Longitudinal Clustering of Psychopathology Across Childhood and Adolescence: An Approach Toward Developmentally Based Classification

Abstract

Current classification systems of psychopathology focus on cross-sectional symptomatology rather than continuity, discontinuity and comorbidity across development. Here, a community sample of 600 youth was assessed every 3 years from early childhood through late adolescence using semi-structured diagnostic interviews. We used longitudinal k-means clustering of joint-diagnostic trajectories to identify 6 distinct clusters (healthy, childhood anxiety, childhood/adolescent ADHD, adolescent depression/anxiety, adolescent depression/substance use, and early childhood disruptive behavior). When comparing psychopathology clusters to the healthy cluster on age 3 predictors (parental education and psychopathology, early environment, temperament, cognitive and social functioning) and age 18 functional outcomes, the clusters captured developmental patterning of psychopathology not apparent in cross-sectional nosology. The study serves as a proof of principle in applying a longitudinal clustering approach to common mental disorders, affording a rich perspective on the unfolding of sequential comorbidity and heterotypic continuity and identifying transdiagnostic subgroups with meaningful clinical, family, and temperamental correlates.

Introduction

Existing classification systems of psychopathology vary in their consideration of development, but they are largely cross-sectional, emphasizing current psychopathology. However, some researchers have called for greater emphasis on development and course to better understand syndromes or symptom dimensions as they unfold and interact with one another (Oldehinkel & Ormel, 2023). An explicit focus on disorder continuities and discontinuities could shed light on distinct patterns of multifinality and equifinality which would inform classification of psychiatric illness, reduce within-disorder heterogeneity, and provide clues to underlying processes and mechanisms.

Following Kraepelin (1919), development and course have been accepted as a key feature of diagnostic validity (Robins & Guze, 1970). In the Diagnostic and Statistical Manual for Mental Disorders, fifth edition (DSM-5; American Psychiatric Association, 2013), age of onset plays a key role in the criteria for some disorders (e.g., attention deficit-hyperactivity disorder, autism) and persistence is critical for others (e.g., persistent depressive disorder, schizophrenia). Yet, existing classification systems fail to provide a comprehensive picture of illness progression from a life course perspective (Maughan & Collishaw, 2015). Individuals often accumulate diagnostic comorbidities, and although disorders may persist over time, people often transition between related or unrelated disorders (Caspi et al., 2020; Copeland et al., 2013).

A life course perspective can contribute to delineating more homogenous groups of disorders. For example, diagnostic continuity over time provides a stronger genetic “signal” than diagnoses at a single time point (Kendler et al., 2023); the unfolding of comorbidities over time may be markers for heterogeneity within diagnostic groups (e.g., alcoholism preceded by anxiety differs in fundamental respects from alcoholism preceded by antisocial personality; Chassin et al., 2013); and the emergence and continuity of disorders at different developmental stages may reflect different conditions (e.g., adolescent-limited versus life course persistent antisocial behavior; Moffitt & Caspi, 2001).

Several more recent frameworks have attempted to address these issues. The clinical staging model (McGorry et al., 2006) takes a transdiagnostic approach which posits that subthreshold psychopathology increases in specificity as it becomes more severe over time. However, this model does not address symptom discontinuities and the accumulation of comorbidities during the progression of illness. To address problems of heterogeneity and comorbidity, the Hierarchical Taxonomy of Psychopathology (HiTOP) is based on factor analytic studies that decompose categorical diagnoses into more homogenous dimensions, many of which cut across multiple disorders (Kotov et al., 2017). These symptom dimensions are aggregated into higher-order spectra, organizing psychopathology in a hierarchical fashion that provides important insights into the patterning of comorbidity. However, HiTOP does not address the emergence and course of symptom dimensions, their continuities and discontinuities, or the patterning of their interrelationships over time. While some studies have considered the dimensionality of psychopathology across several decades (Caspi et al., 2014), the resulting factor structure did not attempt to characterize and differentiate longitudinal patterns.

While factor analytic techniques focus on relationships between variables, an alternative approach that may be more directly suited to classification focuses on relationships between people. Examples of data-driven person-centered approaches are cluster analysis, latent profile (or latent class) analysis, and growth mixture modeling. Cluster and latent profile analysis have rarely been applied to longitudinal data because traditional techniques require modifications to account for their nested structure. However, unlike growth mixture modeling, they can handle complex non-linear relationships with a limited number of waves.

Applying data-driven person-centered approaches across a developmentally informative timeframe, Healy et al. (2022) used a longitudinal extension of latent profile analysis, latent profile transition analysis, to derive patterns of transitions of internalizing and externalizing symptoms in two large cohorts – one followed in childhood and another followed in adolescence. They found 4 profiles: no psychopathology, high levels of psychopathology, internalizing problems, and externalizing problems. About 50% of both cohorts transitioned into one of the three psychopathology profiles at some point in development, with high psychopathology most often preceded by externalizing problems. This study robustly captured the dynamic (and, in some cases, persistent) nature of psychopathology over time. However, the groupings were not truly longitudinal, as they created classes cross-sectionally and examined transitions into and out of these cross-sectionally derived clusters across waves. Longitudinal clustering, in contrast, builds groups based on both cross-sectional and longitudinal variation.

Only a few studies have attempted to apply cluster or latent profile analysis to characterize the development and/or continuity of psychopathology. Using k-means cluster analysis for longitudinal data, Martinek et al. (2023) clustered trajectories of weekly ratings of the course of depression for a year following hospital or clinic discharge and found five subgroups with unique patterns of recovery, relapse, and persistence in an adult sample. The PsyCourse Study (Schulte et al., 2022) focused on patients with schizophrenia and bipolar-spectrum disorders followed for 18 months. Applying the same procedure, they reported five distinct longitudinal clusters that differed in diagnoses and functioning based on three clinical dimensions. To our knowledge, however, no studies have applied this novel approach to a broad range of mental disorders over the course of a longer and developmentally informative timeframe. By examining the trajectories of multiple disorders as they jointly co-evolve over time, it is possible to identify subgroups with unique patterns of comorbidity and homo- and heterotypic continuity during specific developmental periods.

The present study applied a longitudinal clustering approach to common mental disorders in a community sample prospectively assessed triennially from early childhood through the end of adolescence. We identified subgroups of individuals based on multiple simultaneous (or joint) trajectories of diagnostic course from ages 3 to 18 years old. Importantly, we included mental disorders beginning in the preschool years, a period that has been relatively neglected in psychopathology research (Angold & Egger, 2007; Bufferd et al., 2016), but which has important prognostic implications. For example, Finsaas et al. (2018) observed that 48% of preschoolers with a psychiatric diagnosis met criteria for a mental disorder in early adolescence. In using a data-driven longitudinal clustering approach to common mental disorders throughout child and adolescent development, we accounted for the accumulation of comorbidities and continuities and transitions among disorders as an initial step towards a developmentally-based classification framework. Once we identified the optimal cluster solution, we compared the cluster on a set of a priori variables typically used as predictors of later internalizing and externalizing psychopathology, including parental education and psychopathology; parenting and other markers of a child’s early environment; child temperament; and child cognitive and social functioning. We also assessed how the clusters differed based on key functional outcomes, such as interpersonal and academic functioning.

Transparency and Openness

Preregistration

This study was not pre-registered.

Data, Materials, Code and Online Resources

All code and data can be found here:

https://osf.io/d93ks/?view_only=158233500e6e4a0896b1ebd9fdeb4c1e.

Reporting

We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study.

Ethical Approval

Ethical approval was obtained by the Stony Brook University Institutional Review Board and was carried out in accordance with the provisions of the World Medical Association Declaration of Helsinki.

Method

Participants

Participants were drawn from the Stony Brook Temperament Study¹, a longitudinal study of risk factors and pathways to psychopathology from age 3 to age 18 (Klein & Finsaas, 2017). Families with a 3-year-old child living within 20 miles of Stony Brook, NY, were recruited using commercial mailing lists for a larger study of risk for mental disorders; children were excluded if they did not live with a biological parent or had significant medical or developmental disorders (N=559). An additional 50 families were added in the second wave of assessments when children were 6 years old to increase the diversity of the sample. Only one child per family was included. Children were reassessed every 3 years until age 18.

Diagnostic interviews were conducted with a parent when the children were 3 (N=541) and 6 (N=516) years old, and with a parent and the child at ages 9 (N=488), 12 (N=476), and 15 (N=458). At age 18 (N=418), only the youth were interviewed. Participants were included in the study if they had at least one wave of diagnostic information. Of these 600 participants, 45 (6.9%) completed one, 41 (6.3%) completed two, 35 (5.8%) completed three, 48 (8.0%) completed four, 113 (18.8%) completed five, and 318 (53.0%) completed all six assessment waves. When considering the unique participants across all samples, 272 (45.3%) were female, and 477 (87.5%) were White and non-Hispanic. The sample’s demographic and socioeconomic characteristics were representative of the larger county (Bufferd et al., 2011).

Diagnostic Assessments

The Preschool Age Psychiatric Assessment (Egger et al., 1999), an interviewer-based structured diagnostic interview, was administered to parents by telephone at the age 3 wave and in-person at the age 6 wave. Diagnostic interviews with parents about their children conducted in-person and by phone yield similar results (Lyneham & Rapee, 2005). Diagnostic and Statistical Manual for Mental Disorders, fourth edition (DSM-4; American Psychiatric Association, 1994) diagnoses (including modified criteria for preschool depression; Luby et al., 2002) in the past 3 months were derived following the developers’ algorithms. The Preschool Age Psychiatric Assessment has good test-retest reliability over a mean 11 day interval (Egger et al., 2006). Interrater reliability in our study was assessed using audio recordings on sample of 21 interviews at age 3 and 35 interviews at age 6 enriched for psychopathology. At age 3, Kappa was 1.00 for all disorders. At age 6, Kappas were .64 for any depressive disorder, .89 for any anxiety disorder, .64 for attention-deficit/hyperactivity disorder (ADHD), and .87 for any disruptive behavior disorder (DBD), all of which are in the moderate-substantial range (Shrout, 1998).

The Kiddie Schedule for Affective Disorders Schizophrenia Present and Lifetime Version, a semi-structured interview for school-age children and adolescents (Kaufman et al., 1997), was administered to a parent and the child at the age 9, 12, and 15 assessments and to the youth alone at the age 18 assessment. In the age 9 wave, lifetime psychopathology was ascertained but only disorders present from ages 7–9 are included in our analyses. At the ages 12, 15, and 18 waves, psychopathology was assessed since the previous assessment. Interrater reliabilities (indexed by Kappa) was determined using videotapes of interviews of samples of participants enriched for psychopathology. Interrater reliability ratings were obtained using 74 interviews at age 9, 25 interviews at ages 12 and 15, and 34 interviews at age 18. Interrater reliabilities ranged from 0.72 to 0.88 for any depressive disorder, 0.67 to 0.94 for any anxiety disorder, 0.85 to 1.00 for attention-deficit/hyperactivity disorder (ADHD), and 0.58 to 0.91 for any disruptive behavior disorder (DBD), all of which are in the fair, moderate, or substantial range (Shrout, 1998).

All diagnostic interviews were conducted in-person or remotely by clinical psychology graduate students and Masters-level clinicians, supervised by a senior child and adolescent psychiatrist and clinical psychologist. In-person and remote interviews with adolescents and young adults yield comparable results (Rohde et al., 1997). The following DSM-IV diagnoses were examined: depressive disorders (major depressive disorder, dysthymic disorder, depressive disorder not otherwise specified [NOS] ages 3–18), anxiety disorders (specific phobia, social phobia, separation anxiety, generalized anxiety, and panic and/or agoraphobia at ages 3–18 and anxiety disorder NOS at ages 9–18), disruptive behavior disorders (oppositional defiant disorder and conduct disorder at ages 3–18 and DBD-NOS at ages 9–18), ADHD (ADHD at ages 3–18 and ADHD-NOS at ages 9–18), and substance use disorders (SUD; alcohol or drug abuse or dependence at ages 9–18).

Validation Measures

The clusters in the optimal solution were validated against a set of age 3 predictors and age 18 outcomes.

Age 3 Predictors

Parental education and psychopathology.

Parental education, as a proxy for socioeconomic status, was defined as the number of parents with a bachelor’s degree or higher. Parental psychopathology was assessed with the Structured Clinical Interview for DSM-IV, Non-patient version (First & Gibbon, 2004). We determined the number of parents with a lifetime history of any depressive disorder, any anxiety disorder, and any substance use disorder. Kappas for inter-rater reliability (N=30) were 0.93 for mood disorder, 0.91 for anxiety disorder, and 1.00 for substance use disorder, all of which are in the substantial range (Shrout, 1998).

Early environment.

Early environment variables consisted of the mother- and father-reported Dyadic Adjustment Scale (Spanier, 1976), a 32-item questionnaire assessing marital satisfaction (Cronbach’s alphas [α] = 0.94 and 0.95, both in the substantial range). Example items include how often parents have arguments about finances, household tasks, amount of time spent together, career decisions, and other life domains, as well as how often they discuss divorce. Additionally, the life stress scale (Costello et al., 1998), a module of the Preschool Age Psychiatric Assessment, assesses a wide range of life events that might affect the child, including ‘high magnitude’ events associated with PTSD and “low magnitude” events (e.g., parental separation, changing schools). We summed the number of events experienced prior to age 3.

Regarding parenting, each child and one parent participated in a 30-minute structured parent-child interaction session using a modified version of the Teaching Tasks (Egeland & Hiester, 1995). The battery consisted of six standardized tasks, adapted from the Ainsworth Strange Situation Procedure (Ainsworth et al., 1978), that was designed to elicit individual differences in parenting. Tasks were video-recorded and coded for behavioral indices of support (parental expression of positive regard and emotional support), hostility (parent’s expression of anger, frustration, annoyance, discounting, or rejection), and the quality of relationship between parent and child (affective and verbal sharing between child and parent, contingent responding to each other, sensitivity of parent to child’s distress, and effective conflict resolution). Ratings were summed across episodes and reliability was computed via intraclass correlations (ICC; 2 way random effects, absolute agreement) on a random sample of 55 individuals. ICCs ranged from 0.59–0.91, which range from fair to substantial (Shrout, 1998).

In addition, both parents completed the Parenting Styles and Dimensions Questionnaire (PSDQ; Robinson et al., 1995), a 47-item measure comprised of three factor-analytically derived dimensions: authoritative (warmth and involvement; democratic participation), authoritarian (verbal hostility; harsh punishment), and permissive (lack of follow-through, ignoring misbehavior) parenting styles. Alphas ranged from 0.74 to 0.82, in the moderate-substantial range (Shrout, 1998).

Temperament.

The Laboratory Temperament Assessment Battery (Lab-TAB; Gagne et al., 2011), is an observational measure designed to assess child temperament using a series of emotion-eliciting episodes. The child participated in 12 episodes, which were videotaped and coded for facial, vocal, and postural indicators of emotion and several emotion-relevant behaviors. Ratings were z-scored, summed across episodes, and used to derive a number of temperament constructs (see Olino et al., 2010). For this paper, we examined positive emotionality (positive affect and engagement/interest); negative emotionality as a whole and each of its three components (fear, sadness, anger) separately; and impulsivity. Interrater ICCs (2 way random effects, absolute agreement) for the variables included in this paper, based on a random sample of 35 individuals, ranged from 0.73 to 0.89, which are in the moderate-substantial range (Shrout, 1998).

We included three episodes designed specifically to assess temperamental behavioral inhibition. The “risk room” episode had the child explore a set of novel and potentially threatening stimuli (e.g., Halloween mask, black box). The “stranger approach” episode involved a male accomplice approaching the child while left alone and speaking to the child while slowly walking closer. In the “exploring new objects” episode, the child was given the opportunity to explore ambiguous stimuli (e.g., mechanical spider). Coding procedures followed prior literature (Olino et al., 2010; Pfeifer et al., 2002). Briefly, each episode was divided into 20–30 second epochs, and within each epoch, a maximum intensity rating of facial, vocal, or bodily fear was coded on a 4-point Likert scale. Behavioral inhibition was computed as the average of these standardized ratings, along with standardized ratings of latency to fear (reversed); latency to touch objects; total number of objects touched (reversed); tentative play; referencing the parent; proximity to parent; referencing experimenter; time spent playing (reversed); startle; sad facial affect; latency to vocalize; approach toward the stranger (reversed); avoidance of the stranger; gaze aversion; and verbal/nonverbal interaction with the stranger (reversed). Interrater ICC (N=28) was 0.88, which is in the substantial range (Shrout, 1998).

We also assessed behavioral inhibition using parent reports. The Behavioral Inhibition Questionnaire (BIQ; Bishop et al., 2003) was administered to the parent who accompanied the child to the laboratory (typically the mother). This 30-item questionnaire assesses the frequency of the child’s behavioral inhibition across six contexts in the domains of social novelty, situational novelty, and novel physical activities with possible risk of injury (α = 0.96). For example, items assess the child’s comfort in asking other children to play; child’s caution in activities that involve physical challenges; child’s ability and timing in adjusting to new situations; and child’s comfort with being the center of attention. We z-scored the Behavioral Inhibition Questionnaire and the Laboratory Temperament Assessment Battery and summed them to provide a composite index of behavioral inhibition.

The child’s mother completed the Children’s Behavior Questionnaire (CBQ; Rothbart et al., 2001), a 191-item parent report measure designed to assess temperament in young children (αs for 16 subscales ranged from 0.65 – 0.91, in the moderate-substantial range). Items comprise the following scales: activity level; anger/frustration; attentional focusing; discomfort; fear; high and low intensity pleasure; impulsivity; inhibitory control; perceptual sensitivity; positive anticipation; shyness; sadness; smiling/laughter; and soothability. We examined the three higher order factors (surgency, negative affectivity, effortful control) derived from Rothbart et al.’s (2001) factor analyses of the Children’s Behavior Questionnaire subscales.

Cognitive and social functioning.

To index cognitive functioning, each child completed the Peabody Picture Vocabulary Test, third edition (PPVT-III; Dunn & Dunn, 1997) and the Expressive One-Word Vocabulary Test (EOWVT; Brownell, 2000); these tests assess receptive and expressive vocabulary, respectively, based on presentation of various photos (e.g., children are asked to pick the square out of a series of photos of shapes). Additionally, to index social functioning, the 15-item Social Competence scale from the Vineland Screener (Sparrow et al., 1993), a parent-report measure of children’s adaptive behaviors, such as communication, socialization, and daily living skills, was administered. Example items include the child’s ability to make eye contact when meeting new people; child’s caution around things that could burn him/her; and child’s ability to pay attention to 15-minute stories.

Age 18 Functioning Outcomes

The UCLA Life Events Interview (LSI; Hammen et al., 1987) was administered to participants in the age 18 wave. Although the interview was designed to assess episodic and chronic stress, the latter scores can readily be interpreted as reflecting functional impairment (Harkness & Monroe, 2016). Interviewers used behavioral probes to assess functioning over the past year on a 5-point scale, including half-points (higher scores indicate poorer functioning). For the present study, we examined academic/work functioning and interpersonal (family, friends, peers, and romantic partners) functioning. Interrater ICCs for each domain (N=34) included in these summary scores ranged from 0.65 to 0.89, in the moderate-substantial range (Shrout, 1998).

Data Analysis

Cluster Estimation

k-means cluster modeling for longitudinal data (K-means longitudinal 3D (kml3d), <10.32614/CRAN.package.kml3d>), using R package version 4.2.0, was used to identify distinct clusters of psychopathology trajectories over 6 assessment time points. kml3d offers a non-parametric, expectation-maximization algorithm that clusters joint variable-trajectories, capturing the evolution and complex interactions between variables over time (Genolini et al., 2015). To evaluate the optimal number of cluster trajectories, we tested six longitudinal k-means models, with the number of clusters increasing stepwise from 2 to 7 computed using the kml3d algorithm. The k-means algorithm was initialized with the following procedure: “a) choose one center c₀ uniformly at random from among the data points; b) for each data point x, compute D(x), the distance between x and c₀; c) choose one new center c₁ at random using a weighted probability distribution proportional to D(x)²; d) remove c₀ from the list of centers; e) for each data point x, compute D(x), the distance between x and the nearest center that has already been chosen; f) randomly choose a data point as the new center c_i, using a weighted probability distribution where a point x is chosen with probability proportional to D(x)²; g) repeat steps e and f until k centers have been chosen” (Genolini et al., 2015).

The algorithm was run 50 times, with a maximum of 500 iterations if convergence was not reached; individual runs were automatically sorted by best fit. As part of the model, we imputed missing data using linear interpolation, then added a variation to make the trajectory follow the ‘shape’ of the population’s mean trajectory; thus, overall trends are informing cluster proportions, as well as within-person changes across time (Genolini et al., 2013). Because diagnostic data is binary, we utilized the deviance distance metric (rather than the Euclidean distance metric for continuous data) to find the optimal number of centroids. We assessed model fit using the Calinski-Harabasz index (Calinski & Harabasz, 1974), Genolini variant. This variant is notated as C_G(k) = (Trace(B) /Trace(W)) * (n−k / √k−1), where B is the between-cluster covariance matrix and W is the within-cluster covariance matrix. High values of Trace(B) denote well-separated clusters and low values of Trace(W) denote compact clusters (Genolini et al., 2015). This variant has the advantage of jointly considering both parsimony and fit. The Calinski-Harabasz index was found to be the best index in detecting the optimal number of clusters (Milligan & Cooper, 1985).

Cluster Validation

Once clusters were estimated and extracted from kml3d, handling of missing values in the validation measures was performed using multiple imputation using the MICE package (<10.32614/CRAN.package.mice>) in R (van Buuren & Groothuis-Oudshoorn, 2011). 20 imputations were conducted, with 100 as the maximum number of iterations if convergence was not reached. The reverse monotone visit sequence was used, so data from the variables with the greatest amount of missingness was imputed first. Imputation performance was assessed by inspecting density plots of all imputations at once to ensure they follow the same shape of distribution.

For descriptive purposes, chi-square tests were performed at each imputation to assess differences in the prevalence of the diagnostic categories among the clusters at each time point; these values were then pooled across imputations. Chi-square tests were also performed on demographic variables (i.e., sex and race/ethnicity). Race (White, Black or African American, Asian, American Indian or Alaska Native, Native Hawaiian or other Pacific Islander) and ethnicity (Hispanic or non-Hispanic) were measured via self-report. For validation purposes, one-way ANOVAs and post-hoc Holm-Bonferroni corrected pairwise t-tests between the healthy (reference) cluster and every other cluster were computed on each imputed dataset, which were then pooled. Corresponding Cohen’s d values were computed for each comparison.

Results

Table 1 displays respective Calinski-Harabasz (Genolini variant) fit indices for the best-fitting cluster solutions across iterations for the 2–7 cluster solutions, with the 6-cluster solution providing the best fit. Figure 1 displays the trajectories of the 6 clusters as a function of each of the 5 disorders. Cluster A (46.7% of the sample, N=280), demarcated by the red line, is characterized primarily by the absence of psychopathology from early childhood through adolescence (i.e., the ‘healthy’ cluster). Cluster B (13.5% of the sample, N=81), demarcated by the yellow line, is characterized by elevated rates of anxiety disorders, particularly in late childhood and early adolescence, and will be referred to as the ‘anxiety’ (ANX) cluster. Cluster C (12.3%, N=74), demarcated by the green line, is characterized by increasing rates of ADHD through early adolescence, and will be referred to as the ADHD cluster. Cluster D (12%, N=72), demarcated by the light blue line, is characterized by gradually rising rates of anxiety disorders throughout childhood and adolescence along with a rapid increase in rates of depressive disorders beginning in mid-adolescence. We will refer to this cluster as the ‘depression/anxiety’ (DEP/ANX) cluster. Cluster E (8.5%, N=51), demarcated by the dark blue line, is characterized by moderately rising rates of depressive disorders through adolescence along with a very sharp increase in SUD in late adolescence. This cluster will be referred to as the ‘SUD/depression’ (SUD/DEP) cluster. Finally, Cluster F (7%, N=42), demarcated by the pink line, is characterized by high but decreasing rates of disruptive behavior disorders through childhood, with a modest increase in mid-adolescence. This cluster will be referred to as the DBD cluster.

Table 1.

Fit Indices for Best-Fitting Solution for Each k-means model

Cluster Solution	Best-Fitting Calinski-Harabsz (Genolini variant) fit index
2	105.104
3	144.631
4	173.115
5	193.703
6	198.020
7	193.796

Figure 1. Cluster Visualization for Best-Fitting Cluster Solution.

Note. Dep = Any depressive disorder including NOS; Anx = Any anxiety disorder including NOS; ADHD – Attention Deficit/Hyperactivity Disorder; DBD = Disruptive Behavior Disorder; SUD = Substance Use Disorder

Qualitatively, this optimal 6-cluster solution differed from the 5-cluster solution in that the 5-cluster solution did not capture childhood anxiety as its own cluster. Instead, most participants in the cluster B (ANX) group were included in cluster A (the ‘healthy’ cluster). The 7-cluster solution, by contrast, included two clusters characterized by disruptive behavior disorder - one cluster with moderate rates of disruptive behavior disorder throughout childhood and adolescence often accompanied by ADHD (N=22) and the other cluster with a spike in disruptive behavior disorder in adolescence, along with rising rates of ADHD in adolescence and moderate and increasing rates of anxiety across childhood and adolescence (N=17). Thus, although the 7-cluster solution is theoretically interesting, the overall model fit is poorer than that of the 6-cluster solution, and the N for some clusters was quite small.

Table 2 compares the distribution of diagnoses and chi-square results at each wave across the 6 clusters. Except for substance use disorder at age 15 (as only one individual had this diagnosis), all omnibus chi-square values for all disorder categories at all time points were significant. Post-hoc Holm-Bonferroni corrected multiple comparisons show that cluster B (ANX) had significantly higher rates of anxiety disorders than all other clusters at ages 9 and 12. Cluster C (ADHD) had significantly higher rates of ADHD than all other clusters at ages 9, 12, 15, and 18. Cluster D (DEP/ANX) had significantly higher rates of depression than all other clusters at age 18 and significantly higher rates of anxiety than all other clusters except cluster C at age 18. Cluster E (DEP/SUD) also had significantly higher rates of depression than all other clusters (except cluster D) at age 18, as well as significantly higher rates of substance use disorder than all other clusters at age 18. Cluster F (DBD) had significantly higher rates of disruptive behavior disorders than all other clusters at age 3.

Table 2.

Chi-Square Comparison of Diagnoses Among Best-Fitting Cluster Solution

Diagnosis	Age	A (Healthy) N=280		B (ANX) N=81		C (ADHD) N=74		D (DEP/ANX) N=72		E (SUD/DEP) N=51		F (DBD) N=42		X2(5)
		n	%	n	%	n	%	n	%	n	%	n	%
DEP	3	4cde	1.34	1cdf	1.35	2abdef	2.96	0abcef	0.07	0cdf	0.56	4abcde	10.20	18.77**
	6	4bcdef	1.51	5ac	6.76	12abdef	16.60	4ac	6.22	3ac	6.35	2ac	5.10	27.78***
	9	1bcdf	0.49	4ae	4.53	4ae	5.79	4ae	5.16	0bcdf	0.84	1ae	2.72	14.22*
	12	9bcd	3.21	6adef	7.70	8aef	10.55	9abef	11.84	2bcd	4.67	1bcd	3.17	13.50*
	15	11bcdef	3.81	10acdef	11.82	11abdef	15.06	26abcef	35.58	14abcd	27.17	8abcd	18.93	64.14***
	18	18bcdef	6.46	8acdef	10.46	18abdef	24.13	68abcef	94.71	25abcdf	49.58	9abcde	22.34	265.05***
ANX	3	31bcdef	11.24	33acdef	40.15	20abdf	27.41	13abcef	17.92	12abdf	23.25	13abcde	31.63	40.56***
	6	14bcdef	4.97	40acdef	49.03	14abef	19.05	13abef	18.25	5abcdf	10.46	11abcde	25.85	96.51***
	9	26bcdef	9.32	50acdef	62.20	22abe	29.92	19abe	26.79	11abcdf	20.92	13abe	30.61	103.81***
	12	16bcdef	5.54	54acdef	66.84	28abef	37.45	25abcf	34.59	10abcdf	19.89	11abcde	26.42	147.97***
	15	17bcdef	6.00	30acdf	36.74	23abdef	30.82	31abcef	42.53	12acd	23.25	9abcd	21.54	77.63***
	18	25bcdef	8.86	28acdf	34.04	27abd	36.36	43abcef	59.66	15ad	29.51	11abd	26.87	95.94***
ADHD	3	2cef	0.63	0cef	0.06	4abdf	5.21	0cef	0.07	2abdf	4.30	5abcde	12.93	33.76***
	6	8cef	2.76	2cef	2.18	17abdef	22.84	1cef	1.85	2abcdf	4.01	5abcde	12.36	52.10***
	9	12bcef	4.30	5acde	6.76	55abdef	74.84	2bcef	2.85	6abcdf	12.04	3acde	6.58	260.99***
	12	19bcdef	6.62	6acdef	7.41	69abdef	93.69	2abcef	2.78	9abcdf	18.39	5abcde	12.81	322.88***
	15	14bcef	5.02	5acdef	5.64	65abdef	87.52	2bcef	2.84	9abcdf	18.21	4abcde	9.41	317.07***
	18	17cef	6.12	5cef	6.00	54abdef	73.55	4cef	5.75	11abcdf	20.92	4abcde	10.77	221.49***
DBD	3	2bcef	0.65	0acef	0.47	9abdf	12.48	0cdef	0.13	5abdf	10.55	39abcde	92.97	388.24***
	6	10bcef	3.52	4acdef	4.70	12abdf	16.73	2abcef	3.24	8abdf	15.87	14abcde	34.01	58.93***
	9	3.10bcef	1.11	1acdef	1.53	14abdef	18.98	0.10bcef	0.13	4.29abcd	8.40	2abcd	7.03	54.13***
	12	6bcdef	1.99	4acef	4.76	9.48abd	12.81	2acef	3.11	6abd	12.51	5abd	12.24	24.75***
	15	7cdef	2.60	3cdef	3.12	6.38abdf	8.62	0abcef	0.46	4abdf	8.31	7abcde	16.87	23.80***
	18	4bcdef	1.34	2acef	2.29	6.19abd	8.37	1acef	1.46	4abd	7.84	3abd	6.80	17.61**
SUD	15	1	0.48	1	0.65	0.19	0.26	1	1.59	0	0.28	0	0.00	4.60
	18	13cdef	4.57	4cef	5.23	3.86abef	5.21	1aef	1.19	49abcdf	96.64	3abcde	6.12	378.75***

^*p < .05;

^**p <.01;

^***p < .001

Note. Estimates reflect pooled raw numbers of number of participants with the disorder in each cluster. Percentages reflect proportion of individuals with the disorder within each cluster. Squared values indicate significant difference between proportions (Holm Bonferroni-corrected for multiple comparisons). Means with different subscripts differ at the p=.05 level. DEP = Any depressive disorder including NOS; ANX = Any anxiety disorder including NOS; ADHD – Attention Deficit/Hyperactivity Disorder; DBD = Disruptive Behavior Disorder; SUD = Substance Use Disorder.

We also examined how clusters differed on sex and race/ethnicity (see Table 3). Proportion of males to females significantly differed across clusters. Holm-Bonferroni corrected post-hoc multiple comparisons revealed that cluster C (ADHD) had significantly more males than females, while cluster D (DEP/ANX) had significantly more females than males. The clusters did not differ on race or ethnicity; however, these results are underpowered given the relatively homogenous White, non-Hispanic study sample.

Table 3.

Means, Standard Deviations (or Percentages and Sample Sizes for Categorical Variables), and One-Way Analysis of Variance of Each Measure by Each Cluster

Measure	A (Healthy)	B (ANX)	C (ADHD)	D (DEP/ANX)	E (SUD/DEP)	F (DBD)	χ2 (5, N=600)	p
Percent Female	44.3 (124)	51.9 (42)	28.4 (21)	73.6 (53)	37.3 (19)	40.0 (13)	38.18	<.001
Percent Non-White	12.1 (34)	8.6 (7)	16.2 (12)	2.8 (2)	15.7 (8)	9.5 (4)	23.30	.274
Percent Hispanic	13.2 (37)	8.6 (7)	10.8 (8)	9.7 (7)	21.6 (11)	11.9 (5)	5.78	.328
Age 3 Predictors							F(5, 594)	η2
Percent of parents with college degree	52.5 (147)	51.0 (41)	43.3 (32)	54.0 (39)	36.3 (18)	42.9 (18)	6.99***	.14
Percent of parents with lifetime mood disorder	19.7 (55)	29.6 (24)	28.8 (21)	32.4 (23)	23.7 (12)	34.3 (14)	6.24***	.09
Percent of parents with lifetime anxiety disorder	22.3 (62)	29.8 (24)	35.5 (26)	29.8 (21)	25.4 (13)	32.9 (14)	5.14***	.09
Percent of parents with lifetime substance use disorder	26.4 (74)	29.2 (24)	35.8 (26)	31.5 (23)	41.0 (21)	29.3 (12)	4.68***	.09
Mother-reported dyadic adjustment	16.34 (3.76)	16.65 (3.21)	14.77 (4.21)	15.81 (3.63)	15.43 (3.88)	14.86 (4.22)	3.51**	.03
Father-reported dyadic adjustment	16.32 (3.52)	16.81 (3.09)	15.19 (3.87)	15.79 (3.51)	16.51 (3.32)	15.29 (3.32)	2.61*	.02
Parent-report life stress of child	2.39 (1.63)	2.79 (1.68)	3.69 (1.85)	2.76 (1.76)	3.31 (1.90)	2.98 (2.08)	3.67**	.01
Teaching Task support	4.48 (.55)	4.40 (.75)	4.34 (.65)	4.54 (.45)	4.47 (.54)	4.32 (.78)	4.44**	.10
Teaching Task hostility	1.19 (.30)	1.22 (.44)	1.23 (.33)	1.12 (.21)	1.19 (.36)	1.37 (.52)	9.92***	.26
Teaching Task quality of relationship	4.00 (.53)	3.98 (.71)	3.80 (.65)	4.08 (.52)	4.02 (.58)	3.78 (.86)	6.34***	.10
PSDQ mother-reported authoritarian parenting	19.50 (3.97)	19.45 (4.01)	20.90 (4.21)	18.81 (3.95)	20.28 (4.91)	23.66 (4.60)	9.60***	.07
PSDQ father-reported authoritarian parenting	19.85 (4.43)	20.25 (4.72)	21.88 (4.83)	19.13 (4.10)	19.96 (4.82)	23.56 (4.54)	7.73***	.06
PSDQ mother-reported permissive parenting	10.48 (2.97)	10.51 (3.21)	11.27 (3.18)	9.83 (2.75)	11.22 (3.57)	13.61 (3.55)	9.56***	.07
PSDQ father-reported permissive parenting	10.72 (2.94)	11.42 (2.95)	11.38 (3.13)	11.22 (3.36)	12.05 (3.27)	12.71 (4.00)	4.16**	.03
Peabody Picture Vocabulary Test	102.60 (14.53)	104.48 (14.97)	98.21 (13.96)	105.16 (12.38)	102.38 (12.27)	103.83 (11.68)	2.32*	.02
Expressive One Word Vocabulary Test	100.73 (12.64)	100.10 (12.66)	97.80 (15.14)	105.01 (11.90)	101.09 (11.99)	95.46 (13.33)	3.75**	.03
Vineland Screener	19.40 (3.65)	18.59 (3.77)	17.08 (3.49)	20.19 (3.55)	19.18 (3.39)	16.75 (3.49)	9.87***	.08
Lab-TAB positive emotionality	.171 (1.86)	−.142 (2.07)	−.473 (1.63)	−.161 (1.66)	−.168 (1.72)	−.162 (1.59)	1.73	.01
Lab-TAB negative emotionality	.547 (.260)	.631 (.333)	.584 (.246)	.561 (.286)	.623 (.220)	.620 (.283)	10.70***	.36
Lab-TAB anger	.555 (.333)	.652 (.388)	.645 (.350)	.505 (.338)	.564 (.310)	.710 (.365)	11.83***	.26
Lab-TAB sadness	.532 (.312)	.672 (.409)	.565 (.274)	.569 (.317)	.540 (.292)	.587 (.304)	10.48***	.29
Lab-TAB fear	.658 (.345)	.658 (.352)	.640 (.363)	.702 (.389)	.825 (.344)	.657 (.363)	8.53***	.25
Lab-TAB global impulsivity	.675 (.320)	.604 (.335)	.813 (.350)	.593 (.290)	.713 (.332)	.904 (.366)	18.56***	.27
Behavioral inhibition z-score	−.210 (1.54)	.804 (1.66)	−.292 (1.72)	.154 (1.49)	.105 (1.50)	.199 (1.61)	3.50**	.02
CBQ surgency	4.81 (.65)	4.44 (.84)	5.08 (.71)	4.79 (.67)	4.98 (.62)	5.13 (.72)	9.54***	.08
CBQ negative affectivity	3.82 (.55)	4.08 (.54)	3.98 (.62)	3.95 (.60)	3.84 (.58)	4.25 (.48)	9.92***	.11
CBQ effortful control	5.09 (.49)	4.99 (.52)	4.64 (.62)	5.23 (.57)	4.87 (.63)	4.70 (.60)	14.31***	.11
Age 18 Functional Outcomes
LSI Chronic Academic Stress	1.81 (.35)	1.82 (.41)	2.07 (.48)	1.91 (.44)	2.30 (.63)	1.78 (.36)	18.32***	.18
LSI Chronic Interpersonal Stress (parent-included)	2.03 (.34)	2.06 (.31)	2.21 (.44)	2.21 (.46)	2.39 (.46)	2.22 (.48)	17.13***	.21

^*p < .05;

^**p <.01;

^***p < .001

Note. Standard deviations (or sample sizes for categorical variables) are listed in parentheses. PSDQ = Parenting Styles and Dimensions Questionnaire; Lab-TAB = Laboratory Temperament Assessment Battery; CBQ = Children’s Behavior Questionnaire; LSI = Life Stress Interview.

Next, we examined the associations with the external validators to assess validity of the clusters. The external validators included a series of predictors from the initial (age 3) wave, as well as functional outcomes at the final (age 18) wave. Table 3 displays the means and SDs of the validation measures for each cluster and the results from one-way ANOVAs. All ANOVA results were significant, with the exception of positive emotionality from the Lab-TAB. We continued to examine pairwise t-tests for all variables, as it is possible to detect meaningful group differences despite a nonsignificant omnibus test (Tian et al. 2018). Table 4 displays corresponding Holm-Bonferroni corrected t-tests and corresponding p-values and effect sizes (using Cohen’s d) when comparing cluster A (healthy) to each of the other clusters. Significant effect sizes were in the medium-to-large range. Pairwise cluster comparisons with each psychopathology cluster as the reference group can be found in the Supplemental material.

Table 4.

Post-Hoc Comparisons of All Psychopathology Clusters with Cluster A

Measure	B (ANX)			C (ADHD)			D (DEP/ANX)			E (DEP/SUD			F (DBD)
Age 3 Predictors	t(359)	p	Cohen’s d	t(352)	p	Cohen’s d	t(350)	p	Cohen’s d	t(329)	p	Cohen’s d	t(320)	p	Cohen’s d
Number of parents with bachelor’s degree	−.398	.693	.049	−1.86	.071	.240	.116	.818	.031	−3.06	.003	.463	−1.59	.121	.259
Number of parents with lifetime mood disorder	1.92	.056	.261	1.86	.063	.256	2.67	.008	.369	.296	.768	.050	2.26	.024	.399
Number of parents with lifetime anxiety disorder	2.15	.032	.280	3.55	<.001	.461	1.53	.128	.208	.983	.326	.158	1.54	.123	.267
Number of parents with lifetime substance use disorder	.401	.689	.053	1.94	.052	.259	.890	.374	.122	3.11	.002	.486	.226	.821	.039
Mother-reported dyadic adjustment	1.53	.126	.203	−3.45	<.001	.447	−.664	.507	.090	−1.44	.150	.222	−2.40	.017	.397
Father-reported dyadic adjustment	2.94	.003	.369	−2.45	.014	.305	−.051	.959	.007	1.68	.093	.251	−2.21	.028	.358
Parent-report life stress of child	1.25	.213	.166	5.02	<.001	.678	1.40	.162	.194	3.07	.002	.484	2.03	.044	.342
Teaching Tasks support	−1.08	.279	.132	−1.45	.147	.193	.720	.472	.104	.250	.803	.040	−1.56	.120	.256
Teaching Tasks hostility	.562	.574	.069	−0.75	.455	.103	−1.52	.128	.230	−.245	.807	.039	2.96	.003	.473
Teaching Tasks quality of relationship	−.888	.375	.118	−2.71	.007	.385	.739	.460	.112	.019	.985	.003	−2.79	.005	.478
PSDQ mother-reported authoritarian parenting	.092	.927	.011	1.93	.054	.254	−1.39	.164	.188	1.57	.117	.234	5.90	<.001	.973
PSDQ father-reported authoritarian parenting	1.27	.217	.157	3.23	.001	.424	−1.31	.190	.180	4.45	.656	.068	4.35	<.001	.734
PSDQ mother-reported permissive parenting	−.043	.966	.006	2.02	.044	.271	−1.65	.099	.231	2.03	.043	.313	6.38	<.001	1.07
PSDQ father-reported permissive parenting	1.93	.054	.261	1.63	.104	.226	.704	.482	.097	3.50	<.001	.563	3.33	<.001	.562
Peabody Picture Vocabulary Test	1.20	.229	.143	−2.73	.007	.341	1.53	.128	.197	−.024	.981	.004	1.09	.278	.174
Expressive One Word Vocabulary Test	−.163	.871	.021	−1.98	.048	.249	2.69	.008	.360	.363	.717	.056	−2.50	.012	.412
Vineland Screener	−1.66	.098	.205	−4.70	<.001	.611	1.74	.083	.228	−.630	.529	.096	−4.67	<.001	.767
Lab-TAB positive emotionality	−1.42	.155	.173	−3.29	.001	.435	−1.53	.126	.204	−2.07	.039	.314	−.459	.647	.076
Lab-TAB negative emotionality	2.85	.005	.356	1.48	.140	.207	.700	.485	.096	2.28	.023	.375	1.47	.141	.256
Lab-TAB anger	2.48	.012	.310	2.29	.023	.304	−.952	.342	.129	−.396	.692	.063	2.61	.009	.440
Lab-TAB sadness	3.73	<.001	.462	1.22	.223	.172	.768	.443	.107	1.11	.268	.180	.819	.413	.144
Lab-TAB fear	.345	.730	.045	−.300	.765	.040	1.38	.167	.185	3.50	<.001	.553	.191	.849	.033
Lab-TAB global impulsivity	−1.80	.073	.230	3.50	<.001	.461	−1.85	.065	.256	−.121	.903	.019	4.18	<.001	.700
Behavioral inhibition z-score	5.52	<.001	.705	.041	.968	.005	2.06	.040	.281	1.78	.076	.280	1.84	.067	.311
CBQ surgency	−3.93	<.001	.501	2.99	.003	.414	.437	.744	.046	1.05	.296	.174	2.88	.004	.508
CBQ negative affectivity	3.12	.002	.422	2.06	.039	.281	1.41	.159	.196	−.084	.933	.014	4.71	<.001	.845
CBQ effortful control	−1.98	.049	.270	−6.43	<.001	.877	1.73	.083	.245	−3.76	<.001	.603	−4.79	<.001	.848
Age 18 Functional Outcomes
LSI Chronic Academic/Work Stress	1.31	.191	.180	4.99	<.001	.683	2.08	.038	.296	8.09	<.001	1.21	.269	.788	.050
LSI Chronic Interpersonal Stress	1.20	.231	.172	4.05	<.001	.553	4.56	<.001	.624	6.48	<.001	1.04	3.31	<.001	.578

Note. Independent samples t-test comparing parameter estimates between cluster A and each psychopathology cluster. Holm-Bonferroni corrected significant p-values are bolded. Red cells indicate significantly higher than cluster A while blue cells indicate significantly lower than cluster A. PSDQ = Parenting Styles and Dimensions Questionnaire; Lab-TAB = Laboratory Temperament Assessment Battery; CBQ = Children’s Behavior Questionnaire; LSI = Life Stress Interview. Cohen’s d ranges are small: 0.2–0.5; medium: 0.5–0.8; large 0.8–1.20.

Age 3 Predictors

Comparing cluster B (ANX) to cluster A (healthy), cluster B had significantly higher levels of father-reported dyadic adjustment, and higher levels of observed negative emotionality, anger, sadness, and behavioral inhibition. Cluster B also had significantly higher levels of mother-reported negative affect and significantly lower levels of surgency.

By contrast, Cluster C (ADHD), compared to cluster A (healthy), had parents with significantly higher rates of anxiety disorders; lower mother-reported dyadic adjustment; and higher parent-reported life stress affecting the child. Additionally, this cluster had a poorer observed quality of relationship with the parent and higher father-reported authoritarian parenting. Cluster C also had lower receptive vocabulary scores and lower levels of parent-reported social competence on the Vineland screener. This cluster exhibited lower levels of observed positive emotionality, higher observed impulsivity, higher mother-reported surgency, and lower mother-reported effortful control.

Cluster D (DEP/ANX) had parents with significantly higher rates of mood disorders and higher expressive vocabulary scores on the Expressive One-Word Vocabulary Test.

Cluster E (SUD/DEP) had parents with significantly lower education, as well as higher rates of SUD. Those in cluster E also experienced higher levels of life stressors, higher father-reported permissive parenting, greater fear during the Laboratory Temperament Assessment Battery, and lower parent-reported effortful control.

Cluster F (DBD group) had greater observed parental hostility and a poorer quality of parent-child relationship during the Teaching Tasks, as well as higher mother- and father-reported authoritarian parenting and mother- and father-reported permissive parenting. The cluster also had significantly lower expressive vocabulary and social competence scores. They also displayed significantly higher levels of observed anger and global impulsivity. This cluster had significantly higher parent-reported surgency and negative affect, as well as lower effortful control.

Age 18 Functional Outcomes

Cluster B (ANX) did not differ from cluster A (healthy) on academic/work and interpersonal functioning at age 18. In contrast, cluster C (ADHD) displayed poorer functioning in both the academic/work and interpersonal domains compared to cluster A. Cluster D (DEP/ANX) exhibited greater problems in interpersonal functioning. cluster E (SUD/DEP) experienced significantly poorer academic/work and interpersonal functioning. Cluster F (DBD) displayed greater impairment in interpersonal functioning.

Discussion

The importance of development, continuity, and course has long been emphasized in theory and research in psychopathology and developmental psychopathology (Bromet, 2015; Cicchetti & Rogosch, 2002), and is considered one of the criteria for judging the validity of diagnostic constructs (Robins & Guze, 1970). However, the development and course of psychopathology are seldom explicitly incorporated into classification systems (Klein et al., 2015; Tackett & Hallquist, 2022). To our knowledge, this study is the first to apply a longitudinal data-driven clustering algorithm to the range of common mental disorders. We applied this novel approach to rigorously collected data on six occasions from early childhood through late adolescence. The best fitting solution revealed six clusters with distinct patterns of psychopathology at each time point, as well as unique patterns of homotypic and heterotypic continuity across development.

The present study is consistent with prior work reporting both homotypic and heterotypic continuity of common mental disorders through adolescence and showing greater continuity among, than between, internalizing and externalizing psychopathology (Caspi et al., 2020; Copeland et al., 2013; Finsaas et al., 2018; Healy et al. 2022; Oldehinkel & Ormel, 2023). However, we also found evidence of subgroups of individuals with unique patterns of sequential comorbidity (e.g., childhood anxiety that diminished with age vs. childhood anxiety that persisted into adolescence along with the emergence of adolescent-onset depressive disorders vs. adolescent-onset depression and substance use disorders). Moreover, we found that the six clusters revealed in our analyses generally differed regarding sex distribution, early childhood predictors (i.e., parental education and psychopathology, early environment, temperament, cognitive and social functioning) and late adolescent outcomes (i.e., functional impairment). Thus, this approach accounted for comorbidity and change in symptom presentation, creating more homogenous subgroups of transdiagnostic psychopathology that were associated with different antecedents and outcomes.

However, it is worth noting that the optimal cluster solution is derived from a particular prospective longitudinal dataset - this exact solution may not be replicated in other datasets using different samples, age groups, and measures. For example, the sample’s age influences the prevalence of particular disorders (e.g., we would not expect to find much substance use before adolescence). Additionally, the prevalence of disorders will differ depending on whether the sample is selected from the community or a clinical setting. Our community sample has very low rates of psychosis and eating disorders, which might be more common in a clinically referred sample. Thus, we regard this more as a proof of principle rather than as a definitive classification. Specifically, we argue that longitudinal clustering of psychopathology affords unique insights into patterns of continuity, sequential comorbidity, and developmental patterning not otherwise considered by cross-sectional classification systems (Lahey et al., 2014).

Cluster B (ANX) can be thought of as reflecting fear-related internalizing psychopathology, whereas cluster D (DEP/ANX) reflects distress-related internalizing psychopathology. Importantly, these clusters differ in that, compared to cluster A, cluster B displayed higher rates of anxiety disorders at earlier ages and was characterized by early temperamental negative affect and behavioral inhibition according to both observational and parent-report measures. However, cluster B did not exhibit notable comorbidity, and rates of anxiety disorders declined in adolescence, suggesting that this group is the healthiest of the psychopathology clusters. Indeed, this cluster did not differ from cluster A (healthy) on parent psychopathology and parenting, or functional outcomes at age 18. Consistent with this, the 5 factor solution did not produce a cluster like cluster B; most participants in cluster B were included in the healthy cluster in the 5 factor solution.

Cluster D, by contrast, displayed gradually rising rates of anxiety disorders throughout adolescence, and rates of depressive disorders that increased sharply in mid-late adolescence, suggesting a later onset but possibly more pernicious course of psychopathology than cluster B. Cluster D had higher early expressive vocabulary, a higher rate of parental mood disorders, and poorer functional outcomes than cluster A. When directly comparing cluster B and cluster D (see Supplementary Table 1), cluster D had lower father-reported dyadic adjustment; higher parent-reported social competence; lower observed and parent-reported behavioral inhibition; and lower observed anger. Cluster D also had higher mother-reported surgency, higher mother-reported effortful control, and greater impairment in interpersonal functioning at age 18. Thus, cluster D has generally better functioning that declines in adolescence, while cluster B has worse functioning in childhood that improves in adolescence. These differences further strengthen the distinction between two clusters, despite their diagnostic overlap.

Cluster E (DEP/SUD) was also characterized by increasing depression but was quite different from cluster D (DEP/ANX) in that it was additionally marked by substance use disorders, rather than anxiety disorders. Thus, it reflects a mixed internalizing/externalizing presentation. Depression and substance use disorders are frequently comorbid (Swendsen & Merikangas, 2000), and it is posited that those with depression might use substances as a form of self-medication or an attempt to cope with negative mood or stressful life events (Magee & Connell, 2021). Alternatively, substances can lead to depression through their biological effects or indirectly via functional consequences (Boden & Fergusson, 2011). Indeed, in this study, we found that cluster E had the most significant associations with impaired academic and interpersonal functioning, in comparison with cluster A (healthy). Importantly, compared to cluster A, cluster E was also associated with higher levels of stress and temperamental fear in early life despite low rates of psychopathology at earlier ages that did not intensify until age 12. This cluster might reflect the internalizing (as opposed to externalizing) pathway to substance use disorders as discussed by Chassin et al. (2013). Notably, whereas cluster D was associated with a higher rate of parental depression than cluster A, cluster E had a higher rate of parental substance use disorder than cluster A, suggesting some specificity in familial etiological influences. This, using cluster A as the reference group, clusters D and cluster E were distinguished by several early childhood risk factors, despite both being characterized by increasing rates of depression in adolescence. When clusters D and E were compared directly (see Supplementary Table 3), cluster E was characterized by higher mother-reported permissive parenting, lower mother-reported effortful control, and greater academic impairment at age 18. Again, these differences underscore the utility of our longitudinal clustering approach in distinguishing phenotypes that can appear similar at single points in time but are very different from the perspective of risk factors and developmental course.

Clusters C (ADHD) and F (DBD) were the two externalizing clusters. Cluster C exhibited low levels of psychopathology in early childhood which intensified through later childhood and adolescence, manifesting mainly as ADHD with some co-occurring anxiety and depressive disorders and disruptive behavior disorders. Cluster F, by contrast, was characterized by high rates of disruptive behavior disorders in early childhood, which declined but showed a small increase again in adolescence. Those in cluster F also had modest but elevated rates of other disorders over the course of development. Despite the difference in the nature of their symptomatology and the onset and course of psychopathology, compared to cluster A (healthy) at age 3, clusters C and F both displayed a poorer observed quality of parent-child relationship, higher levels of father-reported authoritarian parenting, lower social competence, greater observed impulsivity, lower parent-reported effortful control and greater parent-reported surgency, and poorer verbal abilities (although they differed on whether the problems were receptive or expressive). Additionally, cluster C was marked by a history of anxiety disorders in parents, consistent with some prior literature finding familial coaggregation of ADHD and anxiety (Jarrett et al., 2016). Cluster F, by contrast, was associated with more problematic parenting styles and temperamental anger in early childhood compared to cluster A, consistent with the high rate of disruptive behavior disorders in this cluster. When comparing them directly (see Supplementary Table 2), cluster C and F differed in that cluster C had lower expressive vocabulary abilities at age 18, while cluster F had higher mother-reported authoritarian and permissive parenting and mother-reported negative affect. These differences perhaps point to cluster C having greater verbal learning difficulties, while cluster F was marked by greater mood dysregulation in early childhood. Regarding functional impairment, cluster C displayed poorer academic functioning at age 18 compared to both cluster A and cluster F. This makes sense given the more problematic executive and language functioning in this cluster and the fact that their high rates of psychopathology persisted into late adolescence. Cluster F, on the other hand, did not show later functional impairment, and it appears that their disruptive behavior disorders had largely remitted by late adolescence. In our community sample, we perhaps lack sufficient power to detect a small subgroup with early behavior problems and poor functional outcomes that corresponds to Moffitt (1991)’s life course-persistent subtype of conduct problems. Indeed, a 7 cluster solution produced clusters that were similar to Moffitt’s typology, but the sample sizes were small, and overall model fit was poorer than for the 6 cluster solution.

Overall, the distinction among the derived clusters was supported by a variety of early childhood risk factors and late adolescent functional outcomes. This study is unique in its explicit consideration of developmentally-based transdiagnostic subgroups with meaningful clinical, family, and temperamental correlates. The utilization of longitudinal k-means clustering to psychiatric diagnostic data is particularly novel and has utility for future research using syndrome or symptom-level characterization of psychopathology. While our clusters align for the most part with the existing literature, they provide a richer and more detailed perspective, illustrating different patterns of the development of psychopathology and the unfolding of sequential comorbidity and heterotypic continuity over distinct developmental periods.

However, the present study had several limitations. First, the sample was predominantly White and non-Hispanic and cannot be assumed to generalize to other ethnoracial groups, as the relative homogeneity of the sample precluded an adequately powered analysis of racial and ethnic differences between clusters. Given evidence that rates of psychopathology and many clinical features (e.g., age of onset, severity, course, comorbid conditions) differ by sociodemographic characteristics, greater attention to sampling strategies and ethnoracial diversity is needed to understand how developmental patterning of psychopathology differs across groups (Wilson, 2024). Second, we assessed psychopathology every three years. Triennial assessments may have limited precision, particularly early in development when change is rapid. Thus, we may have missed some onsets and offsets of episodes of psychopathology, particularly during the age 3 and 6 assessments when we focused on the 3 months prior to the interview instead of assessing the entire interval, as we did in subsequent waves. Third, some of the measures (i.e., DBD at age 9, rating of Parent Confidence in the Teaching Tasks) had interrater reliabilities in the upper end of the “fair” range, based on Shrout’s (1998) recommendations. Lower interrater reliability could affect the prevalence of disorders and attenuate prediction by age 3 variables. However, almost all measures had reliabilities that were in the moderate-substantial range. Fourth, the severity of psychopathology was not considered. Clusters were derived based on the presence or absence of diagnoses at each time point. The incorporation of dimensional measures of psychopathology might provide additional nuance or even more distinctive groups. Future research utilizing a dimensional approach to capture variability in symptom trajectories is warranted. Fifth, although this study emphasizes person-centered trajectories to understanding the patterning of homotypic and heterotypic continuity, the interpretation of each cluster is based on the proportion of individuals in that cluster with the specific diagnosis at that time point. Thus, some heterogeneity remains in each cluster, as cluster members do not exhibit a uniform developmental course. Finally, the validation measures were assessed at either the age 3 or age 18 waves, two time points that were included in the clustering algorithm. Hence, they are not entirely independent of the clusters. This likely inflated some of the associations. For example, effect sizes of the association between maladaptive parenting and cluster F assignment were extremely large, very possibly because the parenting measures were derived at age 3 and this cluster was characterized by an elevated rate of disruptive behavior disorders at age 3. Thus, the present cluster solution is offered as illustrative, rather than definitive, and provides a proof of principle that we hope will stimulate further work in longitudinal classification.

In conclusion, the present study suggests that a data-driven person-centered approach to classifying psychopathology that includes development and course over time may be useful in identifying meaningful transdiagnostic groups. These groups differed on preschool assessments of parental education, family history of psychopathology, parenting and life stress, and preschool temperament, cognitive and social development, as well as on functional impairment at 18 years of age. The consideration of development and course in psychiatric classification may be crucial to address issues of heterogeneity and comorbidity and account for the patterning of homotypic and heterotypic continuity (Tackett & Hallquist, 2022).

Figure 2. Heatmap of Effect Sizes for Post-Hoc Comparisons of All Psychopathology Clusters with Cluster A.

Note. Dep = Any depressive disorder including NOS; Anx = Any anxiety disorder including NOS; ADHD – Attention Deficit/Hyperactivity Disorder; DBD = Disruptive Behavior Disorder; SUD = Substance Use Disorder; PSDQ = Parenting Styles and Dimensions Questionnaire; Lab-TAB = Laboratory Temperament Assessment Battery; CBQ = Children’s Behavior Questionnaire; LSI = Life Stress Interview. Red cells indicate higher than cluster A while blue cells indicate lower than cluster A. Effect sizes are expressed using Cohen’s d (small: 0.2–0.5; medium: 0.5–0.8; large 0.8–1.20).