Exploring Transdiagnostic Mechanisms of Youth Externalizing Psychopathology: A Longitudinal Person-Centered Approach

Abstract

The present study utilized an exploratory data analysis approach to consider how executive functioning (EF) relates to the developmental course of externalizing psychopathology using the Oregon ADHD-1000 dataset. Multinomial logistic regressions of EF domains (working memory, processing speed, set shifting, reaction time variability, response inhibition, and vigilance) in predicting symptomatic classes and longitudinal pathways from a recent latent transition analysis (Smith et al., in press) were conducted; predicted probability figures were interpreted. Findings suggest that hyperactivity/impulsivity (HI) was most related to EF impairment in many domains. Inattention contributed to set shifting and processing speed impairment specifically. HI and oppositionality were very aligned with one another in childhood and diverged in adolescence. Youth who were HI in childhood and inattentive in adolescence were distinct in EF impairment from youth who were inattentive across development. Findings reiterate the importance of exploratory, person-centered, longitudinal approaches for understanding heterogeneity, comorbidity, and developmental psychopathology.

Attention-deficit/hyperactivity disorder (ADHD) is the most common mental health disorder of childhood (3–10%; Merikangas et al., 2009; Polanczyk et al., 2015). ADHD is classically heterogeneous and highly comorbid with disruptive behavior problems (DBPs), such as oppositional defiant disorder (ODD) and conduct disorder (CD), occurring among approximately 30–50% of ADHD cases (Maughan et al., 2004; Reale et al., 2017). The importance of etiological research has become increasingly apparent in recent years, as underscored by NIH initiatives (Insel et al., 2010). Specifically, most prominent etiological theories of ADHD, ODD, and CD (Cavanagh et al., 2017; Frick & Nigg, 2012; Pievsky & McGrath, 2018; Sonuga-Barke, 2002) emphasize the role of cognitive and emotional processes. However, ADHD’s notable heterogeneity and high rate of comorbidity generates significant challenges in understanding the disorder’s etiology. While recognition of this problem has led to more transdiagnostic research (Aldao et al., 2016; Sauer-Zavala et al., 2017), this work commonly suffers from two limitations: 1) the literature has largely ignored the role of comorbidity in giving rise to ADHD’s heterogeneity, and vice versa (Feczko et al., 2019); and 2) the current literature consists mostly of cross-sectional studies, which limits examination of how symptoms and their putative mechanisms may develop over time (Karalunas & Nigg, 2020).

A recent paper (Smith et al., in press) explored how a person-centered, longitudinal approach such as latent transition analysis (LTA) can help advance understanding of the developmental course of heterogeneity and comorbidity in externalizing psychopathology. Such an approach allows for an examination of how symptoms co-occur across development and how individuals commonly transition between symptom sets over time. The ideal class solutions at each age¹ were identified as five classes at age 9, four classes at age 12, and five classes at age 15, with a corresponding LTA. At age 9, Class (i.e., “C”) 1 consisted of typically developing (i.e., TD) youth; C2 consisted of youth likely to have symptoms of inattention (i.e., IN) and hyperactivity/impulsivity (i.e., HI), or “IN+HI”; C3 consisted of “IN+HI+ODD” youth; C4 consisted of “IN” youth; and C5 consisted of youth with a moderate probability of endorsing IN and HI, or a subthreshold class (i.e., “ST”). At age 12, C1 was TD; C2 had inattention and moderate representation of HI (i.e., “IN+modHI”); C3 was represented by inattention and moderate representation of HI and ODD (i.e., “IN+modHI+ODD”); and C4 included youth with a moderate representation of inattention symptoms (i.e., “ST/IN”). At age 15, C1 was “TD”; C2 was “IN+HI”, C3 was “IN+modHI+ODD”; C4 was “IN”; and C5 included individuals with moderate levels of HI as well as some ODD symptoms, or “modHI+lowODD”. See Figure 1. The most common LTA paths over time included (three-digit numbers indicate the class assignment at ages 9, 12, and 15, respectively): persistently TD (111); adolescent IN (114); typical ADHD (224); desisters (324); persistently oppositional (333); worsening (442); persistently IN (444); and mixed trajectory (524); see Figure 2. See Smith and colleagues (in press) for a comprehensive overview of these findings. The present study expands upon these foundational phenotypic findings to examine putative etiological mechanisms underlying the observed latent classes and trajectories.

Figure 1.

Final Latent Class Analysis Solutions at Ages 9, 12, and 15 (Rows A-C)

Note. D=difficulty. Figures from Smith and colleagues (in press); please see manuscript for full information.

Figure 2.

Diagram of all Latent Transition Probabilities between Latent Classes at Ages 9, 12, and 15 and Most Common Latent Transition Analysis Paths

Note. TD=typically developing; IN=inattention; HI=hyperactivity/impulsivity; ST=subthreshold; mod=moderate. Percentages displayed on lines between the columns for each age represent the percentage of individuals from within the prior (i.e., left) latent class who transitioned to the indicated (i.e., right) latent class. Percentages indicated on the right side of the figure indicate the percentage of the total sample in a given latent transition path, as depicted by the pair of lines in the indicated color and the corresponding numbers. Three-digit numbers indicate the class assignment at ages 9, 12, and 15 in the latent transition pathway. 111=persistently TD; 114=adolescent IN; 224=typical ADHD; 324=desisters; 333=persistently oppositional; 442=worsening; 444=persistently IN; 524=mixed trajectory. Figure from Smith and colleagues (in press); see manuscript for details.

The present study will explore how person-centered, longitudinal analytic approaches can help elucidate mechanisms of heterogeneity and comorbidity in youth externalizing psychopathology through an exploratory data analysis (EDA) approach (Tukey, 1977). EDA is not one set of procedures, but rather, an approach characterized by seeking to understand patterns conveyed by the data, often relying heavily upon data visualization, sometimes without use of hypotheses or formal tests for significance (Jebb et al., 2017). While EDA has often been conflated with questionable research practices such as p-hacking, fishing, or hypothesizing after results are known (i.e., HARKing), it is in fact the use of confirmatory data analysis to analytically “confirm” exploratory findings that leads to these issues (Fife & Rodgers, 2022; Jebb et al., 2017). By contrast, EDA is intended to lead to identification of potentially meaningful phenomena for future confirmatory testing (Jebb et al., 2017). Importantly, the same methodological limitations that obscure our understanding of the development of symptoms (i.e., methods that parse and control for shared variance between diagnostic groups) may also obscure our ability to understand the relationships between symptoms and mechanisms, described further below, which informed the analyses that were conducted to explore transdiagnostic mechanisms of externalizing psychopathology in the present study.

First, person-centered approaches may be helpful in exploring mechanisms shared between clusters of symptoms. Despite extensive theoretical discussion in the literature that the DSM has parcellated what could be meaningful, etiologically distinct entities into multiple excessively comorbid diagnostic labels and/or combined multiple disease processes under single diagnoses that are excessively heterogeneous (e.g., Caron & Rutter, 1991; Hyman, 2010; Lilienfeld & Treadway, 2016), the literatures on ADHD heterogeneity and DBP comorbidity remain largely separate. Many studies of ADHD recruit pure ADHD samples or statistically control for comorbidity (Beauchaine et al., 2010; Caron & Rutter, 1991). On the other hand, many studies of comorbidity neglect to account for heterogeneity by comparing separate groups or factors that correspond to separate diagnoses (Beauchaine & McNulty, 2013; Feczko & Fair, 2020). These methods may create artificial groups that do not exist in the general population (e.g., attempting to statistically remove the contribution of ODD symptoms to the presentation of the ADHD group) and obscure the common variance that is of key interest to understanding what is shared between disorders (Beauchaine et al., 2010; Beauchaine & Zisner, 2017; Beauchaine & Hinshaw, 2020). These theoretical and methodological issues may contribute to the notable heterogeneity and overlap observed in the literature examining cognitive processes in ADHD, ODD, and CD (e.g., Fair et al., 2012; Frick, 2016; Ghosh et al., 2017; Kofler et al., 2019). While ADHD is often associated with EF deficits at the group level (Pievsky & McGrath, 2018), children with ADHD differ in the domain of their EF impairment, and some youth exhibit no EF deficits at all (Kofler et al., 2019). Further, EF deficits are also present in some children with ODD (Ghosh et al., 2017) and CD (Frick, 2016). Some posit that these symptoms confer additional impairment, while others argue that detected impairments are most likely attributable to the presence of ADHD (Fairchild et al., 2019; Ghosh et al., 2017). In other words, literature focused on only heterogeneity of one disorder or comorbidity of disorders treated as homogenous and distinct neglects to account for how specific symptoms across disorders may relate to one another and share mechanisms (Feczko et al., 2019; Lilienfeld & Treadway, 2016).

Second, many studies focusing on mechanisms of ADHD heterogeneity and DBP comorbidity are cross-sectional. Just as our current nosology may create excessive heterogeneity and comorbidity, it similarly creates apparent equifinality and multifinality. Specifically, youth with different deficits in specific candidate mechanisms may have the same diagnosis (i.e., the symptoms are presumed to be equifinal) while other youth with the same deficits on those same candidate mechanisms may be given different diagnoses given the degree of diagnostic overlap (i.e., various presentations are assumed to be multifinal; Beauchaine & McNulty, 2013). When investigating a single timepoint, it cannot be evaluated whether differences in symptoms or putative etiologies reflect potentially meaningful distinctions or whether it may instead reflect groups of children at different points in the developmental course of one diagnostic pathway (Beauchaine & McNulty, 2013). Further, most longitudinal studies are diagnostic retention studies, considering how many youth retain diagnoses or develop new diagnoses over time. Longitudinal studies that are reliant upon a priori DSM classification may reach similar conclusions (e.g., mechanism X at time 1 led to disorders A and B at time 2, or mechanisms X and Y at time 1 led to disorder B at time 2, even though heterogeneity and comorbidity limits the extent to which youth with “A” are similar and the extent to which youth with “A” and “B” are different). Indeed, the literature on these issues is also inconsistent, likely due in part to methodological issues. Some argue that given the vast heterogeneity present in symptomatic presentation and developmental course, there are likely meaningfully distinct pathways (Steinberg & Drabick, 2015) and that disorders co-occur due to shared risk factors (e.g., both disorders share EF deficits; Ghosh et al., 2017). Others argue that these disorders are related but unique phenomena that increase risk for the development of one another in sequence (Rowe et al., 2010). Further still, some argue that externalizing symptoms are heterotypically continuous, particularly the pathway from HI to ODD (Beauchaine & McNulty, 2013), as evidenced by the typical ages of onset and the fact that children with ADHD are more than ten times as likely to develop DBPs than children without ADHD (Steinberg & Drabick, 2015) However, this does not account for youth with ADHD who do not develop DBPs (Nock et al., 2007; Rowe et al., 2010). Exploratory longitudinal methods that consider how the developmental course of symptoms may be related to different etiological mechanisms are needed to explore these theories of developmental psychopathology.

While there is a robust literature dedicated to understanding mechanisms of ADHD heterogeneity, often utilizing latent profile/class analyses, the majority of such studies are not longitudinal, have not accounted for potential mechanisms shared across ODD and CD, and/or examine DSM disorders present among mechanistically-derived profiles (e.g., Arnett & Flaherty, 2022; Fair et al., 2012; Goh et al., 2020; Martel et al., 2010; Martel, 2016; Morris et al., 2023; Rajendran et al., 2015; Shroff et al., 2024; van Hulst et al., 2015). By contrast, the present study is focused on identifying clusters of symptoms, exploring the mechanisms present among profiles, and considering whether this approach encapsulates more homogenous mechanisms than would be present among DSM-based groupings. Further, while there have been latent approaches dedicated to examining symptoms across disorders (Acosta et al., 2008; Ametti et al., 2022; Elia et al., 2009; Halldorsdottir et al., 2015; Neuman et al., 2001; Ostrander et al., 2008; Racz et al., 2023; Rosa-Justicia et al 2020; Yue et al., 2022), these studies were not longitudinal and/or did not examine mechanisms. Work that has examined mechanisms across externalizing disorders have largely been cross-sectional and utilized variable-centered analyses based on diagnoses (e.g., Antonini et al., 2015; Barkley et al., 2001; Barnett et al., 2009; Baving et al., 2006; Brocki et al., 2007; Connor & Doerfler, 2008; Ezpeleta & Granero, 2015; Glenn et al., 2017; Hummer et al., 2011; Humphreys & Lee, 2011; Kara et al., 2017; Kleine Deters et al., 2020; Leno et al., 2018; Luman et al., 2009; Mayes et al., 2009; Noordermeer et al., 2020; Oosterlaan et al., 2005; Ter-Stepanian et al., 2017; Van Goozen et al., 2004). With regards to longitudinal work, many studies have examined how symptoms of one disorder may confer risk for another, but have not considered mechanisms (Allmann et al 2022; Bolhuis et al., 2017; Burns & Walsh, 2002; Harvey et al., 2016; Mustonen et al., 2023; Whelan et al., 2013). Some studies have utilized latent class growth analysis or similar models, which typically examine the latent trajectories of total continuous scores of a diagnostic construct. Some recent work has utilized longitudinal cross-lagged panel network models, but have either not examined mechanisms (Zhang et al., 2024) and/or have not examined clustering of individual items (i.e., examined subtypes or diagnoses; Freichel et al., 2024; Karalunas et al., 2021). The only extant LTA of youth externalizing psychopathology (Villodas et al., 2015) identified “well-adjusted”, “hyperactive/oppositional”, and “aggressive/rule-breaking” classes in youth aged 4, 8, and 12, but did not examine mechanisms.

In sum, the present study is the first longitudinal person-centered approach which examines profiles of individual ADHD, ODD, and CD symptoms, and the putative EF mechanisms of these profiles, from childhood to adolescence. More specifically, this study is intended to utilize an EDA approach to consider how EF relates to youth externalizing symptomatology when using an analytic approach that examines how symptoms co-occur and develop over time, without use of analytic methods that obfuscate associations between symptoms and mechanisms by reliance on a priori diagnostic grouping and control for variance that is likely meaningful to understanding mechanisms of comorbidity. Due to our EDA approach, we do not have specific hypotheses. Rather, the present study will build upon the LTA conducted in Smith and colleagues (in press) by including EF measures as predictors of class membership and latent transition paths to consider if this analytic method yields unique insights about the mechanisms of the developmental course of ADHD and DBPs and to explore relevant theoretical and quantitative implications.

Transparency and Openness

Preregistration

The present study was not preregistered.

Data, Materials, Code, and Online Resources

Code and syntax used for the present study’s data management and analyses is included within the following Open Science Framework (OSF) link: https://doi.org/10.17605/OSF.IO/F5SZU. Further, additional tables and figures are available in the supplemental materials. For further information about data, code, or materials, contact J. Smith.

Reporting

We believe that our methods are rigorous, that they were conducted with integrity, and are presented transparently. The present study utilizes data from the Oregon ADHD-1000 dataset (Nigg et al., 2023). Please note that the specific information provided throughout the Method section reflects only the data relevant to the present study at the time the data was acquired and was based, in part, upon direct communication with the study team at Oregon Health & Science University (OHSU) with regards to the most recent information, as data collection and processing were still ongoing at the time this study was conducted. Please see Nigg and colleagues (2023) for full information about the finalized, publicly available Oregon ADHD-1000 dataset. We report how we determined the subsample used in the present study, the specific measures used in our secondary data analysis, and all manipulations to utilized variables.

Ethical Approval

The protocol from which the present study derived its data, as well as the secondary data analysis, were each approved by an institutional review board (i.e., OHSU and Florida International University, respectively) carried out in accordance with the provisions of the World Medical Association Declaration of Helsinki.

Method

Procedures and Participants

As stated above, the data described herein are from the Oregon ADHD-1000 dataset (Nigg et al., 2023), and secondary data analysis was approved by the institutional review board at Florida International University. In brief, families were recruited from the community (within a 50-mile radius from OHSU) with ADHD deliberately oversampled and with symptoms of ODD and CD free to vary. A sample of n=2,144 inquiries were screened by phone to establish initial eligibility and interest. Next, an on-site clinical evaluation was conducted for n=1,450, after which best-estimate research diagnoses and final eligibility were established by a team of two clinicians. After the diagnostic screen, n=104 eligible participants withdrew due to lack of further interest, and n=31 excluded participants ultimately participated, either for sub-study needs or for rapport (i.e., sibling of an included participant). Ultimately, the present study sample includes n=849 at baseline, a subset of which (target n=610) were followed up by design. In addition to this planned attrition, some annual waves were missed for some individuals. Please see Smith and colleagues (in press) and Nigg and colleagues (2023) for additional information regarding recruitment procedures, exclusion criteria, and data management.

The total baseline sample (n=849) endorsed a primary race of 85.4% white/Middle Eastern, 6.7% Black, 5.1% Asian/East Indian, 1.8% American Indian/Alaska Native/Eskimo, 0.7% Native Hawaiian/Pacific Islander, and 0.4% declined to answer/not known. Additionally, 9.8% endorsed additional racial identities. The baseline sample was 6.1% Hispanic/Latino/a. Lastly, the baseline sample was 61.8% male and 38.2% female.

With regards to the highest education obtained by the participant’s parents, 34.0% had a bachelor’s degree, 26.4% had a master’s, law, or other 2–3-year degree, 16.1% had some college education but no degree, 9.9% had a doctorate, Ph.D., or medical degree, 9.8% had an associate’s degree, 2.5% had a high school degree or equivalent, and 0.1% had some high school education but no degree (1.2% of participants did not respond). Regarding income, 8.1% earned <25.000, 5.9% earned <35,000, 10.5% earned <50,000, 20.1% earned <75,000, 21.0% earned <100,000, 14% earned <130,000, 5.4% earned <150,000, and 8.1% earned >150,000 (6.8% declined to answer or did not know).

Measures

Kiddie Schedule for Affective Disorders and Schizophrenia (KSADS)

Latent classes were derived using parent-report of current ADHD, ODD, and CD items from the KSADS (Kaufman et al., 1997). These individual items were dichotomized (i.e., symptom endorsed [1] or not endorsed [0], such that subthreshold and threshold responses counted as endorsement).

Measures of Executive Function

See Nigg and colleagues (2023) for full details on the EF tasks administered. In short, the present study utilized the following measures: visuospatial working memory (CANTAB Spatial Span backwards total items correct; higher scores indicate better performance); vigilance (identical pairs continuous performance task d’1 catch trials; higher scores indicate better performance); set shifting (D-KEFS Trail Making Test condition 4 completion time; lower scores indicate better performance); response inhibition (Stop Task stop signal reaction time calculated using the Verbruggen method; lower scores indicate better performance); processing speed (Color-Word color naming control condition completion time; lower scores indicate better performance); and reaction time variability (standard deviation of the reaction time on Stop Task “go” trials; lower scores indicate better performance).

Data Analytic Plan

In Smith and colleagues (in press), individual latent class analyses (LCAs) and a corresponding LTA were conducted at ages 9 (n=398), 12 (n=265), and 15 (n=179) in Mplus 8.9 (Muthén & Muthén; 1998–2017). Before LCAs were conducted, data management was conducted in SPSS. Specifically, KSADS items with n<5% endorsement were removed. As a result, all 18 ADHD items were retained at all ages; 7 of 8 ODD items were retained at all ages (only excluding ODD #8: “spiteful/vindictive”); and 1 of 15 CD items was retained at ages 9 and 12, but not 15 (only retaining CD #1: “bully”). Ultimately, chosen latent class and latent transition solutions were run, and class assignments were hard-classified (see Figure 1, Figure 2, and Smith et al., in press for full details). Next, multinomial logistic regressions with task performance predicting concurrent latent class membership at each age, as well as age 15 task performance predicting experienced symptomatic trajectory (i.e., the eight most common longitudinal latent pathways from 9 to 15), were conducted in R 4.3.0. See Supplemental Table S1-S18 for sample sizes of each task in each class and Supplemental Table S19 for sample sizes of LTA paths and for each task in each LTA path. Lacking a theoretical justification for simultaneously or hierarchically entering covariates into a regression model and to avoid control for potentially meaningful shared variance, each covariate was analyzed separately. Importantly, figures for predicted probabilities will be presented given the limited interpretability of multinomial logistic regression logit coefficients and consistent with the spirit of our EDA approach.

A Note on Interpretation

As previously discussed, our EDA approach is focused on what patterns may be revealed by the data using a novel analytic approach rather than testing specific hypotheses. As such, the results section will focus on describing what was revealed by a series of multinomial logistic regression predicted probability figures. Specific p values are not provided, not only because it is inconsistent with our EDA approach, but also because without specific hypotheses the number of analyses that would need to be conducted is prohibitive (in multinomial logistic regression, statistical significance is provided only in comparison to one chosen reference class at one given score of the dependent variable at a time; for example, 156 p-values would need to be examined to compare every class at just one value of each dependent variable). However, we understand that given existing standards in the field, readers may desire some approximation of differences that are likely to be significant. Thus, we utilized an “inference by eye” approach with confidence intervals (Cumming, 2009) using the MNLpred R package (Neumann, 2021) for the multinomial logistic regressions with task performance predicting concurrent latent class membership at each age. In simulation studies, when confidence interval regions overlap by approximately half of the length of one confidence interval bar, the p value was virtually almost always between p=.03 and p=.05 (Cumming & Finch, 2005). Results where this occurred are indicated with a footnote. Importantly, it should be emphasized that for predicted probabilities, a given height on the Y axis represents both sample size of the latent class and the relationship between the latent class and the dependent variable. Classes may demonstrate different relationships with the dependent variable, yet overlap greatly only due to where lines converge because of sample sizes (or conversely, not overlap due only to sample size differences, without meaningful distinctions in slope). Statistical significance is not the focus of this study, and inference by eye did not change our interpretation of patterns. Confidence intervals and their associated footnotes should only be used to inform interpretation and not as a tool for dichotomous decision making about whether a result is meaningful (Cumming & Finch, 2005). Rather, we encourage future studies to follow up and investigate specific hypotheses with confirmatory methods based upon patterns revealed by this study’s EDA. See Figures 3–4 for predicted probabilities and Supplemental Figure S1-S18 for predicted probabilities with confidence intervals.

Figure 3.

Multinomial Logistic Regression Probabilities of Latent Class Membership Predicted by Tasks of Executive Function

Note. Small back lines on the bottom of each figure show the distribution of scores on the predictor.

Figure 4.

Multinomial Logistic Regression Probabilities of Latent Transition Path Membership Predicted by Tasks of Executive Function at Age 15

Note. 111=persistently TD; 114=adolescent IN; 224=typical ADHD; 324=desisters; 333=persistently oppositional; 442=worsening; 444=persistently IN; 524=mixed trajectory. Small back lines on the bottom of each figure show the distribution of scores on the predictor.

In our approach, we attended to which latent classes have a distinct slope or trend with regards to the dependent variable (e.g., groups that are much more represented as scores are increasingly impaired on a given task) versus classes that did not. Additionally, we attended equally to where classes converged and diverged in their patterns (e.g., if classes A and B are similar to one another and both differ in slope from classes X, Y, and Z, we considered what symptoms are shared by classes A and B that are unique from the symptoms of classes X, Y, and Z). Authors largely attended to differences and similarities between classes at the impaired end of the dependent variable. Additionally, focus was placed on comparing clinical classes (i.e., excluding C1, a TD class at all ages). While all differences between clinical classes according to inference by eye were footnoted, there were many clearly significant differences with C1 that were not footnoted, as it is less of interest that youth without symptoms differed from youth with symptoms. C1 was only mentioned in the event it was useful for interpretation to highlight a TD comparison. Interpretations were made by two authors, discrepancies were discussed, and consensus was reached.

Results

See left column of Figure 3 for predicted probability figures at age 9. For the standard deviation of go reaction time², d’1 catch trials³, and trail making condition 4 completion time⁴, IN+HI and IN+HI+ODD classes were generally similar to one another and most represented at impaired ranges of EF tasks, while IN and ST classes were generally similar to one another and approximately equally represented at all levels of the dependent variable. For d’1 catch trials and trail making condition 4 completion time, it appears that hyperactivity/impulsivity was largely responsible for worse performance on these variables (i.e., vigilance, set shifting) as it was the only set of symptoms shared between C2 (IN+HI) and C3 (IN+HI+ODD) that are unique from C4 (IN) and C5 (ST). Neither inattention nor oppositionality appeared to contribute to worse scores on these tasks given C4 (IN)’s lack of slope and given that C3 (IN+HI+ODD) did not demonstrate additional representation at the impaired end of the variables beyond C2 (IN+HI). However, for standard deviation of go reaction time (reaction time variability), IN+HI+ODD demonstrated a steeper slope in the impaired range than IN+HI, suggesting additional contribution of oppositionality. Further, there were a few EF domains wherein inattention appeared to contribute to worse performance. Specifically, C4 (IN) had the steepest slope for color naming (processing speed), and inattention is the only characteristic shared between all classes with greater representation at the impaired range (i.e., IN+HI, IN+HI+ODD, and IN classes). Additionally, inattention appeared to contribute somewhat to worse scores on spatial span backwards (visuospatial working memory), though hyperactivity/impulsivity contributed further impairment (i.e., steeper slope) for IN+HI and IN+HI+ODD classes. Lastly, inattention appeared to contribute to stop signal reaction time (inhibition) impairment due to the similar slope between IN+HI and IN classes, while oppositionality appeared to confer additional impairment in the IN+HI+ODD class⁵.

See middle column of Figure 3 for predicted probability figures at age 12. Overall, groups did not differ as clearly in their relationships with EF domains as they did at age 9; rather, all clinical groups had a similar slope, such that all clinical groups were more represented at the impaired end of the dependent variables (including spatial span backwards, color naming, trail making condition 4 completion time, standard deviation of go reaction time, d’1 catch trials); classes appeared to differ primarily due to sample size differences. These similar slopes could indicate that inattention is related to EF impairment at this age as it is the only symptom shared by all three clinical classes (i.e., IN+modHI; IN+modHI+ODD; ST/IN); alternatively, findings may simply reflect that youth with any symptomatology perform worse compared to typically developing youth. As differences between clinical groups are most of interest to the present study, we provide minimal interpretation of most age 12 findings. The only exception was stop signal reaction time (inhibition), on which hyperactivity/impulsivity appeared to contribute uniquely to impairment for C2 (IN+modHI) and C3 (IN+modHI+ODD), which both had steeper slope than C4 (IN), while C2 and C3 had similar slopes to one another.

See right column of Figure 3 for predicted probability figures at age 15. It is worth noting that sample sizes are smallest at age 15; interpretation of the implications and reliability of results should be tempered accordingly. C2 (IN+HI) and C3 (IN+modHI+ODD) were again generally similar in their relationships with many EF domains, and therefore, hyperactivity/impulsivity was determined to be important for EF impairment. For example, IN+HI and IN+modHI+ODD classes were highly similar in slope on stop signal reaction time (i.e., inhibition) and were most represented at impaired scores (in contrast to age 9, oppositionality did not confer additional impairment). However, at impaired ranges of spatial span backwards (visuospatial working memory), color naming (processing speed), trail making condition 4 (set shifting), and d’1 catch trials (vigilance), C2 (IN+HI) was slightly more represented than C3 (IN+modHI+ODD). These findings especially emphasize the role of hyperactivity/impulsivity, as these classes symptomatically diverged at age 15 such that C3 (i.e., IN+modHI+ODD) had lower hyperactivity/impulsivity than C2 (IN+HI; see Figure 1 and Smith et al., in press). Conversely, for standard deviation of go reaction time (reaction time variability), IN+modHI+ODD was slightly more represented than IN+HI at worse scores, highlighting some contribution of HI and the exacerbating role of ODD. On most EF domains, the IN class was very slightly more represented at average scores, with a negligible slope from impaired to unimpaired ends. The only exceptions are trail making condition 4 (set shifting) and color naming (processing speed), for which inattention appeared to contribute to impairment (as the IN class had a clear slope for set shifting and a slight slope for processing speed) in addition to the contribution of hyperactivity/impulsivity (as IN+HI and IN+modHI+ODD youth showed similar slope to IN youth on set shifting, and a steeper slope than IN youth on processing speed). Lastly, C5 (modHI+lowODD) appeared generally unassociated with EF; the only exception was trail making condition 4 (set shifting), where this class demonstrated more representation at better scores. The only significant difference for EF between clinical groups at age 15 was in comparison to this class; namely, IN+modHI+ODD youth were more represented than modHI+lowODD at higher standard deviation of go reaction time scores (reaction time variability)⁶. However, it should be noted that C5 at age 15 had a particularly small sample size.

Differences across latent class analysis (LCA) findings may be due to the use of the same tasks that become easier with development, and/or the fact that symptomatically similar classes over time do not necessarily contain the same participants. However, LTA paths allow us to infer the role of the entire past symptomatic trajectory on age 15 EF impairment (see Figure 4). However, as noted above for age 15 LCA findings, sample sizes for LTA paths are relatively small and thus should be interpreted with caution (see Supplemental Table S19). Relatedly, the MNLpred R package (Neumann, 2021) could not estimate confidence intervals where there was limited data, and as such, the “inference by eye” approach was not utilized to estimate likely statistical significance of these multinomial logistic regressions for task performance predicting LTA paths (see OSF for more information). Nevertheless, figures were interpreted in the same manner as other results and as described in A Note on Interpretation. For visuospatial working memory, the worsening (442), persistently oppositional (333), and typical ADHD (224) paths all demonstrated similar slopes such that they were most represented at impaired values. In contrast to less impaired paths, these groups share hyperactivity/impulsivity during at least one point in development, and also had significant levels of symptomatology (of any kind) at age 9; the only exception are desisters (324). For color naming (processing speed), worsening (442) and persistently oppositional (333) paths had the steepest slopes, followed by less notable slopes in persistently IN (444), adolescent IN (114), and mixed trajectory (524) paths, with greater representation at impaired scores. This emphasizes the role of concurrent inattention (i.e., all groups were IN at the time of task performance), as well as the exacerbating role of concurrent hyperactivity/impulsivity (which 442 and 333 share), which is generally consistent with conclusions from LCA findings. Again, an exception was desisters (324), who had a notable slope in the opposite direction, as well as typical ADHD (224), with a slight slope in the opposite direction (i.e., more representation at better performance). For trail making condition 4 (set shifting), worsening (442) youth demonstrated the most prominent slope, followed closely by persistently oppositional (333) and persistently IN (444) youth, followed by a slight slope in typical ADHD (224) youth. While set shifting impairment appeared attributable to hyperactivity/impulsivity at age 9 and to both inattention and hyperactivity/impulsivity at age 15, LTA findings emphasize the role of inattention in all impaired paths for these tasks, especially given the clear association with task performance of persistently IN (444) youth. Stop signal reaction time (response inhibition) had relatively flat slopes overall; however, persistently oppositional (333) and worsening (442) paths have slightly steeper slopes, perhaps related to hyperactivity/impulsivity and oppositionality (generally consistent with LCA findings). For d’1 catch trials (vigilance), persistently oppositional (333) youth had the most marked slope, highlighting the role of oppositionality, followed by worsening (442), mixed trajectory (524), and persistently IN (444) paths, which share inattention. This contrasts with LCA findings, which highlighted the role of hyperactivity/impulsivity in vigilance. This may be due to the fact that persistently oppositional youth are also more persistently hyperactive/impulsive over time (see Smith et al., in press). Lastly, for standard deviation of go reaction time (reaction time variability), LTA findings strongly emphasize the role of oppositional defiant disorder symptoms, with only persistently oppositional (333) youth having the steepest slope. While LCA findings highlighted the role of both hyperactivity/impulsivity and oppositionality, this again may be due to the fact that classes with oppositionality also had greater hyperactivity/impulsivity over time (see Smith et al., in press) and/or hyperactivity/impulsivity and oppositionality’s slight divergence in adolescence and thus, the increasingly unique contributions of oppositionality to task performance at 15.

Discussion

Extant research investigating mechanisms of ADHD and DBP comorbidity largely uses analytic methods that neglect to consider the role of heterogeneity in comorbidity and vice-versa, and/or largely relies on cross-sectional studies. A recent paper (Smith et al., in press) utilized LTA to explore how a person-centered, longitudinal approach can help advance our understanding of the developmental course of externalizing psychopathology. The present study expands upon these foundational findings to consider how this approach may aid in exploring mechanisms.

One of the primary takeaways of the present study that contrasts with what has been found previously is that hyperactivity/impulsivity, and not inattention, is responsible for many domains of EF impairment (Pievsky & McGrath, 2018). It is surprising that inattention was unassociated with reaction time variability, as it is an EF domain with close theoretical ties to inattention in the literature (e.g., Leth-Steensen et al., 2000; Tamm et al., 2012), though consistent with some empirical evidence that RTV is especially associated with HI symptoms (Kofler et al., 2013). Notably, even with slight decreases in hyperactivity/impulsivity for C2 over time (i.e., classes IN+HI at 9 and 15 and IN+modHI at 12) and moderate decreases in hyperactivity/impulsivity for C3 over time (i.e., classes IN+HI at 9 and IN+modHI+ODD at 12 and 15; see Figure 1, Figure 2, and Smith et al., in press), hyperactivity/impulsivity was still most contributory to many domains of EF impairment in adolescence. This is especially interesting as all classes maintain high levels of inattention, and the IN class becomes more prominent in adolescence (see Figure 2 and Smith et al., in press). This adds further evidence to the idea that fewer hyperactivity/impulsivity symptoms in adolescence may be necessary as even moderate levels in adolescence was clearly contributory to EF impairment. One potential reason for these findings is our analytic approach. That is, most analytic approaches rely on comparing means or effect sizes and find a severity heuristic (i.e., increasingly more symptoms are associated with increasingly more impairment). Indeed, the descriptives for task performance in our data are consistent with this and suggest that TD youth are least impaired, followed by classes with inattention only, followed by classes with inattention in addition to hyperactivity/impulsivity and oppositionality (see Supplemental Table S1-S18); however, the predicted probability figures demonstrate that inattentive youth are nearly equally or normally distributed across scores on many dependent variables.

Further, primarily inattentive classes were surprisingly aligned in many ways with subthreshold classes (i.e., C5 or the ST class at age 9, and C5 or the modHI+lowODD class at age 15). On many EF tasks at age 9 and 15, IN and ST had a similar slope, and at age 12, IN and ST youth were symptomatically combined in a latent class (i.e., C4 at age 12 was ST/IN). The one distinction between these classes is that both ST youth at age 9 and modHI+lowODD youth at age 15 lacked high levels of inattention symptoms. These youth also lacked EF impairment in the two domains for which inattention was often contributory – processing speed and set shifting – and some of the functional impairment observed in IN youth (i.e., learning and family in C4 at age 15; see Smith et al., in press). These findings suggest that the absence of inattention specifically may be a sign of non-clinical externalizing behavior. Instead, ST (C5) youth may instead represent normative hyperactivity at age 9, and modHI+lowODD (C5) youth may represent normative impulsivity/oppositionality at age 15.

Our results suggested that HI and ODD were generally more aligned with each other than HI and IN were, contrary to what the DSM nosology suggests. HI and ODD were generally aligned in EFs in childhood (i.e., age 9), except that ODD conferred additional impairment for reaction time variability and response inhibition. This is in contrast to a great deal of literature that attributes these impairments to ADHD (Barkley, 1997; Pievsky & McGrath, 2018). HI and ODD diverged moreso in adolescence, such that the IN+modHI+ODD class has fewer HI symptoms than the IN+HI class, though still more than other classes (see Figure 1 and Smith et al., in press). This also means that for tasks on which the IN+HI class exceeded the IN+modHI+ODD class in representation at impaired scores, this representation is especially attributable to the higher levels of hyperactivity/impulsivity (e.g., visuospatial working memory, vigilance). Interestingly, the tasks for which oppositionality conferred additional impairment beyond hyperactivity/impulsivity may have elicited an emotional response such as frustration (Smith et al., 2023). Specifically, reaction time variability and stop signal reaction time are both from the Stop Task, during which a beep sounds to instruct the participant to stop a prepotent response; the task is designed to cause a certain fail rate for each participant, inherently blocking goal attainment (Leibenluft, 2017). Relatedly, while the present study did not evaluate the role of emotion, irritability, and mood symptomatology, it is possible that a part of the divergence between HI and ODD is due to an unmeasured pathway from ODD to mood disorders (i.e., an affective dimension of ODD), a distinct branch from the pathway from ODD to continued externalizing behaviors (i.e., a behavioral dimension; Burke et al., 2021; Burke & Loeber, 2010). Additionally, findings are consistent with the idea that it is perhaps subthreshold or unaccounted for ODD in ADHD youth that may in part drive heterogeneity in findings focused on ADHD youth (Dolan & Lennox, 2013; Noordermeer et al., 2020).

Collectively, there are several insights that can be made regarding theories of developmental psychopathology. A great deal of work has focused on the pathway from childhood ADHD (particularly hyperactivity/impulsivity) to ODD and later CD as distinct but related phenomena which confer risk for one another in sequence (e.g., Ghosh et al., 2017; Rowe et al., 2010); while our study did not have sufficient variability in CD, our findings suggest that HI and ODD co-occur from childhood and that the prevalence of ODD decreases with time (see Figure 2 and Smith et al., in press). With the added consideration of mechanisms, we can also consider that HI and ODD are largely aligned in EF impairment from a young age, diverging slightly in adolescence as the role of EF in HI becomes more pronounced as levels of HI decreases in oppositional youth. While HI and ODD have a high rate of apparent co-occurrence due to numerous shared risk factors under our current nosology, our findings do not suggest that they are separate and relate to one another in sequence, or even that they represent heterotypic continuity (i.e., that one diagnostic entity shifts in presentation over time; Beauchaine & McNulty, 2013). Instead, our findings suggest that, with respect to the literature focused on the pathway of externalizing psychopathology specifically, HI and ODD may be different behavioral manifestations of the same underlying mechanisms, perhaps representing one etiological entity, at least in childhood and with respect to EF. Looking forward in age, our findings are roughly aligned with multifinality, such that HI and ODD share mechanisms and symptomatic presentations in childhood, but may become more distinct or follow different trajectories over time (i.e., oppositionality becomes less associated with HI and thus less associated with the EFs that are associated with HI). Certainly, however, this is only one exploratory study which should be followed up with further research to specifically test the hypotheses posed here. Additionally, as briefly mentioned above, the present study did not evaluate the role of putative emotional mechanisms or mood disorders; the heterogeneity and heterotypic continuity of ODD to mood concerns and the comorbidity across externalizing and internalizing psychopathology complicates these pathways further, and more transdiagnostic work is needed to explore these issues.

There are many insights that were offered by this analytic approach. Most importantly, a problem with cross-sectional studies is that it cannot be determined whether differences (in symptoms or mechanisms) reflect meaningful distinctions or merely groups of children at different places in the course of one symptomatic pathway (Beauchaine & McNulty, 2013). Our analytic approach allowed us to consider this. For example, the typical ADHD path (224) was distinct from the persistently IN path (444). While both paths presented with inattention only in adolescence, the previous hyperactivity/impulsivity in the typical ADHD path appeared to contribute to some EF impairment at age 15. This underscores the potential etiological distinctiveness of a persistently IN class compared to ADHD-combined youth who normatively decrease in HI with age. Conversely, if 15-year-olds were considered in a cross-sectional study using DSM nosology, it is likely that some EF impairment would have been detected in inattentive adolescents and attributed to inattention, when perhaps EF deficits were actually attributable to prior hyperactivity/impulsivity in childhood. Further, our exploratory approach allowed us to attend to patterns in the data with a greater amount of nuance; while this was at the expense of concision, the aggregate results offered by many analytic studies may be overly simplistic. For example, our study was able to detect the lack of clear association between IN symptoms and many EFs (i.e., the flat slopes for IN classes in many predicted probability figures). As another example, our approach allowed us to consider the extent to which certain EF deficits (e.g., reaction time variability) were primarily driven by HI and exacerbated by the presence of ODD. Studies of ADHD only likely would have attributed these findings to HI only, and studies of DBPs only likely would have attributed the contributions of HI to ODD.

Nevertheless, this study has several limitations that should be outlined. First, our study was limited in part by sample size. A larger sample size would have allowed for classes and LTA paths to be larger and more robust; sample sizes were especially small at age 15. Larger sample sizes and broader recruitment of symptoms would have also likely allowed for more investigation into ODD and CD heterogeneity. Further, there were even smaller sample sizes on specific tasks that were missing for some participants (see Supplemental Table S1-S18). Future studies should utilize much larger samples, as well as samples exploring other commonly comorbid symptoms (e.g., anxiety, depression; disruptive mood dysregulation and bipolar symptomatology). Additionally, our study investigated youth from 9 to 15. While this is a wider range than many studies that investigate only young childhood or only adolescence, there are several important insights that are missed. For one, it is possible that a transition from HI to ODD occurs in younger children that our study did not include (e.g., Burns & Walsh, 2002; Harvey et al., 2016). It is also possible that interesting patterns may unfold in late adolescence and early adulthood, particularly if HI and ODD symptoms continue to diverge, which future studies should investigate. Moreover, the present study did not investigate the role of emotion regulation, which may be especially associated with oppositional symptoms (Cavanagh et al., 2017; Leibenluft, 2017). Additionally, the present study did not investigate any extra-individual factors (e.g., socioeconomic status; family, neighborhood, or community factors) that are associated with ADHD and ODD symptoms and their course (Noordermeer et al., 2017; Rowland et al., 2018; Russell et al., 2015; Sharp et al., 2021). Relatedly, the present study utilized a predominantly white, non-Hispanic, middle-class sample, which limits generalizability. Further, the parent education level in the present sample is higher than typical; this relatively advantaged sample may limit the observed rates of ODD (Noordermeer et al., 2017; Rydell, 2010). Lastly, the role of treatment participation on latent classes and longitudinal latent paths was not considered.

This study not only provides unique insights that can be tested with future confirmatory data analysis, but also serves as a proof-of-concept for the importance of both EDA and person-centered, longitudinal methods for understanding mechanisms of comorbidity. Most of the literature is focused on identifying what is different between DSM groups rather than considering what is shared, which is surprising given the extraordinarily high rate of comorbidity. While analyses that parse or adopt a “splitting” approach may produce more homogenous groupings that are meaningful in one way or another (e.g., in treatment response, in mechanisms), these approaches will also create more comorbidity between these groupings, which will ultimately further obscure the relationships between symptoms and mechanisms if future studies utilize these groupings and control for the shared variance between them. While we recognize that “lumping” (i.e., Beauchaine et al 2010; Meehl, 1987) may conversely generate fewer but more heterogeneous groupings, we discourage unquestioning reliance on DSM groupings that hinder the advancement of basic science seeking to understand the etiology of psychopathology, and instead encourage unfettered exploration of the relationships between symptoms and putative etiological mechanisms.

Supplementary Material

Please see the end of this document, or the journal’s website, for supplemental materials.