairwise meta-analyses and related heterogeneity are reported in the appendix (pp 459–71). Results of the network meta-analyses of primary outcomes at 12 weeks are shown in figure 3, Table 1, Table 2, and the appendix (pp 472, 473). Table 1, Table 2 also show the confidence of estimates for every comparison. Figure 4 summarises data for efficacy (in 10 068 children and adolescents and 8131 adults) and tolerability (in 11 018 children and adolescents and 5362 adults).Figure 2 Network of eligible comparisons for efficacy and tolerability

472, 473). Table 1, Table 2 also show the confidence of estimates for every comparison. Figure 4 summarises data for efficacy (in 10 068 children and adolescents and 8131 adults) and tolerability (in 11 018 children and adolescents and 5362 adults).Figure 2 Network of eligible comparisons for efficacy and tolerability The width of the lines is proportional to the number of trials comparing every pair of treatments, and the size of every circle is proportional to the number of randomly assigned participants (sample size). The number of trials for pairs of treatments ranged from 22 (eg, studies of tolerability of methylphenidate vs placebo in children and adolescents) to one (several comparisons). Figure 3 Forest plots of network meta-analysis results Plots include all trials for efficacy and tolerability and are compared with placebo as reference. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. ADHD=attention-deficit hyperactivity disorder. OR=odds ratio. SMD=standardised mean difference. Table 1 Effect of ADHD drugs in children and adults at timepoints closest to 12 weeks in terms of efficacy, as rated by clinicians and teachers

Plots include all trials for efficacy and tolerability and are compared with placebo as reference. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. ADHD=attention-deficit hyperactivity disorder. OR=odds ratio. SMD=standardised mean difference. Table 1 Effect of ADHD drugs in children and adults at timepoints closest to 12 weeks in terms of efficacy, as rated by clinicians and teachers Atomoxetine Bupropion Clonidine Guanfacine Methylphenidate Modafinil Placebo Children Adults Children Adults Children Adults Children Adults Children Adults Children Adults Children Adults Amphetamines Clinicians −0·46 (−0·65 to −0·27)* −0·34 (−0·58 to −0·10)* −0·06 (−0·81 to 0·68)† −0·33 (−0·77 to 0·11)* −0·31 (−0·81 to 0·18)* .. −0·35 (−0·59 to −0·10)* .. −0·24 (−0·44 to −0·05)* −0·29 (−0·54 to −0·05)* −0·39 (−0·67 to −0·12)* −0·94 (−1·43 to −0·46)‡ −1·02 (−1·19 to −0·85)‡ −0·79 (−0·99 to −0·58)‡ Teachers .. .. .. .. .. .. .. .. .. .. .. .. .. .. Atomoxetine Clinicians .. .. 0·40 (−0·34 to 1·14)* 0·01 (−0·41 to 0·42)* 0·15 (−0·33 to 0·63)* .. 0·11 (−0·09 to 0·32)* .. 0·22 (0·05 to 0·39)* 0·04 (−0·14 to 0·23)‡ 0·07 (−0·17 to 0·31)* −0·61 (−1·06 to −0·15)* −0·56 (−0·66 to −0·45)* −0·45 (−0·58 to −0·32)* Teachers .. .. 0·00 (−0·90 to 0·90)† .. .. .. 0·31 (−0·79 to 1·42)† .. 0·50 (−0·11 to 1·10)* .. 0·44 (− 0·19 to 1·07)* .. −0·32 (−0·82 to 0·18)† .. Bupropion Clinicians .. .. .. .. −0·25 (−1·12 to 0·62)† .. −0·28 (−1·04 to 0·47)† .. −0·18 (−0·90 to 0·54)† 0·04 (−0·38 to 0·45)* −0·33 (−1·10 to 0·43)† −0·62 (−1·20 to −0·03)* −0·96 (−1·69 to −0·22)‡ −0·46 (−0·85 to −0·07)* Teachers .. .. .. .. .. .. 0·31 (−0·92 to 1·55)† .. 0·50 (−0·17 to 1·17)* .. 0·44 (−0·38 to 1·26)* .. −0·32 (−1·07 to 0·43)† .. Clonidine Clinicians .. .. .. .. .. .. −0·03 (−0·53 to 0·46)† .. 0·07 (−0·42 to 0·56)† .. −0·08 (−0·59 to 0·43)† .. −0·71 (−1·17 to −0·24)‡ .. Guanfacine Clinicians .. .. .. .. .. .. .. .. 0·11 (−0·13 to 0·34)* .. −0·05 (−0·32 to 0·23)* .. −0·67 (−0·85 to −0·50)‡ .. Teachers .. .. .. .. .. .. .. .. 0·18 (−0·86 to 1·22)† .. 0·12 (−0·93 to 1·18)† .. −0·63 (−1·62 to 0·35)† .. Methylphenidate Clinicians .. .. .. .. .. .. .. .. .. .. −0·15 (−0·41 to 0·10)* −0·65 (−1·11 to −0·19)* −0·78 (−0·93 to −0·62)‡ −0·49 (−0·64 to −0·35)‡ Teachers .. .. .. .. .. .. .. .. .. .. −0·06 (−0·53 to 0·42)† .. −0·82 (−1·16 to −0·48)* .. Modafinil Clinicians .. .. .. .. .. .. .. .. .. .. .. .. −0·62 (−0·84 to −0·41)* 0·16 (−0·28 to 0·59)* Teachers .. .. .. .. .. .. .. .. .. .. .. .. −0·76 (−1·15 to −0·37)† .. Data are standardised mean difference (95% CI) between treatments.

. .. .. .. .. .. .. .. .. −0·06 (−0·53 to 0·42)† .. −0·82 (−1·16 to −0·48)* .. Modafinil Clinicians .. .. .. .. .. .. .. .. .. .. .. .. −0·62 (−0·84 to −0·41)* 0·16 (−0·28 to 0·59)* Teachers .. .. .. .. .. .. .. .. .. .. .. .. −0·76 (−1·15 to −0·37)† .. Data are standardised mean difference (95% CI) between treatments. Results in bold are significant. Negative values favour the treatment in the row and positive values favour the treatment in the column. Drugs are reported in alphabetical order. Results are based on network estimates. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. No teacher ratings were available for clonidine. ADHD=attention-deficit hyperactivity disorder. * Low quality of evidence. † Very low quality of evidence. ‡ Moderate quality of evidence. Table 2 Effect of ADHD drugs in children and adults at timepoints closest to 12 weeks in terms of tolerability

Results in bold are significant. Negative values favour the treatment in the row and positive values favour the treatment in the column. Drugs are reported in alphabetical order. Results are based on network estimates. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. No teacher ratings were available for clonidine. ADHD=attention-deficit hyperactivity disorder. * Low quality of evidence. † Very low quality of evidence. ‡ Moderate quality of evidence. Table 2 Effect of ADHD drugs in children and adults at timepoints closest to 12 weeks in terms of tolerability Atomoxetine Bupropion Clonidine Guanfacine Methylphenidate Modafinil Placebo Children Adults Children Adults Children Adults Children Adults Children Adults Children Adults Children Adults Amphetamines 1·54 (0·79–3·01)* 1·40 (0·54–3·66)† 1·53 (0·17–13·88)† 1·28 (0·14–11·40)† 0·51 (0·08–3·27)† .. 0·87 (0·35–2·16)† .. 1·60 (0·94–2·73)* 1·36 (0·54–3·43)† 1·72 (0·64–4·59)† 0·81 (0·23–2·93)† 2·30 (1·36–3·89)‡ 3·26 (1·54–6·92)‡ Atomoxetine .. .. 0·99 (0·11–9·15)† 0·91 (0·11–7·77)† 0·33 (0·05–2·14)† .. 0·57 (0·22–1·47)† .. 1·04 (0·55–1·94)† 0·97 (0·47–2·02)* 1·11 (0·40–3·09)† 0·58 (0·18–1·93)† 1·49 (0·84–2·64)* 2·33 (1·28–4·25)* Bupropion .. .. .. .. 0·33 (0·02–5·51)† .. 0·57 (0·06–5·77)† .. 1·05 (0·12–9·14)† 1·07 (0·13–8·92)† 1·12 (0·11–11·62)† 0·64 (0·06–6·37)† 1·51 (0·17–13·27)† 2·55 (0·33–19·93)† Clonidine .. .. .. .. .. .. 1·71 (0·24–12·22)† .. 3·14 (0·51–19·33)† .. 3·36 (0·46–24·64)† .. 4·52 (0·75–27·03)† .. Guanfacine .. .. .. .. .. .. .. .. 1·83 (0·74–4·57)† .. 1·97 (0·63–6·16)† .. 2·64 (1·20–5·81)* .. Methylphenidate .. .. .. .. .. .. .. .. .. .. 1·07 (0·41–2·83)† 0·60 (0·19–1·92)† 1·44 (0·90–2·31)* 2·39 (1·40–4·08)§ Modafinil .. .. .. .. .. .. .. .. .. .. .. .. 1·34 (0·57–3·18)† 4·01 (1·42–11·33)‡ Data are odds ratio (95% CI). Values above 1 favour the treatment in the column and values below 1 favour the treatment in the row. Results in bold are significant. Drugs are reported in alphabetical order. Results are based on network estimates. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. ADHD=attention-deficit hyperactivity disorder.

reatment in the row. Results in bold are significant. Drugs are reported in alphabetical order. Results are based on network estimates. No data for clonidine and guanfacine in adults are reported because no studies identified by our search tested these two drugs in adults. ADHD=attention-deficit hyperactivity disorder. * Low quality of evidence. † Very low quality of evidence. ‡ Moderate quality of evidence. § High quality of evidence. Figure 4 Two-dimensional graphs of efficacy versus tolerability in studies in children and adolescents and adults Effect sizes for individual drugs are represented by coloured nodes, with bars representing corresponding 95% CIs. With respect to ADHD core symptoms rated by clinicians in children and adolescents, all drugs were superior to placebo (figure 3, table 1). In adults, amphetamines, methylphenidate, bupropion, and atomoxetine were superior to placebo, but modafinil was not superior to placebo; no data were available for guanfacine and clonidine. In children, adolescents, and adults, amphetamines were significantly superior to modafinil, atomoxetine, and methylphenidate (table 1). Additionally, in children and adolescents, amphetamines were superior to guanfacine and methylphenidate was superior to atomoxetine. In adults, methylphenidate, atomoxetine, and bupropion were superior to modafinil. By contrast, according to teachers' ratings of children's ADHD core symptoms, only methylphenidate and modafinil were superior to placebo (no data were available for amphetamines and clonidine; table 1).

date was superior to atomoxetine. In adults, methylphenidate, atomoxetine, and bupropion were superior to modafinil. By contrast, according to teachers' ratings of children's ADHD core symptoms, only methylphenidate and modafinil were superior to placebo (no data were available for amphetamines and clonidine; table 1). With respect to tolerability, in children and adolescents, only guanfacine and amphetamines were less well tolerated than placebo (figure 3, table 2). In adults, modafinil, amphetamines, methylphenidate, and atomoxetine were inferior to placebo (no data were available for guanfacine and clonidine). No differences in tolerability were noted between active drugs, in children, adolescents, and adults. In children and adolescents, the common heterogeneity SD for efficacy (teachers' and clinicians' ratings) and tolerability was 0·355, 0·188, and 0·268, respectively. In adults, the common heterogeneity SD for efficacy rated by clinicians and tolerability was 0·178 and 0·282, respectively. The test of global inconsistency did not show any significant difference for the primary outcomes. Additional details are reported in the appendix (p 474).

s 0·355, 0·188, and 0·268, respectively. In adults, the common heterogeneity SD for efficacy rated by clinicians and tolerability was 0·178 and 0·282, respectively. The test of global inconsistency did not show any significant difference for the primary outcomes. Additional details are reported in the appendix (p 474). Parents' ratings of their child's ADHD core symptoms and adults' self-ratings of their own ADHD core symptoms, with respect to efficacy of active drugs versus placebo, were similar to clinicians' ratings. Exceptions were guanfacine, which was not superior to placebo according to parents' ratings (SMD −0·23, 95% CI −0·90 to 0·45), and bupropion, which was not superior to placebo with respect to parents' ratings (0·24, −0·44 to 0·92) and adults' self-reports (−0·30, −0·61 to 0·01; appendix pp 475–76). In children and adolescents, all compounds were superior to placebo on the CGI-I scale, except for clonidine (OR 2·78, 95% CI 0·91–8·53). In adults, amphetamines (4·86, 3·30–7·17), bupropion (3·43, 1·45–8·14), and methylphenidate (3·08, 2·04–4·65) were superior to placebo on the CGI-I scale (appendix p 476).

Parents' ratings of their child's ADHD core symptoms and adults' self-ratings of their own ADHD core symptoms, with respect to efficacy of active drugs versus placebo, were similar to clinicians' ratings. Exceptions were guanfacine, which was not superior to placebo according to parents' ratings (SMD −0·23, 95% CI −0·90 to 0·45), and bupropion, which was not superior to placebo with respect to parents' ratings (0·24, −0·44 to 0·92) and adults' self-reports (−0·30, −0·61 to 0·01; appendix pp 475–76). In children and adolescents, all compounds were superior to placebo on the CGI-I scale, except for clonidine (OR 2·78, 95% CI 0·91–8·53). In adults, amphetamines (4·86, 3·30–7·17), bupropion (3·43, 1·45–8·14), and methylphenidate (3·08, 2·04–4·65) were superior to placebo on the CGI-I scale (appendix p 476). Weight was decreased significantly by amphetamines (in children and adolescents, SMD −0·71, 95% CI −1·15 to −0·27; in adults, −0·60, −1·03 to −0·18), methylphenidate (in children and adolescents, −0·77, −1·09 to −0·45; in adults, −0·74, −1·20 to −0·28), atomoxetine (in children and adolescents, −0·84, −1·16 to −0·52), and modafinil (in children and adolescents, −0·93, −1·59 to −0·26), compared with placebo (appendix pp 476, 477). Systolic blood pressure was increased with use of amphetamines (SMD 0·09, 95% CI 0·01–0·18) and atomoxetine (0·12, 0·02–0·22) in children and adolescents, and with use of methylphenidate (0·17, 0·05–0·30) in adults, compared with placebo (appendix p 477). Use of amphetamines (0·21, 0·12–0·31), atomoxetine (0·28, 0·18–0·37), and methylphenidate (0·24, 0·14–0·33) in children and adults, and atomoxetine (0·19, 0·08–0·30) and methylphenidate (0·20, 0·08–0·32) in adults, significantly increased diastolic blood pressure compared with placebo (appendix p 478).

appendix p 477). Use of amphetamines (0·21, 0·12–0·31), atomoxetine (0·28, 0·18–0·37), and methylphenidate (0·24, 0·14–0·33) in children and adults, and atomoxetine (0·19, 0·08–0·30) and methylphenidate (0·20, 0·08–0·32) in adults, significantly increased diastolic blood pressure compared with placebo (appendix p 478). For acceptability, compared with placebo, methylphenidate (OR 0·69, 95% CI 0·52–0·91) in children and adolescents and amphetamines (0·68, 0·49–0·95) in adults were significantly better (appendix p 478). In subgroup and sensitivity analyses, data were sufficient to assess the effect of study length, comorbidities, IQ, crossover design, unfair dose comparisons, and data imputation. Findings of these analyses were generally robust (appendix pp 479–91). Because of a paucity of data, we could not assess the effect of gender, age (children vs adolescents), low risk of bias, medication status, and industry sponsorship. Sensitivity analyses investigating the effect of different maximum doses confirmed the results of the primary dose analysis (appendix pp 492–575).

79–91). Because of a paucity of data, we could not assess the effect of gender, age (children vs adolescents), low risk of bias, medication status, and industry sponsorship. Sensitivity analyses investigating the effect of different maximum doses confirmed the results of the primary dose analysis (appendix pp 492–575). Post-hoc analyses separating lisdexamfetamine from other amphetamines highlighted some differences. In children, lisdexamfetamine was less well tolerated compared with placebo (OR 2·69, 95% CI 1·40–5·16), whereas tolerability of the other amphetamines was slightly better (1·83, 0·84–4·02); in adults, the opposite pattern emerged (vs placebo: lisdexamfetamine, 2·74, 0·80–9·30; other amphetamines, 3·66, 1·36–9·87). Network meta-analyses heterogeneity for the dose and post-hoc analyses are reported in the appendix (pp 576–78). Data for network meta-analyses inconsistency and SUCRA and mean rank are reported in the appendix (pp 579–85). Empirical heterogeneity variance for continuous outcomes for drug versus placebo comparisons was 0·05 (50% percentile) and 0·24 (75% percentile); for binary outcomes it was 0·12 (50% percentile) and 0·34 (75% percentile). Funnel plots are shown in the appendix (pp 630–32). We retained only a few studies—all in adults—with reported outcomes closest to 26 weeks or 52 weeks (appendix pp 586–88); therefore results for outcomes at these timepoints were deemed not informative.