Early‐stage profiles of adolescent mental health difficulties and well‐being: A systematic review of cluster analyses in large school and community samples

Abstract

Traditional diagnostic and services pathways often overlook the nuanced ways that mental health problems and strengths appear in community settings. Some researchers have therefore used person‐centered statistics—or clustering analyses—to identify profiles of socioemotional and behavioral difficulties and well‐being traits in preclinical settings such as schools and communities. The objective of this review was to synthesize common adolescent mental health profiles within the literature and examine the state of the science. A systematic review of the literature was completed. Only studies that assessed multiple types of difficulties and/or strengths across community and/or school samples were included. A total of 3960 studies were screened, and k = 13 were included. Data extraction focused on the types of clusters in each included study, along with associated information like standardized scores, qualitative descriptions, sample size, and demographic characteristics. Data were integrated using a narrative synthesis, and meta‐analysis was used to investigate the prevalence of each cluster. Data were reviewed from n = 103,098 adolescents in 10 countries across 3 World Health Organization (WHO) Regions with a mean (SD) age of 13.72 ± 1.76 years. A total of 59 clusters were identified, and these consisted of 6 main cluster types. The 4 most prevalent cluster types described patterns of Flourishing, Moderate Mental Health, Struggling, and Mental Health Problems. The other 2 cluster types were less prevalent and associated with Languishing or Asymptomatic presentations. Cluster types also differed with respect to the intensity, complexity, and depth of the core profile features. This review describes commonly identified mental health profiles in large representative samples of adolescents. The quality of included studies was generally acceptable, but the literature suffers from variance in how clusters are generated and how mental health is assessed. Overall, this review can guide the development of new classifications of youth mental health and inform early‐stage intervention approaches in community settings.

INTRODUCTION

Adolescence is a critical developmental window for addressing emerging mental health difficulties—yet this understanding does not always translate into clinical practice. A large representative sample of more than 9000 US adults found that around 50% of psychiatric disorders related to mood and emotion begin before age 14 years (Kessler et al., 2005). A review of the WHO World Mental Health Survey also found that the onset of many adult mental health problems can be traced back to childhood and adolescence (Kessler et al., 2007). Despite these patterns of early onset, delays in accessing supports were also evident. Interventions often begin in adulthood once symptoms reach extreme levels of severity (Ford, 2008; Kessler et al., 2007). This finding is consistent with reactive treatment practices that still dominate mental health services. There are, however, clear advantages to adopting early‐stage perspectives on mental health—ones that focus on the experiences of adolescents and the role of healthcare solutions that mitigate future problems. Indeed, the recent 2024 Lancet Psychiatry Commission on Youth Mental Health urgently calls for high‐quality and effective prevention approaches to counter the rising global incidences of youth mental health problems (McGorry et al., 2024).

An adolescent‐centric perspective on mental health requires models and taxonomies that capture the experiences of this stage. However, the compendia of classification systems like the Diagnostic and Statistical Manual (American Psychiatric Association, 2013) and International Classification of Diseases (World Health Organization, 2019) are biased toward one developmental and clinical extreme. They are largely built around findings from psychiatric samples of adults with complex clinical presentations (Colligan, 1985; First et al., 2015; Frances et al., 1989; Regier et al., 2009). Our definitions of mental health (and the services they inform) can therefore struggle to appreciate the full scope of adolescent experiences. For example, there exist large groups of young people whose needs go underserved until problems escalate and a psychiatric diagnosis is required (Buckman et al., 2018; Compton & Shim, 2020; Paksarian et al., 2016; Shah, 2015). This systematic review therefore aimed to synthesize evidence from the literature and identify adolescent mental health and well‐being profiles in preclinical settings. These findings can inform the creation of developmentally appropriate taxonomies and guide early‐stage supports and interventions.

Clinicians and researchers have a long‐standing interest in improving the conceptualization of mental health and how distinct profiles are differentiated (Hayes et al., 1996). The hope is to find a classification approach that can represent the varying dimensions of problems—from low to high severity—(Kotov et al., 2017) and the patterns of well‐being and strength that occur in the general population (Keyes, 2002), while providing a valid typology that allows healthcare providers to streamline services by grouping individuals together (Everitt et al., 1971). This research has gained renewed momentum in recent years with the increased availability of classification analyses. This refers to statistical tools that divide a large sample into sub‐communities, or clusters, of shared traits. Popular methods include latent class analysis (which identifies unobserved “taxa” that multidimensional datasets) and k‐means clustering (which partitions individuals into “k” number of subgroups based on their similarity).

Statistical clustering techniques have been described in detail elsewhere (Weller et al., 2020; Zakharov, 2016). However, they are essentially “person‐centered” statistical models that capture the relationship between individuals. This approach has promise within mental health services, which are inherently concerned with the ways individuals are connected (Everitt et al., 1971); if a person with one mental profile benefited from a particular intervention, then we hope that this approach would help others with a similar profile. Recent years have seen an increase in the number of studies applying classification analyses to explore profiles of psychological strengths and difficulties within the general population (Fonseca‐Pedrero et al., 2020; Forbes et al., 2023; Ortuño‐Sierra et al., 2017; Petersen et al., 2019).

One commonly cited example of where cluster analysis has been used during preclinical stages comes from Bradshaw and Tipping (2010). They applied k‐means clustering to a large community sample of over 5000 children aged 2–5 years. This analysis identified five distinct subgroups based on mental health traits. The two largest clusters were characterized by above‐average prosocial tendencies and low rates of emotional, peer, and conduct difficulties—although one of these groups also experienced heightened hyperactive/inattention difficulties. The next two clusters were characterized by below‐average prosocial tendencies—one of which also experienced heightened emotional, conduct, and peer relations difficulties. The smallest clusters were characterized by elevated difficulties in conduct, peer relationships, emotional functioning, and hyperactivity/inattention, while maintaining relatively intact prosocial tendencies. This study provides a clear example of how clustering techniques can distil discrete profiles from complex, multidimensional datasets. It also highlights the importance of selecting measurements that represent the full range of mental health experiences, consistent with Keyes' dual‐continua model of mental health (Keyes, 2002, 2014). It is necessary to capture not only symptoms and problems (such as behavioral and socioemotional symptoms) but also strengths and traits associated with mental well‐being (such as prosocial tendencies).

Although Bradshaw and Tipping (2010) sampled school‐aged children, the potential for person‐centered approaches across development is evident. Many researchers have since applied similar statistics to adolescence—a time when persistent symptoms become more frequent. For example, one study explored mental health profiles based on data collected via the Danish National Youth Study (Andersen et al., 2021). A mixture of variables capturing psychological strengths (e.g., life satisfaction, well‐being and self‐esteem) and problems (e.g., stress, irritability, low mood, and loneliness) were clustered using latent class analysis, and a four‐profile model was identified. Findings suggested adolescents could be best organized along a continuum from flourishing clusters (with profiles of high well‐being and low problems) to emotionally challenged clusters (with profiles of elevated irritability, nervousness, and feelings of sadness) with clusters of moderate mental health and languishing nested between.

The current study

Recent years have seen a rise in the availability of person‐centered methods in youth mental health (Petersen et al., 2019). These techniques have the potential to support new approaches toward adolescent mental health; ones that are sensitive to the full range of early‐stage mental health strengths and weaknesses commonly seen in adolescence. This systematic review therefore explored common profiles by integrating findings from this literature.

The primary aim was to synthesize evidence on how adolescents cluster together based on profiles of psychological difficulties and strengths. The review focused on studies that recruited adolescents and applied different types of cluster analyses. This is because adolescence is a key developmental period for the onset of early‐stage mental health challenges. Only studies that sampled community and/or school‐based adolescent populations were included. The rationale here was twofold. First, such samples increase the generalizability of findings to wider adolescent populations. Second, these contexts are enriched for early onset and subthreshold mental health problems and less biased toward clinically severe presentations in clinical samples. A secondary aim was to examine methodological trends across the literature—such as the clustering techniques used, the nature of input variables, and how strengths and difficulties were operationalized and measured. This information is important to support future research efforts.

A narrative synthesis, guided by the framework of Popay and colleagues, was used to integrate the different types of clusters identified across studies (Popay et al., 2006; Popay & Mallinson, 2010). This method is well‐suited to synthesizing findings from studies that vary in measurement tools, sample sizes, and criteria used to define and distinguish mental health. Specifically, a narrative synthesis was deemed appropriate given our aim of integrating conceptual information about the various mental health profiles identified during adolescence. A meta‐analysis was not feasible as our primary method since clustering studies do not produce statistical effect sizes or standardized tests of group differences. However, meta‐analyses were used to estimate the prevalence of mental health profiles across adolescent samples.

The study was preregistered on PROSPERO, including the review questions, eligibility criteria, and data extraction and synthesis plan. The guiding research questions were: (1) What shared clusters of socioemotional and behavioral difficulties and strengths have been identified using classification analyses in community and school‐based adolescent samples?; and (2) What are the methodological trends across studies—for example, the prevalent clustering approaches and input inventories used?

METHOD

Study design

The present systematic review was informed by the recommendations of the Preferred Reporting Items for Systematic Reviews and Meta‐Analysis (PRISMA. See Figure 1 for PRISMA flow chart).

FIGURE 1.

PRISMA diagram. This figure illustrates the number of studies identified, screened, excluded, and included at each stage of the systematic review.

Eligibility criteria

Eligibility criteria are detailed in Table S1. Studies were included if they (1): were cross‐sectional or single‐wave of a longitudinal study that (2) generated data‐driven clusters using person‐centered statistics. Studies were excluded if they calculated profiles based on life experiences rather than mental health experiences (e.g., trajectories of difficulties associated with abuse, significant events like COVID‐19 or terrorist incidents). Studies were excluded if their statistical approach was variable‐centered, for example, ANOVA, correlation analysis, t‐tests, or regression. Only published studies in peer‐reviewed journals were included. Preprints were excluded as were any book chapters, conference symposiums, reviews, or thesis dissertations.

The third inclusion criterion (3) was that studies recruited a sample of adolescent participants aged 11–19 years old. This age range was chosen over alternatives like neurological adolescence because it closely matches the definition of adolescence adopted in health care settings in the UK and Ireland. If the minimum and/or maximum age fell outside this range (e.g., 8–18 years or 15–20 years), then the study was only included if the mean fell within the age of 11–18 years old. Another inclusion criterion (4) was that participants were recruited from community or school‐based samples. This was to increase the likelihood that clusters were drawn from appropriately powered and representative samples of adolescents enriched for preclinical difficulties. This means a study was deemed ineligible if it reported a convenience sample or no details on the recruitment protocol were provided. Studies were also excluded if they recruited from infants, children, and/or adults, from primary, secondary, or tertiary health care settings, as well as forensic, juvenile, or other specialist services. Studies sampling target groups were also excluded—this was done to optimize the generalizability of findings to general adolescent samples (e.g., adolescents with a specific shared experience—such as trauma or bullying).

Studies were included that used: (5) two or more measures of socioemotional and behavioral difficulties, interpersonal problems, and/or mental health symptoms; or (6) two or more measures of psychological well‐being and/or strengths; or (7) a combination of both well‐being and mental ill‐health assessments. This ensured that the clusters would capture the full spectrum of mental health—from wellness through to symptoms of mental health difficulties. Finally, studies were only included if they (8) used continuous measurements that were standardized with available psychometric norms that were (9) self‐reported by adolescents only. This was because of the validity of self‐report data over parent and/or teacher reports (Cantwell et al., 1997). Studies were excluded if they only included a single symptom dimension; that is, if they described how adolescents might cluster based on a single mental health difficulty and/or strength. Also excluded were studies reliant only upon measures unrelated to common psychiatric symptoms, such as data on personality, temperament, attitudes and opinions, cognitive task performance, any biometric data, and emotion regulation techniques like (problem‐solving, acceptance, mindfulness).

Search strategy

Three databases were electronically searched for included studies on June 17, 2023. An updated search was conducted on June 6, 2024, leading to the inclusion of one additional study. The databases included were Web of Science, PsycINFO, and PubMed. These were selected for their relevance, scope, and broad accessibility. This combination of databases was chosen to provide a comprehensive overview of peer‐reviewed literature across relevant disciplines. PsycINFO allowed us to capture literature from the psychological, behavioral, and developmental sciences. PubMed provided extensive access to biomedical, health science, and psychiatry research. Finally, Web of Science extended our search to interdisciplinary, humanities, and social science journals beyond the scope of the other two databases. The citation sections of all included studies were also reviewed to identify other eligible studies. There were no geographical, language, and publication date/type restrictions. The search string included terms related to person‐centered statistics like clustering, profiles, and unsupervised machine learning; mental health and specific terms like mood, emotion, anxiety, and symptoms; psychological well‐being; and adolescents, youth, and childhood. The search strings are provided in Table S2; these were developed and refined across three rounds of piloting to give a rich spread of search hits; for example, hits that capture different traits associated with mental health strengths and difficulties, as well as different statistical clustering techniques.

Screening, eligibility assessment, and data extraction

This systematic review was conducted across multiple steps from piloting, screening, full‐text review, data extraction, and through to quality assessment. This work was guided by PRISMA standards. A full PRISMA flow diagram outlining the number of records screened, excluded, and retained is provided in Figure 1.

Identification

Various search string combinations were initially piloted to ensure a robust search strategy that met our goals. On the penultimate search, two rounds of pilot screening were conducted by three authors using approximately 30% of randomly selected abstracts. This step aimed to develop a shared understanding and consistent approach to our inclusion and exclusion criteria. The final search of the systematic review was then completed (July 17, 2023), and a process of screening, extraction, and synthesis began. Throughout this process, authors met regularly to discuss conflicts and disagreements. An updated search was conducted on June 6, 2024, leading to the inclusion of one additional study.

Screening

After the final database search, k = 9835 records were imported into Covidence, and 6145 duplicates were removed. k = 3690 titles and abstracts were then screened for relevance. The first author screened all titles and abstracts. A total of 20% were independently screened by a second author. Double‐screened studies were randomly selected in Covidence. Any conflicts were resolved through discussion with a third author.

Eligibility

Next, the full text of k = 42 studies was downloaded and assessed against our full inclusion/exclusion criteria. The first author reviewed each of these studies, and 20% were once again independently assessed by a second author. Like before, conflicts were resolved through discussion with a third author. The reference sections of any eligible studies that passed full‐text screening were reviewed to identify other potential studies. In total, k = 13 studies passed the inclusion/exclusion criteria and moved forward for data extraction and quality assessment.

Included

For each included study, we extracted the mental health profiles identified using clustering analyses. The first author completed data extraction of each of these studies, and 7 were independently completed by a second author. Conflicts were resolved through collaborative discussion along with a third author. Extracted information included the label designated by the authors of a study to the cluster of adolescents, any qualitative descriptions provided, and available statistical or quantitative characteristics. Relevant quantitative characteristics included the type of clustering analysis applied, mean scores for each psychometric inventory within a cluster, the cluster size, and/or proportion estimates relative to the study sample included in its analysis. As an example of the type of findings extracted from included studies, Kim et al. (2019) reported on a “flourishing” cluster, described as showing “the highest strengths and lowest distress,” comprising 119 adolescents out of a total sample of 848 in their final analysis. The above data were extracted for every cluster reported within each selected study.

Extracted data were compiled into a master dataset of cluster names, descriptive summaries, and associated traits for each profile from each included study. Data extraction was managed using Covidence, using a purpose‐made data extraction template. This master dataset then fed into a narrative synthesis phase.

Quality assessment

The second research question related to the state of the science and trends within the literature. A quality assessment was therefore essential, but none, as far as the authors are aware, exist for the systematic review of person‐centered approaches (for review see Ma et al., 2020). A Quality Assessment Tool was therefore created using an extant one for predictive machine learning—the Probast tool (Moons et al., 2019; Wolff et al., 2019)—as well as the NIH Quality Assessment Tool for Observational Cohort and Cross‐Sectional Studies (NIH, 2013). Additional items were included based on available reviews on the use of machine learning and clustering algorithms in mental health research (Dalmaijer et al., 2022; Gillan & Whelan, 2017). This quality assessment was independently completed for all included studies by the first and second authors.

Data analysis

Narrative synthesis of mental health profiles

This review aimed to identify common mental health profiles during adolescence, based on representative community and school‐based samples. The clusters of adolescents (identified during data extraction) were grouped based on patterns in the language and themes used to describe their mental health profile. That is, mental health profiles were grouped into “cluster types” based on qualitative descriptions and interpretations made by each study's authors.

To synthesize the extracted clusters, a narrative synthesis was applied. This approach was: (1) grounded in a thematic integration of cluster labels and descriptions; (2) prioritized over more quantitative methods due to the substantial heterogeneity in psychometric measures and clustering analyses used (noted during an initial piloting phase); and (3) done in accordance with guidance from Popay et al. (2006). The process was as follows: To begin, all extracted clusters were coded according to the labels provided by study authors (e.g., “flourishing,” “high anxiety with low behavioural concerns,” “mental health”). Where labels were not available—qualitative descriptions were leveraged (e.g., “low symptom reports across different dimensions and high self‐reported well‐being traits”) and a descriptive code was then assigned de novo. Across successive rounds of coding, similar or overlapping labels and descriptors were grouped into higher‐order codes that captured broad cluster types (e.g., “flourishing,” “thriving,” “mental health” would be grouped into one cluster type called “Flourishing”). This iterative process continued until it was no longer possible to concatenate groups without compromising the meaningful distinctions between the cluster types.

This process allowed the identification of commonly recurring cluster types within the literature. Each cluster type captured a profile of mental health strengths and difficulties that emerged across independent samples in the literature. The narrative synthesis was managed using purpose‐built MATLAB scripts. Additional themes identified during this synthesis were documented and noted for discussion, consistent with a narrative synthesis approach.

Meta‐analysis of proportion estimates

Data extraction provided information on the number of adolescents placed in each cluster relative to the overall sample included in a study's analysis—we refer to this as a “proportion estimate.” If, for example, a cluster comprised of 200 adolescents and the analysis was calculated using a sample of n = 1000, then the proportion estimate was 0.2. These estimates allowed us to investigate the prevalence of cluster types within adolescent school and community‐based samples. To accomplish this, we first extracted the number of adolescents in each cluster and the total number of adolescents included in the cluster analysis. This was completed for each cluster in each study, giving a dataset of raw proportion estimates for clusters across all included studies. Next, these raw proportion estimates were pooled according to the higher‐order cluster types as defined during the narrative synthesis; that is, proportion estimates from similar cluster types were aggregated.

To achieve this, a meta‐analysis of proportion estimates for each cluster type was performed using the metafor, meta, and dmetar packages in R. Models were calculated for each cluster type. Proportion estimates for each cluster were logit‐transformed for variance stabilization. Random‐effects models were fitted using restricted maximum likelihood (REML), and pooled estimates were back‐transformed using the inverse logit function to aid interpretation along the original proportion scale. Heterogeneity was assessed using the I ², τ ², and Q statistics. The final pooled estimates provide an indication of the relative prevalence of each cluster type across studies.

RESULTS

Study selection

Thirteen included studies from 3690 search results were identified across two searches (June 2023 and June 2024; Figure 1 for PRISMA flowchart of each review process stage). The agreement between the co‐authors was acceptable at each review stage (percentage of agreement = 96% and Cohen's kappa = 0.73 at title/abstract screening; percentage of agreement = 100% and Cohen's kappa = 1 at full‐text review).

Study characteristics

Sample

Thirteen included studies were identified with a sample of N = 103,098 adolescents. Table 1 provides an overview of these 13 studies as well as of their methodology and findings. Eleven studies recruited via school‐based sampling, including a total of 352 schools. Samples were also drawn from 10 countries, capturing 3 out of 6 World Health Organization regions, illustrated in Figure S1. This included the European, Western Pacific, and Americas regions. The WHO regions not represented were the African Region, South‐East Asian Region, and Eastern Mediterranean Region. The mean ± SD percentage of females (relative to males) across these studies was 53.94 ± 4.81% (95% CI = 49.98–56.10); none reported nonbinary gender identities. The mean age was 13.72 ± 1.76 years, and the range was 6–19 years. Age data are presented in Figure S2.

TABLE 1.

Overview of the key findings from eligible classification studies.

Study ID	Location	Cohort name	Sample (n schools; grade)	N (% female)	Age M (SD)	Person‐centered analysis type (PCA)	Input variables (socioemotional and behavioral problems; SEB)	Input variables (psychological strengths and well‐being)
					Range
Moore et al. (2019)	USA	–	Schools sample (n = 1; Grades 9–12)	963 (50.4%)	NR	Latent Class	Externalizing, Internalizing	Belief‐in‐self (e.g., self‐efficacy), belief‐in‐others (e.g., peer support), emotional competence and engaged living
Yang & McGinle (2023)	Taiwan	–	Schools sample (NR; Grades 7 or 10)	1473 (57.9%)	14.87 (1.47) 12–19 years	Latent Class	Impulsivity/self‐control, interpersonal problems	Prosocial behavior
Kim et al. (2019)	South Korea	–	Schools sample (n = 6; Grades 4–6)	1530 (51%)	14 (0.81) NR	Latent Class	Behavioral/conduct, mood/affect difficulties	Life Satisfaction, school‐specific well‐being
Gustafsson et al. (2023)	Finland	Health behavior in school‐aged children study	Schools sample (NR; Grades 5, 7 and 9)	3149 (49.9%)	13.44 (1.69) NR	Two‐step clustering	Psycho‐somatic symptoms, loneliness, psychological complaints	Life satisfaction
Andersen et al. (2021)	Denmark	Danish National Youth Study	Schools sample (n = 119; Upper secondary)	60,612 (61.76%)	17.5 (1.15) NR	Latent Class	Anxiety, stress, mood/affect, loneliness	Self‐esteem, self‐efficacy, autonomy
Cannon & Weems (2006)	USA	–	Community sample	225 (50.76%)	11.5 (NR) 6–17 years	K‐means	Depression, anxiety	No well‐being measures
Noel et al. (2013)	New Zealand	–	Schools sample (n = 143; Year 9–13)	9107 (46%)	14.8 (NR) 12–19 years	Latent Class	Depression, behavior/conduct, substance‐use/addiction, suicidal behavior, unsafe sex and risky behavior	No well‐being measures
Arbeit et al. (2014)	USA	4‐H Study E10:E11 of Positive Youth Development	Schools sample (n = 61; Grades 6–12)	4391 (46%)	12.1 (0.61) NR	Latent Class	Depression, substance‐use/addiction, restrictive eating, unsafe sex	No well‐being measures
Riglin et al. (2016)	UK	School Transition and Adjustment Research Study	Schools sample (n = 10; NR)	1648 (NR)	NR 11–12 years	Latent Class	Depression, behavior/conduct	No well‐being measures
Galle‐Tessonneau et al. (2019)	France	–	Schools sample (n = 6; Second grade)	469 (59.9%)	12.1 (1.1) 10–16 years	Latent Class	Internalizing, externalizing, school absenteeism	No well‐being measures
Picoito et al. (2021)	UK	Millennium Cohort Study	Population survey	17,216 (48.86%)	11 and 14 years	Latent Class	Behavior/conduct, hyperactivity, inattention, mood/affect; interpersonal problems	No well‐being measures
Zhou et al. (2020)	China	Project for the Longitudinal Development of Chinese Adolescents' Mental Health	Schools sample (n = 4; Grade 7)	1009 (50.7%)	12.97 (0.67) 10–15 years	Latent Class	Depression, anxiety, school stress	Life satisfaction, self‐esteem, satisfaction in school
Sun et al. (2023)	China	–	School sample (n = 2)	1306 (53.4%)	13.98 (1.19) 10–18 years	Latent Class	Depression, anxiety	No well‐being measures

Person‐centered analyses

Across the 13 included studies, 23 person‐centered analyses were calculated (Table 2). K = 7 studies calculated a single model across the entire adolescent sample (Cannon and Weems, 2006, Galle‐Tessonneau et al. 2019, Gustafsson et al., 2023; Noel et al., 2013, Riglin et al., 2016; Sun et al., 2023; Yang & McGinley 2023) (giving 7 analyses); K = 4 studies split their samples according to age/grade, running separate models for each (giving 14 analyses) (Arbeit et al., 2014; Moore et al., 2019; Picoito et al., 2021; Zhou et al., 2020); K = 2 studies split the sample by sex, running separate models for each (giving 2 analyses) (Andersen et al., 2021; Kim et al., 2019). Studies included a mix of clustering approaches, with latent class analysis being the most common (K = 10), followed by k‐means (n = 1; Cannon & Weems 2006) and two‐step clustering (K = 1; Gustafsson et al., 2023).

TABLE 2.

Synthesis of individual clusters identified in each of the eligible studies.

	Cluster name (#)	Within‐study level	Cluster description	Cluster size (n)	% of sample	Cluster type
Moore 2019	Complete mental health (1)	Grade 9	High well‐being and low distress	107	23.5	Flourishing
		Grade 10		220	25.5
		Grade 11		161	21.0
		Grade 12		113	36.9
	Moderately mentally healthy (2)	Grade 9	High‐average well‐being and low distress	213	46.7	Moderate mental health
		Grade 10		422	49.0
		Grade 11		338	44.0
		Grade 12		103	33.7
	Symptomatic but content (3)	Grade 9	Average to high‐average well‐being and average to above‐average distress	91	19.9	Struggling
		Grade 10		162	18.8
		Grade 11		221	28.8
		Grade 12		65	21.2
	Troubled (4)	Grade 9	Average to low‐average well‐being and above‐average distress	45	9.8	Mental health problems
		Grade 10		58	6.7
		Grade 11		48	6.2
		Grade 12		25	8.3
Yang & McGinley 2023	High self‐control and high in all attachment relationships (5)	Whole sample	High scores in self‐control as well as in parent, teacher, and peer attachments	618	42.0	Flourishing
	Low self‐control and high peer attachment (6)		Low levels of self‐control, high levels of peer attachment, and low to moderate levels of parent and teacher attachments	190	12.9	Moderate mental health
	Low self‐control and low in all attachment relationships (7)		Low levels of self‐control and also low scores in parent, teacher, and peer attachments	65	4.4	Mental health problems
	Low self‐control and moderate in attachment relationships (8)		Low levels of self‐control, and moderate levels of parent, teacher, and peer attachments	600	40.7	Struggling
Kim 2019	Flourishing (9)	Males	Highest strengths and lowest distress	119	14.0	Flourishing
		Females		148	17.0
	Languishing (10)	Males	Low strengths and high distress	147	17.0	Mental health problems
		Females		69	8.0
	Moderate flourishing (11)	Males	Moderate strengths and moderate distress	242	29.0	Moderate mental health
		Females		339	38.0
	Moderate languishing (12)	Males	Low strengths and moderate distress	340	40.0	Struggling
		Females		339	38.0
Gustafsson 2023	Good mental health (13)	Whole sample	Low prevalence of psychological and somatic complaints. They were highly satisfied with their lives and the majority were normative social media users. All adolescents in this profile reported low loneliness	1375	44.0	Flourishing
	Mixed psychosocial health (14)		Life satisfaction was moderate. The highest percentage of adolescents reporting one to two frequent psychological complaints whereas the prevalence of somatic complaints was low. One fourth reported high loneliness and nearly half were risky social media users	628	20.0	Moderate mental health
	Poor mental health (15)		Frequent psychological or somatic complaints. The highest percentage of adolescents reporting high loneliness, the lowest mean value of life satisfaction, and the highest percentage of problematic social media users	675	21.0	Mental health problems
	Somatically challenged (16)		Frequent somatic complaints. Around half reported low prevalence of psychological complaints and the rest reported one to two frequent psychological complaints. Their mean life satisfaction was 7.92 (sample M ± SD = 7.72 ± 1.81), and all reported low loneliness. The majority were normative social media users, and the rest were risky social media users	471	15.0	Struggling
Andersen 2021	Emotionally challenged (17)	Females	Regular emotional symptoms but otherwise scored relatively high on the positive mental health indicators. Elevated probabilities of irritability, nervousness and seven out of 10 females reported sadness on a weekly basis	10,581	28.0	Struggling
		Males		3477	15.0
	Flourishing (18)	Males	High on all positive mental health indicators and had low probabilities of frequent occurring emotional symptoms	12,749	55.0	Flourishing
		Females		14,224	38.0
	Languishing (19)	Females	Low probabilities of positive mental health and high probabilities of negative mental health across all indicators. The probability of high life satisfaction was less than 5%, about one third of the students reported weekly stress and the students had very elevated probabilities of loneliness and feeling sad	7112	19.0	Mental health problems
		Males		2318	10.0
	Moderate mentally healthy (20)	Females	Reduced probabilities of life satisfaction and self‐esteem. However, their scores on symptoms were only slightly elevated	5615	15.0	Moderate mental health
		Males		4636	20.0
Cannon& Weems 2006	Anxious (21)	Whole sample	NR	54	24.0	Struggling
	Comorbid (22)			32	1.3	Mental health problems
	Depressed (23)			37	16.4	Struggling
	Normal (24)			102	45.1	No symptoms
Noel 2013	Healthy group (25)	Whole sample	Risk behaviors and emotional health concerns among students in this group was low. The majority of students in this group identified no health concerns	7089	79.6	No symptoms
	Multiple (26)		High levels of both risk behaviors and emotional health concerns	319	3.6	Mental health problems
	Risky group (27)		High rates of risk behaviors but relatively low rates of emotional health concerns	963	10.8	Struggling
	Distressed youth (28)		High levels of depressive symptoms, and 48% had made a suicide attempt in the past 12 months	530	6.0	Mental health problems
Arbeit 2014	Alcohol and aggression (29)	Grade 6	Some kind of problematic behavior, with mid‐range probabilities of engaging in aggressive behaviors such as beating people up and also using alcohol and marijuana	NR	20.0	Struggling
		Grade 7		NR	16.0
		Grade 8		NR	21.0
		Grade 9		NR	21.0
		Grade 10		NR	15.0
		Grade 11		NR	9.0
		Grade 12		NR	21.0
	High drive for thinness (30)	Grade 9	Elevated drive for thinness accompanied by depressive symptoms and bulimic symptoms and body dissatisfaction scores increasing. Very low probabilities engaging in almost all problematic behaviors, except for drinking alcohol and having protected sex	NR	24.0	Struggling
		Grade 10		NR	21.0
		Grade 11		NR	27.0
		Grade 12		NR	26.0
	High risk (31)	Grade 6	Highest probabilities of engaging in problematic behaviors. Youth in this profile at (Grade 6) were not quite as engaged with problematic behaviors across the board, except for smoking cigarettes, beating people up, and drinking alcohol	NR	3.0	Mental health problems
		Grade 7		NR	2.0
		Grade 8		NR	4.0
		Grade 9		NR	8.0
		Grade 10		NR	4.0
		Grade 11		NR	2.0
		Grade 12		NR	6.0
	Low risk (32)	Grade 6	Low probability of engaging in any of the problematic behaviors	NR	67.0	No symptoms
		Grade 7		NR	67.0
		Grade 8		NR	63.0
		Grade 9		NR	37.0
		Grade 10		NR	46.0
		Grade 11		NR	49.0
		Grade 12		NR	41.0
	Mental health and other risks (33)	Grade 7	Mental health struggles combined with a mix of problematic behaviors, although the specific pattern of behaviors varied from grade to grade	NR	6.0	Mental health problems
		Grade 8		NR	5.0
		Grade 9		NR	6.0
		Grade 10		NR	8.0
		Grade 11		NR	9.0
	Mental health struggles (34)	Grade 6	High levels of depressive symptoms. Levels of problematic behaviors were also slightly elevated. Evidence of disordered eating attitudes and behaviors was also strong. Low levels of problematic behaviors except for using tobacco and alcohol in Grade 12	NR	10.0	Mental health problems
		Grade 7		NR	9.0
		Grade 8		NR	7.0
		Grade 9		NR	4.0
		Grade 10		NR	4.0
		Grade 12		NR	6.0
Grade 11		NR		4.0
Riglin 2016	Comorbid (35)	Whole sample	Depressive symptoms, which met a clinical cut‐point with co‐occurring high symptoms of conduct problems	70	4.0	Mental health problems
	Conduct Problems A (36)		NR	70	4.0	Struggling
	Conduct Problems B (37)		NR	29	2.0	Struggling
	Depressive (38)		Depressive symptoms, which met a clinical cut‐point in the absence of conduct symptoms	49	3.0	Mental health problems
	Depressive comorbid (39)		High levels of depressive symptoms, scoring more than double the clinical cut‐points. In the school sample this was accompanied by co‐occurring moderate symptoms of conduct problems and therefore labeled depressive comorbid	22	1.0	Mental health problems
	Elevated (40)		NR	334	20.0	Struggling
	Moderate depressive (41)		Moderate levels of depressive symptoms in the absence of conduct symptoms	152	9.0	Struggling
	Normative (42)		Relatively low levels of all symptoms	922	56.0	No symptoms
Galle‐Tessonneau 2019	High problems (43)		Borderline level of internal problems and external problems, and higher levels compared with youths in the other clusters	155	33.0	Mental health problems
	High‐absence (44)		Highest amount of school absence. A level of internal and external problems within the norm	111	23.7	Struggling
	Low‐absence (45)		Marked by a lower amount of school absence from school compared to HA	57	12.2	Struggling
	Low‐absence‐Low‐problems (46)		Very low level of internal and external problems	146	31.1	No symptoms
Picoito 2021	High externalizing + High emotion (47)	Age 11	High externalizing + High emotion	487	3.8	Mental health problems
	High internalizing + Moderate emotion (48)	Age 14	High internalizing + Moderate emotion	351	3.1	Mental health problems
	Low symptoms (49)	Age 11	Low scores on the symptoms scales and high prosocial behavior	9305	72.6	Flourishing
	Low symptoms	Age 14		726	68.6
	Moderate emotion (50)	Age 14	Moderate emotion	726	6.4	Struggling
	Moderate externalizing (51)	Age 11	Moderate externalizing	1284	13.8	Struggling
		Age 14	Moderate externalizing	1859	16.4
	Moderate internalizing (52)	Age 14	Moderate internalizing	1168	10.3	Struggling
	Moderate peer problems (53)	Age 11	Moderate peer problems	1256	9.8	Struggling
Zhou 2020	Flourishing youth (54)	T1	Low depressive and anxiety symptoms, high self‐esteem and life satisfaction	518	51.2	Flourishing
		T2		356	40.0
		T3		238	36.8
	Troubled youth (55)	T1	High depressive and anxiety symptoms, low self‐esteem and life satisfaction	82	8.5	Mental health problems
		T2		117	13.4
		T3		85	13.9
	Vulnerable youth (56)	T1	Low depressive and anxiety symptoms, low self‐esteem and life satisfaction	409	40.3	Languishing
		T2		421	46.6
		T3		330	49.3
Sun 2023	Healthy group (57)	Whole sample	Fewest symptoms of depression and anxiety	805	61.4	No symptoms
	Anxiety disorder group (58)		Anxiety symptoms without clinical depression	309	23.9	Struggling
	Depression‐anxiety disorder group (59)		Severe self‐reported symptoms of depression and anxiety	192	14.7	Mental health problems

Measurements

Not all studies included measures of psychological well‐being. K = 5 collected measures of socioemotional, behavior, and/or interpersonal problems only (Arbiet et al., 2014; Cannon & Weems, 2006; Galle‐Tessonneau et al. 2019; Noel et al., 2013; Riglin et al., 2016). K = 7 collected measures of these difficulties along with indices of psychological well‐being and strengths (Andersen et al., 2021; Gustafsson et al., 2023; Kim et al., 2019; Moore et al., 2019; Picoito et al., 2021; Yang & McGinley 2023; Zhou et al., 2020) (Table 1). Common difficulties measured included (1) anxiety, depression, and/or stress, (2) interpersonal problems, (3) behavioral and conduct problems, and (4) mood and/or affective difficulty. In contrast, domains of well‐being were relatively fewer and more specific—with measures focusing on (1) self‐efficacy, self‐esteem, and/or autonomy, (2) prosocial behavior, (3) life satisfaction, and (4) mental well‐being. An overview of all the psychometric inventories identified in the literature is presented in Figure S3 and Table S3. Findings indicate marked variability in measurement approaches within classification research. The SDQ was the most common measure for both socioemotional and behavioral difficulties and strengths.

Adolescent mental health clusters

Across included studies, 59 mental ill health and well‐being clusters were identified. Table 2 provides an overview and a numeric identifier for each. The mean (SD) number of clusters identified by the given clustering analysis was 4.53 (±1.5; range = 3–8 clusters). There was no association between a study's sample size and the number of clusters identified (Spearman's rho = .4, p = .15). Although the samples were large, ranging from n = 225 to 60,612.

It is worth noting that some studies calculated more than one cluster analysis. In half of the studies, multiple clustering models were calculated by stratifying the samples (e.g., the overall sample might be split by age or sex) and running separate analyses for each subgroup (Table 2). This means some cluster profiles were repeated within the same study. One, for example, identified the same set of clusters after they split their sample by school grade (Arbeit et al., 2014). When counting each cluster analysis individually—a total of 118 clusters were identified across the included studies. However, many of these were overlapping due to the stratified subgroups. To avoid any double counting, we focus on the 59 nonrepeating cluster profiles.

Themes and clusters characteristics

One theme related to intensity of a cluster profile; that is, clusters varied with respect to the intensity of psychological difficulties and strengths. K = 2 studies provided standardized scores to describe the mental health profile of different clusters. Z‐scores provided an indication of how each cluster rated on the Childhood Depression Inventory and Revised Child Anxiety and Depression Scale (Cannon & Weems, 2006; Sun et al., 2023). The remaining studies provided qualitative description of each cluster (Table 2) (additional information provided in Table S4). Studies characterized clusters using descriptors like “low,” “moderate,” “elevated,” “frequent,” and/or “high” (e.g., elevated depression). Only K = 2 provided clear, operational definitions of these qualitative descriptors (Cannon & Weems, 2006; Noel et al., 2013). Based on the studies text and/or any available standardized scores, we coded the mental health profile of a cluster as being either “high,” “moderate,” or “low” intensity. Information on the intensity of characteristics was coded based on the studies language, available qualitative descriptors or any available standardized scores. For instance, clusters with mental health difficulties “high” or “above clinical cut off” were all coded as “high.” Clustered with difficulties described as “elevated,” “moderate,” “above average” were all coded as “moderate.” Clusters with difficulties described as “low” or “absent” were all coded as “low.” This theme is represented in Figure 2a.

FIGURE 2.

Topography of the adolescent mental health clusters across included studies. This figure provides a visual summary of the common cluster types and key themes from the narrative synthesis. It builds on data extracted from each study (Table S5) and was developed using MATLAB's scatterplot and scatter3 functions. A jitter (0.8) was applied to minimize overlap. Each dot represents a unique adolescent mental health profile identified using cluster analysis. The dot size represents the proportion of adolescents allocated to this cluster relative to the study sample. (a) Clusters are positioned based on the intensity of psychological difficulties (x‐axis) and strengths such as mental well‐being (y‐axis). Intensity was coded as low, moderate or high. (b) Granularity indicates whether clusters involved more general and broad domains (e.g., internalizing/externalizing) or deeper, more specific domains (e.g., self‐harm, substance‐use, self‐esteem). Granularity was coded as “broad” or “specific.” Another theme related to the complexity of presentations. This is noted by color. Some clusters involved isolated problems (blue) while others were characterized by widespread and co‐occurring problems (green). Other clusters displayed no difficulties (purple).

Another theme related to the breadth and specificity of the mental health difficulties associated with a cluster—here we refer to this as granularity. More granular clusters represented narrow and specific difficulties. Examples from included studies include patterns of unsafe sex, low negative mood, or marijuana use. Less granular clusters represented broader and widespread difficulties. Examples from included studies include patterns of heightened internalizing and/or externalizing. To synthesize these data, a cluster was coded as “broad” if general domains were captured or “specific” if narrow symptom concerns were captured. This theme is represented in Figure 2b.

A final theme that emerged was related to the co‐occurrence of difficulties within a cluster. Here we refer to this as a cluster “complexity.” In some clusters, adolescents were described by widespread challenges and elevated scores across multiple domains, for example, depression along with conduct problems. Other clusters were described by focused difficulties, isolated to specific domains, for example, depression in the absence of any other challenges (Riglin et al., 2016). To synthesize these data, a cluster was coded as “co‐occurring” if concurrent difficulties were documented or as “discrete” if isolated difficulties were documented. This theme is represented in Figure 2b. The names and codings for each cluster are provided in Table S5

Cluster types within the literature on adolescents

Flourishing clusters

Seven of the 59 clusters captured patterns of flourishing. Adolescents in this cluster type reported high psychological well‐being and strengths, while reports of socioemotional and behavioral difficulties were low. This flourishing type cohort included clusters #1, 5, 9, 13, 18, 49, and 54 (Table 2). For example, cluster #13 was described by high life satisfaction and low reports of psychological distress, somatic complaints, and loneliness (Gustafsson et al., 2023).

This cluster type was identified in 7 of the 13 included studies and observed a total of 15 times across these studies (Table 2). The average proportion of adolescents assigned to cluster type in a given sample was 25.19%–44.50% (95% CI; M ± SD = 34.84% ± 17.44%) with a mean of 52.06% being female (SD = 9.48). Relative to the full sample of n = 103,098 adolescents included in this reviewed, around 39.75% were classified into the flourishing cluster (n = 40,977).

Moderate mental health clusters

Five of the 59 clusters captured patterns of moderate mental health. This included clusters #2, 6, 11, 14, and 20 (Table 2). Adolescent reports of psychological well‐being and strengths were moderately high, while reports of socioemotional and behavioral difficulties were often elevated. This suggests a cluster of adolescents experiencing some difficulty but with enacted psychological well‐being and strengths. For example, cluster #6 captured adolescents experiencing good overall interpersonal relations but low self‐control (Yang & McGinley, 2023). Cluster #11 was characterized by moderate life satisfaction but elevated reports of psychological distress (Kim et al., 2019).

This cluster type was identified in 4 of the 13 included studies and observed a total of 10 times across these studies (Table 2). The average proportion of adolescents assigned to this cluster type in each sample was 21.22–40.44 (95% CI; M ± SD = 30.82 ± 13.43) with a mean of 63.3% female (SD = 10.18%). Considering the full sample of n = 103,098 adolescents, around 12.34% were classified into the moderate mental health cluster (n = 12,726).

Languishing and asymptomatic clusters

One study described a vulnerable cluster characterized by low self‐esteem and low socioemotional concerns (i.e., low depression and anxiety) (Zhou et al., 2020). This absence of positive mental health here might reflect a languishing‐type cluster. This was cluster #56, and it was observed 3 times in this study (see Table 2). Overall, the average proportion of adolescents assigned to this cluster in this sample was around 33.92–56.87% of this sample (95% CI) (M ± SD = 45.50 ± 4.61). Relative to the overall sample of n = 103,098 adolescents in this review, around 1.13% were classified into this languishing cluster (n = 1160)

Six of the 59 clusters captured patterns of symptom absence; that is, low reports of any socioemotional and behavioral difficulties. The asymptomatic cluster type included clusters #24, 25, 32, 42, 46, and 57. This cluster type was exclusive to studies not using measures of mental well‐being and strengths. Indeed, studies that excluded measures of positive mental health traits could only report on the mere absence of mental health difficulties. This cluster type was identified in 6 of the 13 included studies and observed a total of 15 times across these studies (Table 2). The average proportion of adolescents assigned to this cluster type in each sample was 43.24–65.51% (95% CI; M ± SD = 54.37 ± 17.51) with 50.34% female (SD = 6.51%). Considering the full sample of n = 103,098 adolescents, around 8.79% were classified into the asymptomatic cluster (n = 9046).

Struggling

The largest proportion of clusters captured a pattern of “struggling,” characterized by elevated socioemotional and behavioral difficulties (e.g., moderate levels of internalizing or externalizing). This included clusters #3, 8, 12, 16–17, 21, 23, 27, 29, 30, 36, 37, 40, 41, 44, 45, 50–53, and 58 (Table 2). One theme that emerged was around the discrete nature of these mental health difficulties wherein specific domains of mental health symptoms were impacted. For example, cluster #36–37 captured patterns of conduct and behavioral problems in the absence of any emotional difficulties (Riglin et al., 2016) while clusters #17, 21, 23, and 41 captured patterns of heightened anxiety or depression symptoms in the absence of other difficulties (Andersen et al., 2021; Cannon & Weems, 2006; Riglin et al., 2016). Also of note was the level of mental health strengths in this cluster. Specifically, psychological well‐being and strengths were somewhat intact, falling in a low‐moderate range. For example, cluster #3 described adolescents experiencing average well‐being with above‐average distress.

This cluster type was identified in 12 of the 13 included studies and observed a total of 20 times across these studies (Table 2). The average proportion of adolescents assigned to cluster type across the samples was 15.00–22.19% (95% CI; M ± SD = 18.60 ± 10.63) with a mean of 51.39% female (SD = 29.76). Considering the full sample of n = 103,098 adolescents included in this review, around 24.01% were classified into the struggling cluster (n = 24,756).

Mental health problems

The next largest portion of clusters captured clear patterns of mental health problems; these difficulties were characterized by co‐occurring socioemotional, behavioral, and/or interpersonal difficulties. In addition, the intensity of symptom reports was also comparatively higher than in the other clusters, with authors assigning terms like “very high,” “high,” “clinical,” and/or “severe.” This included clusters #4, 7, 10, 15, 19, 22, 26, 28, 31, 33, 34, 35, 38, 39, 43, 47, 48, 55, and 59 (Table 2). Across each of these clusters, psychological well‐being and positive mental health traits were consistently low or absent. This was therefore a cluster type characterized by problems in the absence of protective psychological reserve. For example, cluster #55 described adolescents experiencing high depressive and anxiety symptoms, low self‐esteem, and low life satisfaction (Zhou et al., 2020). A similar pattern was observed in clusters #10 and 19 (Andersen et al., 2021; Kim et al., 2019).

This cluster type was identified in each of the 13 included studies and observed a total of 42 times across these studies. The average proportion of adolescents assigned to this cluster type in a given sample was 6.85–13.31% (95% CI; M ± SD = 10.08 ± 7.46) with around 52.06% female (SD = 26.72%). Considering the full sample of n = 103,098 adolescents included in this review, around 12.66% were classified into the flourishing‐type cluster (n = 40,977).

Meta‐analyses

Pooling of proportion estimates

A meta‐analysis of proportion estimates for each cluster type was performed. Pooled proportion estimates for each cluster type are presented in Figures S4–S8. Heterogeneity was very high across studies, with statistical parameters presented in Table S6. This likely reflects differences in clustering methods as well as variability in psychometric inventories administered. It was therefore deemed inappropriate to pool proportion estimates and generate summary estimates.

Differences in prevalence of cluster type

A follow‐up analysis investigated whether prevalence estimates varied across cluster types; that is, were certain cluster types more or less common than others across included studies? A random‐effects meta‐regression was therefore calculated using cluster type as a categorical variable to predict logit‐transformed prevalence estimates. Proportion estimates differed across the cluster types, QM (df) = 74.1 (5), p < .0001. This effect was driven by the mental health problems (β = −1.75, t(df) = −5.96 (72), p < .001) and struggling (β = −0.97, t(df) = −3.34 (72), p < .001) cluster types, which were both associated with smaller proportion estimates. However, the heterogeneity across studies was high (I ² = 99.75%, Tau² (SE) = 0.68 (0.11)).

Differences in symptom complexity across cluster types

One theme identified was around the complexity of mental health challenges experienced. Some clusters were characterized by widespread “co‐occurring difficulties” across multiple symptom domains. Others were characterized by “discrete difficulties” that were isolated to specific symptom domains. We investigated how complexity varied across cluster types using chi‐squared tests. A Chi‐square test was calculated to examine the association between complexity and cluster type. The “mental health problems” cluster type was more likely to be complex (61.66%) than discrete (38.88%). In contrast, the “struggling” cluster type was more likely to be discrete (88.23%) than complex (11.76%), χ ² (df) = 10.03 (1), p = .002.

Quality assessment

The quality of studies was found to be acceptable—most studies fell above a score of 75% (Figure 3) (M ± SD = 72.30 ± 11.29%, range = 45–80%). K = 7 studies scored 80%; K = 3 studies scored 70–75%, and K = 4 studies scored 45–65%. Methodological strengths included the consistent use of standardized measures. Also, studies were appropriately powered to extract 4–5 clusters (Dalmaijer et al., 2022).

FIGURE 3.

Outcome of quality assessment. This figure illustrates the quality rating for each study based on a purpose‐made quality assessment tool for cluster analysis research studies.

Limitations were as follows. First, K = 0 reported both clear inclusion and exclusion criteria. This may relate to the use of large community and school‐based samples with minimal selection criteria. Second, K = 5 studies investigated cluster validity. K = 1 used clusters to prospectively predict school dropout (Andersen et al., 2021). K = 1 used clusters to predict independent measures of positive youth development (Arbeit et al., 2014). K = 2 used hold‐out samples to evaluate if clusters could be replicated. In one, a random proportion of the adolescents' data was set aside to internally validate adolescent mental health clusters (Cannon & Weems, 2006). In the other, a separate twins sample was used to externally validate the adolescent mental health cluster (Riglin et al., 2016). Finally, K = 1 used the clusters to predict performance on self‐rated executive function (Sun et al., 2023).

Finally, the impact of study quality on how adolescents were clustered was examined. We investigated whether there were systematic differences in how lower quality studies estimated the prevalence of different cluster types. To address this, included studies were first coded as “high” or “low” based on our quality assessment tool. “High” quality included studies scoring >74% on the quality assessment tool, and “low” quality included studies scoring <75% on the quality assessment. Next, a fixed‐effects meta‐regression tested whether proportion estimates differed by study quality. Findings indicated that study quality was not significantly associated with differences in proportion estimates or heterogeneity. Findings are presented in Table S7.

DISCUSSION

Youth mental health supports typically begin when problems become significantly intense and complex (Ford, 2008). However, difficulties emerge gradually, with sub‐clinical symptoms frequently predating more severe and impairing clinical presentations (Johnson & Wolke, 2013; Roberts et al., 2015). For example, one longitudinal study of over 1000 adolescents (aged 17–18 years) found that a cohort with sub‐clinical depression symptoms had similar long‐term outcomes to peers with a diagnosis of major depression disorder, including persistent symptoms and future suicidal behaviors (Fergusson et al., 2005). This suggests a disconnect between the developmental trajectories of mental ill health and the criteria for accessing care. Although problems evolve over time, many young people struggle to avail themselves of supports until the criteria for a psychiatric diagnosis are reached.

Reactive approaches toward mental health, built around strict thresholds, are less than optimal. Many young people will experience psychological distress without accessing effective supports (Jinnin et al., 2016; Smith et al., 2018). In England, around 26% of referrals to Child and Adolescent Mental Health Services (CAMHS) were rejected in 2018–2019 (Education Policy Institute, 2020). In Ireland, anywhere between 16% and 62% of CAMHS referrals are rejected (Mental Health Commission, 2023). Meanwhile, in Scotland, CAMHS rejection rates have increased from 17% to 30% since 2020 (Ball et al., 2023). Such referrals often reflect persistent socioemotional and behavioral concerns that become more complex across development. For example, re‐referrals amounted to more than a third of all CAMHS referrals in Denmark (Hansen et al., 2021). And once adolescents access support, the severity of distress can necessitate intensive intervention including in‐patient admissions and/or regular appointments with clinicians from a broad team of disciplines. This high‐intensity input can be extremely demanding, with some adolescents critiquing the lack of more acceptable mental health supports that begin in schools and community settings (Plaistow et al., 2014). Overall, the reactive approaches that dominate youth mental health services can obfuscate treatment access while exacerbating the underlying challenges.

In response, there have been recent appeals for proactive approaches in youth mental health—ones that mitigate long‐term problems by providing targeted supports during earlier stages of development (Uhlhaas et al., 2023). The 2024 Lancet Psychiatry Commission on Youth Mental Health, for instance, recognizes the need for early‐stage solutions like interventions that prevent and mitigate mental ill health (McGorry et al., 2024). However, a barrier to progress is a lack of models outlining the mental health needs among adolescents in preclinical settings—where challenges are not yet at peak intensity and complexity. We sought to address this gap. Through a systematic review of the literature, we synthesized evidence of how adolescents in large community and school‐based samples cluster based on profiles of mental health strengths and difficulties. By identifying recurrent profiles, this review offers a framework to better understand the early‐stage mental health needs of adolescents. This framework has the advantages of being developmentally grounded and empirically informed while being inclusive of adolescents who presently struggle to access support despite their elevated risk. This includes adolescents with sub‐thresholds of clinical symptoms or those languishing in poor mental well‐being and quality of life Keyes, 2002; Keyes, Dhingra, and Simoes, 2010).

Our review identified 59 mental health clusters within 13 studies with a large sample of community and school‐based adolescents (N = 103,098) from 3 WHO regions. A synthesis of these data suggests that clusters could be grouped into six broad cluster types. These varied with respect to positive indicators of mental well‐being and negative indicators of socioemotional and behavioral difficulties. The key findings of this review are described below.

Two commonly identified clusters captured profiles of psychological well‐being and strengths. The first included adolescents high in psychological well‐being and low difficulties (Flourishing). The other included adolescents with moderate well‐being with somewhat elevated difficulties (Moderate Mental Health). A large proportion of adolescents were captured by these two cluster types (28.99%–39.42% and 21.22%–40.44%, respectively). These findings suggest that most adolescents experience moderate to high mental well‐being, but some can still manifest a need for support. For example, adolescents with suboptimal well‐being due in part to slightly elevated symptoms. Such individuals could be selectively targeted for low‐intensity supports such as online or app‐based interventions designed to enhance coping and psychological regulation heightened symptom (Lee et al., 2024; McCloud et al., 2020).

One study described a vulnerable cluster of adolescents who were neither flourishing nor experiencing difficulties (Languishing). Although this cluster type appeared only once, it is difficult to accurately estimate its prevalence. This is because almost 40% of studies excluded measures of psychological well‐being. Such studies could only discuss the absence of symptoms rather than the absence of well‐being (Asymptomatic). The asymptomatic cluster type accounted for a large proportion of around 46.46%–64.78% of adolescents. This suggests that the languishing adolescents could be considerably more common. Future research is needed to explore this gap. From a practical perspective, this is a cluster that might benefit from target supports aimed at boosting mental well‐being and positive affect (Froh et al., 2008).

The remaining cluster types captured profiles of socioemotional, behavioral, and interpersonal difficulties. One cluster type included adolescents with moderately intense problems alongside low to moderate well‐being (Struggling). The other cluster type comprised those experiencing intense problems and a total absence of well‐being (Mental Health Problems). Interestingly, problems tended to be more discrete among adolescents in the Struggling group—symptoms were usually isolated to a specific clinical domain. In contrast, those in the Mental Health Problems cohort reported intense symptoms across multiple domains. Future research might consider whether disorder‐specific interventions targeting single domains—such as CBT for depression (Beck et al., 2024)—are best suited to adolescents in the Struggling group, where problems are more circumscribed. Such solutions may be less appropriate for the complex presentations observed in the Mental Health Problems group. Better outcomes might instead accrue from transdiagnostic psychological interventions that target distress more broadly (Gonzalez‐Robles et al., 2018; Radunz et al., 2025)

Most of the 59 clusters identified were associated with poor mental health. Thirty‐eight clusters were characterized by patterns of symptom presentations, while 12 were characterized by positive traits of mental well‐being. However, profiles based on problems tended to be small; fewer adolescents were assigned to the Mental Health Problems and Struggling cluster types in comparison to the others. The relative size of cluster types is interesting since traditional research and practice focus on those with clinically significant distress. Approaches built around such extreme but rare cases (and the health systems they inform) risk overlooking experiences and needs that exist across the full spread of mental health. This can include Languishing adolescents and those with Moderate Mental Health, who may be at risk of long‐term negative outcomes (Buckman et al., 2018; Paksarian et al., 2016). This review, therefore, makes an important contribution by synthesizing evidence of, and giving definition to, these underrepresented cohorts.

Much of this literature stems from the dual‐factor model of mental health (Keyes, 2002). This approach posits that mental health is not a unitary concept characterized solely by the absence of symptoms. Instead, it involves two dimensions—one based on traits of well‐being and the other of psychopathology. This model was seminal in establishing concepts of languishing and flourishing, and it has guided much of the classification research on youth mental health (Petersen et al., 2019). It also predicts four clusters since individuals can be high or low on these two dimensions (Suldo et al., 2016). Our findings of this review are broadly consistent with this perspective. Person‐centered analyses typically identify four clusters, and the most common cluster types were flourishing (or good mental health), moderate flourishing (or moderate mental health), struggling/surviving (or symptomatic but content) and languishing (or troubled).

This review also extends the literature. We found that mental health experiences were represented at varying levels of depth. Some studies identified clusters using broad indices of mental health, like patterns of externalizing. Others identified clusters at a more granular level and explored specific symptom sets, like suicidal behavior, risky sex, or eating difficulties. This implies that mental health can be arranged hierarchically, with some traits sitting on top of others. This is consistent with research on dimensional mental health taxonomies like the Hierarchical Taxonomy of Psychopathology (Kotov et al., 2017; Ringwald et al., 2023). Here, mental health is arranged such that one super‐spectra dimension (like internalizing) sits above sub‐spectra difficulties (like fear/distress, somatoform problems, social maladjustment) (Holmes et al., 2021; Michelini et al., 2019). Future research might use clustering techniques that capture this hierarchical structure. One study, for example, employed hierarchical clustering to show how cognitive abilities (like literacy, numeracy, and memory) can be organized across different layers. The top layer captured broad abilities levels (e.g., relatively impaired), while the lower layers split into more granular presentations (e.g., impaired in memory only) (Siugzdaite et al., 2020).

Included studies were of good quality. They were all powered to observe at least 30 adolescents per cluster (Dalmaijer et al., 2022). Also, although nonstandardized items featured in many studies, so too did standardized inventories with available psychometric norms. A challenge is a lack of consistent measurements and methods. Primarily, many studies lack measures of well‐being and strengths. This means the literature can be agnostic to cluster types characterized by the presence or absence of positive indicators like well‐being. Also, there was considerable heterogeneity in how symptoms are measured (Mew et al., 2020). This meant it was not feasible to pool clusters from separate studies and quantitatively synthesize clusters from different samples. Future research will benefit from a more consistent approach. One recommendation is to employ a combination of measures that proxy (1) day‐to‐day strengths and difficulties typical of community samples (e.g., SDQ) as well as (2) more significant clinical presentations (e.g., RCADS and CES‐D) and (3) elements of psychological well‐being and strengths (WEMWBS).

Some limitations should be mentioned. A small number of studies were included in this review (k = 13). This number reflects the use of stringent inclusion and exclusion criteria. These criteria were designed to capture high‐quality studies reporting on comprehensive profiles based on a broad range of measures. So even though 13 studies were included, each recruited a highly representative sample of community and school‐based adolescents. The combined sample size was also large (n = 103,098) and spanned 3 WHO regions. This suggests our findings are highly generalizable despite the small number of studies included. A second limitation relates to the cross‐sectional focus of this review. Understanding the stability and development of mental health profiles is critical for a clinical staging approach. However, an important first step is to identify common profiles that exist in any one moment in time. The logical next step is to build on these foundations and explore the evolution of these profiles over time as well as factors that influence how adolescents transition from one group to another.

In summary, this review describes 13 studies that used person‐centered statistics to study mental health profiles in large community and school‐based samples of adolescents. Studies were only included if the researchers administered multiple measures of socioemotional, behavioral, and/or interpersonal problems along with measures of psychological well‐being and strengths. A six‐cluster model of adolescent mental health in community settings was found. It consisted of Flourishing, Moderate Mental Health, Languishing, Asymptomatic, Struggling, and Mental Health Problems groups. These cluster types were also found to vary with respect to the intensity, concurrence, and depth (or granularity) of traits. Overall, these findings can contribute to the development of new classifications of youth mental health that can guide early‐stage and tailored interventions in community settings.

AUTHOR CONTRIBUTIONS

Marc Bennett: Conceptualization; data curation; formal analysis; investigation; methodology; project administration; resources; software; validation; visualization; writing – original draft; writing – review and editing. Claire O'Dwyer: Data curation; formal analysis; methodology; project administration; validation; writing – original draft; writing – review and editing. Varsha Eswara‐Murthy: Data curation; formal analysis; methodology; project administration; validation; writing – original draft; writing – review and editing. Tim Dalgleish: Conceptualization; project administration; supervision; writing – original draft; writing – review and editing. Foiniki Nearhou: Conceptualization; data curation; formal analysis; investigation; methodology; project administration; resources; software; validation; visualization; writing – original draft; writing – review and editing.

FUNDING INFORMATION

Contributions by MB, COD, and VEM were supported by funding from the Health Service Executive, Ireland. FN is supported by a UCD Ad Astra Scholarship.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

ETHICS STATEMENT

Ethics approval was not required for this study because it is a systematic review of previously published research.

PATIENT CONSENT

Patient consent was not required because this study is a systematic review of previously published research.

Supporting information

Appendix S1.

Bennett, M. , O’Dwyer, C. , Eswara‐Murthy, V. , Dalgleish, T. , & Nearchou, F. (2025). Early‐stage profiles of adolescent mental health difficulties and well‐being: A systematic review of cluster analyses in large school and community samples. Journal of Research on Adolescence, 35, e70061. 10.1111/jora.70061

DATA AVAILABILITY STATEMENT

Partial data supporting the findings of this study (e.g., summaries of extracted data) are presented within the main manuscript and Supporting Information. The full data extraction master sheet is available from the corresponding author upon reasonable request.