Using Hot and Cool Measures to Phenotype and Predict Functional Outcomes Across Dimensions of ADHD and Typical Development in Adolescents

The present study aims to (a) use a person-centered approach to identify differential profiles of hot and cool processes and (b) characterize differences in sex, diagnosis, and performance among these profiles and relate them to areas of critical everyday functioning. We hypothesized that distinct profiles would emerge within the sample, beyond diagnostic categories—and that not all clinical participants would be rated as high on emotional variables. We then considered the relationship between profile membership and everyday functioning. This study focused on a sparsely studied group, adolescents with ADHD, Combined presentation symptoms, and TD adolescents, to enable the exploration of hot and cool processes at a time in development associated with rapid changes between these processes and in youth with heightened risk for impulsiveness and risk-taking behavior. We used latent profile analysis (LPA) to parse heterogeneity in ADHD and identify clinically meaningful subgroups. Furthermore, we applied the Research Domain Criteria framework to examine how phenotypic features associated with hot (e.g., emotional lability, irritability, risk-taking) and cool (e.g., PS, WM) functions presented beyond diagnostic categories in ADHD and TD youth, using a dimensional approach across a range of the hot versus cool measures.

This study draws from the Mapping Impulsivity's Neurodevelopmental Trajectory (MINT) longitudinal investigation of the neurodevelopmental trajectory of impulsivity (PI: Schweitzer; Elliott et al., 2022; Kahle et al., 2021; Mukherjee et al., 2021, 2022) in adolescents and young adults. MINT collected imaging, behavioral, clinical, and academic measures for ADHD, Combined Presentation, and TD adolescents. Toward the end of the study, MINT also recruited participants with a range of ADHD symptoms between TD and the criterion for clinical diagnosis of ADHD to study ADHD symptoms on a continuum. The University of California, Davis Institutional Review Board approved this study, with written informed consent obtained from parents and written assent from participants.

Participants were recruited from the MIND Institute participant recruitment pool, MIND Institute and Departments of Psychiatry and Pediatric specialty clinics, local middle and high schools and universities, community flyers, and social media sites. Two licensed psychologists (JBS and JFD) evaluated phone screen data regarding symptoms and functioning to determine study eligibility. Eligible participants and one of their parents/caregivers completed a full psychiatric interview (Diagnostic Interview Schedule for Child and Adolescents or the M.I.N.I. International Neuropsychiatric Interview—Kid for DSM-5); parents and teachers completed rating scales (i.e., Parent [Conner-3 Parent Rating Scale—CPRS-3]) and [Teacher (Conners-3 Teacher Rating Scale—CTRS-3]). A licensed psychologist (JBS or JFD) determined whether participants met DSM criteria for ADHD, the presence of any other major psychiatric disorder, or would be best classified as “subthreshold” or TD. Subthreshold ADHD refers to a classification in which individuals exhibit the presence of symptoms characteristic of ADHD but do not meet the full diagnostic criteria as outlined in the DSM-5. Specifically, individuals categorized as subthreshold ADHD in our study present with 3–5 symptoms related to either inattentive or hyperactive/impulsive behaviors, which are representative of ADHD symptomatology. Moreover, these symptoms are accompanied by substantial functional impairment across multiple settings, indicating their significant impact on daily activities, responsibilities, and interactions.

Initially, DSM-IV-TR was used, and then DSM-5 upon its publication. However, evaluation criteria for ADHD for participants entering the study under DSM-IV-TR were re-reviewed for DSM-5 criteria for ADHD for all included participants. Inclusion criteria included an IQ ≥ 80 with additional inclusion criteria for the ADHD group meeting DSM-5 criteria for ADHD Combined presentation. Exclusion criteria included the presence of a math or reading learning disability; a history of head trauma, neurological disorder, or major medical problem as reported by the participant or their parents; taking psychoactive medication other than stimulants or atomoxetine; and meeting criteria for Axis I diagnosis except for ADHD, oppositional defiant disorder, or conduct disorder. Participants prescribed stimulant medication or atomoxetine for ADHD abstained from taking the medication for five half-lives before being tested on the Balloon Analogue Risk Task and Picture Order Memory Paradigm. The current study focuses on the adolescents in our study and includes 152 youth between 12 and 17.9 years old (99 boys, 53 girls), utilizing the data from their baseline visit.

We selected a range of measures that tap into cognitive control and reward/emotion processes and functions that integrate the two. We used multi-informant reporters, including parent and self-report scales and objective measures, representing trait and state functioning.

We first used LPA to identify distinct patterns related to emotional and reward functioning and cognitive control based on CPRS-3, BIS, ZTPI, and BART scores. The maximal reliability (H) for the eight variables included in the LPA was 0.92 (95% confidence interval 0.88 to 0.95). Models were estimated using full-information maximum likelihood, allowing us to include the participants with missing data under the missing-at-random assumption. We fitted and compared models with increasing numbers of classes, starting with one and determining the optimal number of classes based on statistical goodness-of-fit criteria, considering whether the classes capture clinically meaningful features and the proportion of participants represented in the classes (Masyn, 2013; Nylund et al., 2007). Goodness-of-fit criteria included Bayesian information criterion (BIC) and sample-size adjusted BIC, Akaike Information Criterion (AIC), consistent AIC and corrected AIC, approximate weight of evidence criterion (AWE, Masyn, 2013), correct model probability (cmP), Vu-Lo-Mendell-Rubin, Lo-Mendell-Rubin adjusted (LMR), and Parametric Bootstrapped likelihood ratio tests (BLRT, Lo et al., 2001; Nylund et al., 2007). Smaller AIC, BIC, and AWE values indicate better fit. cmP allows a researcher to compare a set of more than two latent class models, and higher values indicate more robust evidence for the candidate model compared to other models (Masyn, 2013). The likelihood ratio tests compare the fit of the specified class solution to models with one fewer class, and a significant p-value indicates that the specified model is preferred. The local maximum problem was addressed using up to 2,000 starting points to replicate each model.

Each LPA model identifies the number of latent classes (subgroups) within the sample and estimates posterior probabilities for each participant’s assignment to each latent class. For descriptive analyses, the highest posterior probability from the best-fitting model was used to assign each participant to the most likely subgroup. For subsequent analyses using latent subgroup membership (i.e., examination of differences in diagnosis, sex, impulsivity, and achievement), multiple pseudo-class draws (Bandeen-Roche et al., 1997) were used to reduce bias by accounting for the uncertainty in class assignments. Differences in categorical characteristics (e.g., sex, ADHD diagnosis) across latent classes were assessed using χ² tests. Differences in dimensional measures of functioning were assessed using general linear models, accounting for age and sex, as appropriate. Age and sex were not retained in the reported models if they did not contribute significantly to the model. Transformations were employed for variables that violated the normality assumption. We used univariate and bivariate residual plots and univariate summaries of residuals to check the normality and homoscedasticity of residuals. Participants were randomly classified into latent classes 100 times based on their distribution of posterior probabilities from the best-fitting LPA model. The subsequent analyses were performed 100 times (i.e., for each draw), and results were combined across draws using standard methods for multiple imputations for missing data (Rubin, 1987). After combining results across draws, two approaches were used to control for multiple comparisons when assessing significant differences between groups. First, overall, three-degree-of-freedom F-tests for the group were adjusted for multiple comparisons using the Benjamini–Hochberg method. Then, for all variables with significant overall F-tests for group after adjustment, a second Benjamini–Hochberg procedure was used to identify all pairs of latent groups that were significantly different out of all possible pairs. LPA was performed in Mplus version 8.0 (Muthen & Muthén, 1998-2017). All other analyses were implemented using SAS Version 9.4 (SAS Institute Inc., Cary, NC). All tests were two-sided, and p-values < 0.05 were considered statistically significant.

Fit indices for one-class to six-class solutions are summarized in Table 1. They provided support for the four-class as the optimal solution. Except for AIC and sample-size adjusted BIC indices, which never increased with added classes (though substantially smaller improvements were garnered after the four-class solution), the AWE, which supported a two-class solution, and the BLRT, which continued to support the larger model up to six classes, all the other criteria suggested that a four-class solution was optimal (four-class was better than three-class, and five-class was not better than four-class). In latent profile analyses, AIC and BIC may not increase with additional parameters, but the resulting models may have additional classes that are not meaningful. For example, in the five-class model, one class with impairments across modalities was differentiated into two classes that were not meaningfully different. Moreover, the five- and six-class models identified a class that included 5% or less of the sample. Thus, the four-class solution was selected as the most parsimonious model that still provided adequate fit and the most clinically meaningful distribution of classes. Based on the pattern of profiles, the four classes were named High-Complex Challenges (17.0%), Moderate-Mixed Challenges (19.9%), Non-Emotive Impulsivity (30.1%), and High Regulation and Control (32.9%). The four-class model provided good classification quality, with an entropy value of 0.87 and high average posterior probabilities of class membership: 0.98, 0.93, 0.89, and 0.91, respectively. Figure 2 illustrates the profiles for the four-class solution. To aid interpretation, before graphing the profiles, we standardized the scores. For all eight variables included in the LPA, there were significant differences across classes (Kruskal–Wallis ps < 0.01). The High-Complex Challenges showed impairment on almost all measures, but particularly on emotional lability and irritability. Challenges are also present in PS, high risk-taking, poor self-control, and low future orientation. The one area in which they displayed less impairment was on a self-rating for oversensitivity to immediate rewards (i.e., present hedonism). The Moderate-Mixed Challenges class is characterized by strong negative emotionality but to a less intense degree than the High-Complex class. This class also still displays elevated issues with self-control, including being drawn to more immediate rewards, yet somewhat higher future orientation and less risk-taking. This Moderate-Mixed class may represent a combination of moderate challenges in cognitive control and reward-related processes. The Non-Emotive Impulsivity class is consistent with immature development of cognitive control, appearing highly drawn toward immediate rewards, high risk-taking, low future orientation, and yet, strong emotional control. The High Regulation and Control class excelled in cognitive control, reward, and emotional control measures.

Demographic and clinical characteristics for all classes are presented in Table 2. The overwhelming majority of the High-Complex Challenges class comprises ADHD-diagnosed adolescents. The Moderate-Mixed class predominantly includes adolescents diagnosed with ADHD, with somewhat lower scores on Conners’ scales than the High-Complex Challenges class. The Non-Emotive Impulsivity class comprises 59% ADHD-diagnosed adolescents, 2% subthreshold ADHD, and, interestingly, 39% TD. Finally, the High Regulation and Control class is primarily, but not exclusively, TD (82%).

Participant sex significantly differed across class membership (p = 0.03). A higher proportion of girls were classified in High Regulation and Control classes. However, the four classes were similar in age, race and ethnical composition, and parental education (Table 2).

Next, we examined whether the four LPA classes differed in functional outcomes. Table 3 shows the CPRS-3, POMP, and WIAT scores for the four classes, and Fig. 3 summarizes the standardized average scores for the four groups across these variables. Table 4 shows the estimated class differences after adjusting for multiple comparisons.

The classes differed significantly on all CPRS scales examined (all p< 0.001). After adjusting for multiple comparisons, adolescents in the High-Complex class displayed significantly higher DSM ADHD Hyperactive/Impulsive scores compared to two classes, Non-Emotive Impulsivity and High Regulation and Control (p< 0.001; Table 4). The Moderate-Mixed class had significantly higher scores than the Non-Emotive Impulsivity and High Regulation and Control classes (p = 0.03 and < 0.001, respectively). After adjusting for multiple comparisons, the High-Complex class did not significantly differ from the Moderate-Mixed class (p = 0.08). Finally, the Non-Emotive Impulsivity class had significantly higher scores than the High Regulation and Control class (p< 0.001). A similar pattern was observed for Learning Problems, with the High-Complex class scoring higher than the Non-Emotive Impulsivity and High Regulation and Control (both p < 0.001). The Moderate-Mixed class showed more Learning Problems than the Non-Emotive Impulsivity and the High Regulation and Control classes (p = 0.005 and < 0.001, respectively). Finally, the Non-Emotive Impulsivity class had higher scores on Learning Problems than the High Regulation and Control class (p = 0.002). Peer Relations from the CPRS-3 revealed significant differences with the Moderate-Mixed class, indicating higher Peer Relations scores than the High Regulation and Control class (p < 0.001). Additionally, the Non-Emotive Impulsivity class had higher Peer Relations scores than the High Regulation and Control class, but this difference was not statistically significant (p = 0.07). Finally, the High-Complex class had higher scores on the Peer Relations subscale compared to the Non-Emotive Impulsivity and High Regulation and Control classes (both p< 0.001). The pattern was similar for EF, with the High-Complex class scoring higher than the Non-Emotive Impulsivity and High Regulation and Control classes (p< 0.001). Additionally, the Moderate-Mixed class scored higher on EF than the Non-Emotive Impulsivity and High Regulation and Control classes (p = 0.007 and < 0.001, respectively). The Non-Emotive Impulsivity class scored higher on EF than the High Regulation and Control class (p < 0.001).

The groups differed significantly on the squared transformed WM task scores (p = 0.004). After further adjusting for multiple comparisons, the results showed that adolescents in the High-Complex and Moderate-Mixed classes had lower scores compared to the High Regulation and Control class (p = 0.01 and 0.004, respectively). Finally, we examined whether the LPA classes differed in academic achievement based on WIAT-III standardized Reading Comprehension and Math Problem Solving scores. As expected, the classes differed on the log-transformed WIAT-III standardized subscale scores (both p = 0.04 after adjusting for multiple comparisons). After further adjusting for multiple comparisons, the results indicated that adolescents in the High-Complex class performed significantly worse in Reading Comprehension than those in the High Regulation and Control class (p = 0.03, Table 4). For the Math Problem Solving scores, the results suggested that adolescents in the High-Complex class demonstrated worse performance than those in the High-Regulation and Control class in Math Problem-solving. Still, these differences did not reach statistical significance after adjusting for multiple comparisons (both p = 0.055, Table 4).