Bipolar disorder: Trimodal age‐at‐onset distribution

Abstract Objective Bipolar disorder (BD) is a chronic mental health disorder with significant morbidity and mortality. Age at onset (AAO) may be a key variable in delineating more homogeneous subgroups of BD patients. However, no known research has systematically assessed how BD age‐at‐onset subgroups should be defined. Methods We systematically searched the following databases: Cochrane Central Register of Controlled Trials, PsycINFO, MEDLINE, Embase, CINAHL, Scopus, Proquest Dissertations and Theses, Google Scholar and BIOSIS Previews. Original quantitative English language studies investigating AAO in BD were sought. Results A total of 9454 unique publications were identified. Twenty‐one of these were included in data analysis (n = 22981 BD participants). Fourteen of these studies (67%, n = 13626 participants) found a trimodal AAO distribution: early‐onset (µ = 17.3, σ = 1.19, 45% of sample), mid‐onset (µ = 26.0, σ= 1.72, 35%), and late‐onset (µ = 41.9, σ= 6.16, 20%). Five studies (24%, n = 1422 participants) described a bimodal AAO distribution: early‐onset (µ = 24.3, σ = 6.57, 66% of sample) and late‐onset (µ = 46.3, σ = 14.15, 34%). Two studies investigated cohort effects on BD AAO and found that when the sample was not split by cohort, a trimodal AAO was the winning model, but when separated by cohort a bimodal distribution fit the data better. Conclusions We propose that the field conceptualises bipolar disorder age‐at‐onset subgroups as referring broadly to life stages. Demarcating BD AAO groups can inform treatment and provide a framework for future research to continue to investigate potential mechanisms of disease onset.

homogeneous subgroups of BD patients. 4 To date no research has systematically validated what the various AAO subgroups should be, and there is no concurrence across studies regarding what is meant by 'early onset'. 5,6 In recent years, it has been acknowledged that AAO in bipolar disorder is not a simple unimodal distribution, but can better be explained by a mixture of distributions. Evidence has suggested that BD aggregates either into a bimodal distribution with two subgroups (early vs. late AAO), or a trimodal distribution with three subgroups (early vs. mid vs. late AAO). 3,4,[7][8][9][10] However, it is not known which of these distribution modalities are more reliable and consistent. A better understanding of the distribution of age at onset in BD could provide an insight into the causes and mechanisms of illness, anticipate disease trajectory, and guide appropriate timeframes for primary and secondary prevention. 11 Understanding the age-at-onset distribution of bipolar disorder over the life course also has implications relating to the conduct of clinical and epidemiological research, and health service provision and planning.

| Objective
The aim of this systematic review was to investigate age-at-onset distributions in bipolar disorder, and correspondingly what constitutes an early age at onset. Only studies that use a data-driven approach to define AAO groups were included in data synthesis, as segregating BD into AAO groups using pre-defined cut-offs is an inherently biased approach. One of the most popular analysis approaches for determining AAO groups is admixture analysis, as it explores the theoretical model that best fits the observed distribution of a continuous variable.
All studies had to meet four criteria: 1) include participants who were recruited with a primary diagnosis of BD I, II or not otherwise specified (NOS); 2) report on the distribution of bipolar disorder AAO using a data-driven analysis approach (e.g. admixture analysis); 3) be an original article including epidemiological, cohort, longitudinal, cross-sectional, survey or observational studies and 4) be an English language article.
Animal research, single case studies, duplicates, conference abstracts or articles with unobtainable missing data were excluded.

| Search strategy
In February 2019 we carried out searches of the following data- Search strategies were developed using medical subject headings (MeSH) and text words related to bipolar disorder, age at onset and study type (Supplement S1). The syntax and subject headings of the search strategies were adapted for each database, and Boolean operators and truncation were used to extend the search terms. No date limits were imposed on the searches.

| Study selection
The web-based systematic review software, DistillerSR, 12 was used to complete screening and data extraction. Two reviewers (SB, JW) independently screened titles and abstracts. A third reviewer (KS) resolved any eligibility conflicts. Following title and abstract screening, we conducted full-text reviews of eligible articles. Where necessary, we sought additional information from study authors to resolve questions about eligibility and obtain missing data. We did not quality assess included studies as accepted standards of quality assessing non-randomised studies are lacking, 13 and our articles employed a broad range of study designs with diverging methodologies and reporting standards.

| Data analysis
We extracted data from eligible studies using a standardised data extraction form. This included data on diagnoses, recruitment strategies, demographics and details of age-at-onset groups, including means, standard deviations (SDs) and age ranges. From our extracted data, we computed summary statistics for each study. These summary statistics related to participant characteristics (sample size, age range and gender ratio); diagnostic criteria used; age-at-onset definition; recruitment settings (clinic, community and hospital) and study locations. Two studies 14,15 recruited mixed samples which included participants with schizoaffective disorder; where possible, we only included participants with a BD diagnosis in analyses and samples with schizoaffective disorder participants were excluded.
We separated the studies into those reporting a trimodal AAO distribution, a bimodal distribution and those investigating cohort effects on AAO. For each study, we then extracted the average AAO per subgroup-for those studies reporting a trimodal AAO distribution we extracted the mean and SD for the early-, mid-and late-onset groups, and for those studies reporting a bimodal AAO distribution, we extracted means and SDs for the early-and late-onset groups. These averages were used to plot probability density functions and boxplots for each AAO group in studies reporting a trimodal versus bimodal AAO distribution. To do this, we used the ggplot2 16 data visualisation package in RStudio (version 1.2.1335) 17 -the analysis code can be found on our Open Science Framework (OSF) webpage (https://osf. io/5c89s/). Our data are also openly available via the OSF. 18

| RE SULTS
Our search produced 14,129 results. After duplicates were removed, we screened the titles and abstracts of 9454 articles for relevance and considered 74 eligible for full-text review (PRISMA Diagram Figure 1; see Supplement S2 for a list of studies excluded at full-text review). Twentyfour articles met full-text eligibility criteria, and 21 articles were included in data synthesis. We excluded 3 of the 24 eligible studies 19-21 due to missing data, which were unobtainable after contacting the authors.

| Study characteristics
All included studies were conducted from 2001 to 2017, with the majority (n = 15, 71%) published from 2009 onwards.

| Participants
Across all studies, there were a total of 22,904 bipolar disorder participants, with an average sample size of 1,094 participants per study. In total, there were 22,165 (96.78% of total) participants with a diagnosis of BDI, 653 (2.85%) with BDII, 12 (0.05%) with BD-NOS and 74 (0.32%) with schizoaffective disorder. There were more female than male participants, with an average of 59.9% female participants across all studies.

| Age of participants at study entry
Several studies did not report age ranges or average age of their samples; those that did (n = 15) had an overall average age of 43.2 years.

| Age-at-onset definitions
Heterogeneous definitions of AAO were used across studies including: age at which diagnostic criteria for an affective episode was first met according to medial case notes, interviews or selfreport 5,7,9,[23][24][25][26][27]29,30,[32][33][34][35] ; age at first impairment due to an affective episode according to self-report 8 ; age at first contact with psychiatric services for symptoms of mania 14,22 ; age at first treatment for an affective disorder 28 and age at first psychiatric hospitalisation. 36 Across all studies, AAO was determined retrospectively using information gathered from medical records and/or interviews with participants and their relatives.

| Recruitment setting
Seven studies recruited patients from a clinic setting only, 9,15,23,27,29,34,35 two from community settings only 8,32 and two from inpatient hospital settings only. 25,36 Three studies recruited from both the clinic and the community, 24

| Study locations
The largest of the included studies collected data on 4037 bipolar patients across 36 collection sites in 23 countries throughout Asia, Africa, Europe, North and South America and Australia. 24 Of the remaining studies, eleven were conducted in Europe, 5,7,8,22,23,25,28,29,31,34,35 six in North America, 9,14,15,32,33,36 one in Australia 27 and two articles combined data from collection sites in both North America and Europe. 26,30

| Age-at-onset distributions
There were three separate types of distributions found for bipolar disorder age at onset across the 21 articles. Fourteen studies showed a trimodal distribution, 5

| Trimodal age-at-onset distribution
Fourteen (67%) of the included 21 articles, including 59% (n = 13,549) of all participants, reported a trimodal age-at-onset distribution with three subgroups: early onset, mid-onset and late onset ( Table 1). Eight of these studies were conducted in Europe, three in America, two in both North America and Europe, and one in Australia. Of the fourteen studies, nine included participants with a diagnosis of BDI only, three with a diagnosis of BDI, BDII or BD-NOS, and two with a diagnosis of BDI, BDII or schizoaffective disorder.
Of the two studies including schizoaffective disorder patients, 14,35 Javaid et al (2011) report their findings including and excluding participants with schizoaffective disorder. We use the results of the 'bipolar only' sample in our analyses.
Two of these fourteen studies had a partial overlap in their samples 34,35 (Table 1)

| Bimodal age-at-onset distribution
Five studies (24%), representing 6% (n = 1422) of all participants, described a bimodal age-at-onset distribution with two subgroups: early onset and late onset ( Table 2). Two of these studies were conducted in Europe and three in North America.
Three of the studies included participants with a diagnosis of BDI only, and two included those with a diagnosis of BDI, BDII or BD-NOS.
Across these five studies, the average age of early onset was 22.5 years (SD = 7.32) and late onset was 40.8 years (SD = 16.89).
Results indicated that an average of 63% out of a total of 1422 participants across the five studies displayed early onset, compared to 37% late onset ( Figure 3).

| Effect of birth cohort
The remaining 2 of the 21 included articles examined cohort effects on AAO (Table 3). Both studies examined the effect of birth cohorts on age at onset in samples of BDI patients (total n = 7,933) recruited from clinical and community settings.
When the effect of birth cohort was not modelled, both studies found a trimodal bipolar disorder AAO distribution.
When birth cohort was adjusted for, both studies reported that a bimodal distribution fit the data better. Across all cohorts in both studies, the overall mean ages for early, mid and late onset were 18.7 (SD = 1.52), 25.5 (SD = 1.47) and 29.4 (SD = 2.21)

| Age-at-onset distributions by study location and diagnostic criteria
Prior research has suggested that study location and BD diagnosis may influence AAO distributions. [37][38][39][40][41][42] Location Of the eleven studies conducted in Europe, eight found a trimodal AAO distribution, two found a bimodal distribution and one reported cohort effects. There was an even split between studies reporting bi-and tri-modal distributions (3 vs. 3) in North American samples.
Studies conducted in both Europe and North America found a trimodal AAO distribution. The one study conducted in Australia reported a trimodal distribution.

Diagnosis
Two thirds of studies included participants with a diagnosis of BDI only (n = 14, 67%). Nine of these studies (64%) found a trimodal AAO distribution, compared to three reporting a bimodal distribution (21%). Five studies (25%) recruited samples with BDI, BDII and BD-NOS. Three of these studies reported a trimodal distribution and two a bimodal distribution.
Two studies included schizoaffective disorder as a diagnostic category, and both of these studies reported a trimodal AAO distribution.
The impact of diagnosis and study location does not appear to affect AAO in a meaningful way (see Supplement S3: Supplementary   Table S2).

| DISCUSS ION
This is the first systematic review of age at onset (AAO) in bipolar disorder. The aim of this review was to provide a more reli-

| Defining early-onset
Our findings offer a more robust understanding of when bipolar dis- cases onset in early-life, from the ages of 14-21 years, with an average onset of 17.3 years. As it is customarily used, the term 'earlyonset' implies an 'earlier than expected AAO', whereas throughout our included studies the 'early-onset' group is the most common age range for the onset of BD. We therefore recommend that the term early onset should be reconceptualised to represent life stage rather than as a comparator. 'Early-onset' in the sense it is traditionally referred to is, thus, best described as onset before the age of 14 years.
This distinction has the potential to aid the interpretation of existing treatment guidelines, which currently offer recommendations for treating 'early-onset and early-stage' BD without providing corresponding definitions. 43 The diagnosis of pre-pubertal bipolar disorder, which is prevalent in North America, 41,44 has long been viewed as contentious due to high rates of comorbidities and elevated levels of symptom overlap with other juvenile psychiatric disorders. 45,46 Our findings do not directly refute the diagnosis of paediatric bipolar disorder, but they do suggest that pre-pubertal onset is rare. This assertion is strengthened as the included studies used samples from both Europe and North America, and is concordant with a recent meta-analysis reporting no differences in rates of youth BD between North American and European samples. 47 The lack of support for childhood onset may reflect the low diagnostic stability associated with very-early-onset BD. Evidence from longitudinal studies of high-risk offspring suggests that manic-like symptoms in very young children without a confirmed history of BD are not predictive of a later BD diagnosis. 48 Additionally, epidemiological findings suggest that individuals diagnosed with BD-NOS in childhood do go on to receive an adult BD diagnosis. 49,50 As our included studies assessed AAO retrospectively in adult samples with a confirmed BD diagnosis, any individuals who received a diagnosis of childhood BD which did not persist into adulthood will have been overlooked.

| Treatment and diagnosis
A corollary to forming a more robust definition of 'early-onset' BD is that clinical trajectory can be better anticipated, as early-life-onset is thought to confer a more severe and remitting course. 2,3 For instance, early-onset BD is associated with comorbid anxiety disorders and substance abuse 2,51 ; clinicians should be mindful of this when assessing and treating early-onset patients. Demarcating these AAO groups, thus, has implications for treatment provision, with the potential to guide appropriate junctures for intervention across the lifespan.
We found no substantive impact of diagnosis (BDI vs. BDII) on AAO distribution. However, most of the included studies (67%) recruited samples of BDI participants only (97% of total sample), while only 25% of studies included samples with BDII participants (3% of total sample). In light of this disparity, our findings must be interpreted with caution, especially as prior meta-analytic research found evidence partially supporting an earlier AAO in BDI compared to BDII patients.
As hypomanic episodes are not always recognised clinically, the true AAO in BDII is likely to be more difficult to determine and less reliable than BDI. Prospective follow-up of youth at high-risk of BD may therefore be the most reliable way to measure AAO of BDII (and/or BD-NOS). Such longitudinal monitoring would facilitate early identification of incipient or sub-threshold (hypo)manic and depressive symptoms.

| Late life onset
In contrast to the finding that 45% of cases onset in the 'early' subgroup, only 20% of cases were deemed 'late-life-onset' (>40 years of age; Figure 2). This indicates that these two subgroups may be aetiologically distinct forms of the same disorder, as suggested by prior research. 37

| Putative mechanisms
Our results indicate that a three-component model (early, mid-and late onset) best describes the AAO distribution of bipolar disorder.
As with the majority of psychiatric disorders, the interaction between genes and environment is likely to underpin the manifestation of this trimodal distribution in bipolar disorder AAO.
There is strong evidence for a genetic predisposition in bipolar disorder, with heritability estimates ranging from 60% to 85%, but the influence of genetics on AAO in bipolar is comparatively under-studied. 53,54 Initial evidence suggests that there is genetic homogeneity within AAO subgroups and heterogeneity between groups. [55][56][57] It is thought that early onset may be a more heritable form of BD than late onset, with studies demonstrating differences in transmission patterns and more pronounced familial aggregation in early-compared to late-onset BD. 4,6,8,54,57,58 Genetics does not explain the whole picture, however, and there are environmental and neurobiological factors that are thought to interact with various susceptibility genes to influence the AAO of BD. 59 It is thought that exposure to childhood trauma interacts with genes that are involved in pathways relating to neuroplasticity, inflammation and calcium signalling to influence AAO. [60][61][62][63] Epigenetic modifications in gene function may play an important role in the mechanism underlying the relationship between childhood trauma and a younger AAO of BD. 64 it has been suggested that immune system dysregulation, puerperal hormones and genetic factors may activate disease pathways. 75,77 Late-onset bipolar disorder is associated with increased rates of cerebrovascular disease, more medical and psychiatric comorbidities, and a weaker family history of psychiatric problems. 78,79 However, without employing detailed prospective longitudinal methodologies, it is unclear whether all of these environmental factors are a cause or consequence (or both) of incipient BD.

| S TRENG TH S AND LIMITATI ON S
This is the only known systematic review investigating age at onset in bipolar disorder. To ensure we captured all relevant studies on bipolar disorder AAO, we used a search strategy with broad criteria and included several different databases and grey literature searches. We were unable to assess risk of bias due to the broad range of reporting standards and methodologies used in our included studies.
Several limitations must be considered when interpreting our findings. It has been suggested that admixture analysis is sensitive to the sample size and the characteristics of the data. 80 The studies that reported a bimodal AAO distribution had smaller sample sizes on average compared to those reporting a trimodal AAO distribution. Interestingly, both cohort studies found a bimodal AAO distribution when they included birth cohorts in their models, but a trimodal AAO distribution when analysing the whole sample. This may be due to the fact that including birth cohorts in AAO analysis can artificially truncate the data, making the results of admixture analysis unreliable.
Inter-study differences in findings may further be attributed to the inconsistency in the definitions used for AAO of BD, as research has suggested that admixture analysis is further sensitive to the criterion used to define groups. 80 It has been proposed that the most valid definition for bipolar disorder AAO is the 'first affective episode meeting diagnostic criteria', as it does not preclude relevant episodes of depression prior to manic onset. 4 However, this does not overcome the limitation of recall bias, which was mitigated in some of the included studies by referring to case notes and interviews with family members rather than relying solely on self-report. Yet, BD patients may be more likely to recall depressive compared to manic episodes or even fail to recognise hypomanic episodes pre-diagnosis as pathological. 81,82 Deciding what constitutes pathology in retrospective studies is further distorted by the fact that potential symptoms in youth are viewed through the lens of an adult diagnosis. As a gold standard, therefore, future research investigating AAO in bipolar disorder should aim to employ prospective longitudinal methodologies, using the age at 'first affective episode meeting diagnostic criteria' as the standardised definition for the point of disease onset.
Results will also have been influenced by factors including interand intra-country differences in diagnostic practices, evolving diagnostic criteria, varying degrees of stigma surrounding mental illness and availability of healthcare provision. Yet, the fact that the clear majority of studies displayed a trimodal AAO despite this heterogeneity suggests that we can consider it a robust finding.

| FUTURE DIREC TIONS
Notably, one of the largest international BD cohorts-the Systematic Treatment Enhancement Program for Bipolar Disorder cohort (STEP-BD 83 )-was not included in this systematic review. This is because the STEP-BD articles identified by our search strategy used pre-defined cut-offs to define AAO groups (e.g. <13, 13-18, >18 years old) and, therefore, did not meet our eligibility criteria.
The field would benefit from a data-driven approach (such as admixture analysis) to defining BD AAO in this well-characterised cohort. Similar analyses in other large, phenotypically detailed cohorts should also be prioritised in future research (e.g. the Flourish Canadian prospective high-risk offspring cohort 84 ).

| CON CLUS ION
The results of this systematic review indicate that bipolar disorder has a trimodal age-at-onset distribution, segregating into early-, mid-and late-onset subgroups with the most common average age at onset being 17.3 years. We propose that the field conceptualises these subgroups as referring broadly to life stage and moves towards a consistent definition of bipolar AAO as 'the first affective episode meeting diagnostic criteria'. Providing valid evidence for three ageat-onset subgroups in BD will help to delineate more homogeneous subgroups of BD. Demarcating bipolar disorder AAO groups in this way can provide a framework for future research to continue to investigate potential mechanisms and, thus, inform treatment targets.

DATA AVA I L A B L I T Y S TAT E M E N T
The data that support the findings of this study are openly available