Computational Modeling of Attentional Impairments in Disruptive Mood Dysregulation and Attention Deficit/Hyperactivity Disorder

Abstract

Objective:

Computational models provide information about cognitive components underlying behavior. When applied to psychopathology-relevant processes, they offer additional insight to observed differences in behavioral performance. Drift diffusion models have been successfully applied to investigate processing efficiency during binary choice tasks. Using these models, we examine the association between psychopathology (irritability and inattention/hyperactivity) and processing efficiency under different attentional demands.

Method:

N=187 youth with ADHD, DMDD, both disorders, or no major psychopathology (age M=13.09 years, SD=2.55; 34% female) completed an Eriksen Flanker task. Of these, 87 youth provided complete data on dimensional measures of the core symptom of DMDD (irritability) and those of ADHD, inattention and hyperactivity.

Results:

In a categorical diagnosis-based analysis (N=187), we found significant interactive effects between ADHD, DMDD, and task condition on processing efficiency, whereby changes in processing efficiency between conflict and non-conflict conditions were larger in youth without psychopathology relative to patients. Analysis of symptom severity (n=87) across diagnosis similarly revealed an interaction between symptom dimensions and task condition on processing efficiency. Irritability moderated the magnitude of association between inattention symptoms and difference in processing efficiency between conflict and non-conflict conditions.

Conclusion:

Adapting processing efficiency to cognitive demands may represent a shared cognitive endophenotype for both ADHD and DMDD. Highly irritable and/or inattentive youth may have difficulty adjusting processing efficiency to changing task demands possibly reflecting impairments in cognitive flexibility.

Introduction

Increased reaction time variability on cognitive tasks is among the most replicated behavioral alterations in ADHD ^1,2. Recent psychophysical theories suggest that such variability can be explained by processing inefficiency, which explains behavior that is not just more variable in speed but also slower and likely to result in errors ³. However, reaction time variability may not be specific to ADHD as it is also observed in other psychiatric phenotypes ⁴. Particularly, very little work to date has examined reaction time variability in youth with severe, chronic irritability. Here, we provide evidence from both a categorical and continuous, transdiagnostic approach to demonstrate associations between irritability and attention deficits, reaction time variability, and processing efficiency.

In studying affective psychopathology, examining response time variability may yield important insights into relevant cognitive processing mechanisms in affective ⁵ and nonaffective contexts ⁶. The high rate of comorbidity between ADHD and DMDD, as well as co-occurrence of their core symptoms, irritability and inattention/hyperactivity, in clinical populations ^7-10, has impeded our examination of the relative contributions of irritability and attentional symptoms to reaction time variability. As noted above, response time variability is a stable feature of ADHD, present across diverse tasks (e.g., reward, cognitive control), with some evidence for moderation of group differences by task difficulty ¹¹. However, advances in computational modeling and clinical phenotyping allow us to examine the associations between irritability, ADHD symptoms, and reaction time variability.

Computational models measure latent cognitive constructs and can reveal mechanismbased psychopathology-related differences in behavioral performance. With regards to response time variability in ADHD, the drift diffusion model (DDM) has been particularly impactful ^12,13. The DDM accounts for reaction time, reaction time variability, and accuracy in speeded, binary choice tasks ¹⁴. It is well-defined with strong theoretical underpinnings and excellent explanatory power for human behavior ¹⁵. In the DDM, latent constructs explaining speed-accuracy trade-off effects are represented by parameters coding the strength of evidence entering the decision process, the amount of accumulated evidence required to make a decision to explain speed-accuracy trade-off effects, as well as time for motor preparation and output ¹⁶. Because it represents reaction time variability while accounting for all behavior (accuracy and reaction time), the primary parameter of interest for the present study is the drift rate, v, generally interpreted as processing efficiency (i.e., a large value represents more rapid, less variable, and error-resistant responses).

In the current study, we compared youth diagnosed with ADHD, DMDD, both, or no major psychopathology on their performance on an Erikson Flanker task. A subsample of participants also provided dimensional measures of the core symptom of DMDD, irritability, and those of ADHD, inattention and hyperactivity, for a complementary transdiagnostic analysis of symptoms. Based on previous findings ^12,17, we expect that youth with ADHD will have lower drift rates relative to healthy volunteers. The Flanker task includes attentional conditions which may facilitate processing or introduce interference; we expect interference to slow processing by reducing drift rates ¹⁸. Thus, the attentional demands of the Flanker task not only allow an examination of associations between irritability and ADHD symptoms and drift rate but also the change in these associations with changing attentional demands. This is the first study to systematically investigate response time variability and processing efficiency in youth with DMDD; thus, we are agnostic in our predictions about cognitive control impairments in youth with DMDD only, as the contribution of comorbid ADHD to attention processing in this population remains unclear.

Method

Participants

Two-hundred twenty-one youth with and without psychopathology (healthy volunteers; HVs) participated in this study. Youth were recruited both locally and nationally through practitioner referrals and IRB approved advertisements. Written informed consent was obtained from parents and assent was obtained from children. The study was approved by the NIMH Institutional Review Board.

Diagnoses were made by master’s or doctoral level clinicians using a the Schedule for Affective Disorders and Schizophrenia for School-Age Children—Present and Lifetime Version K-SADS-PL ¹⁹; with an additional module to assess DMDD ²⁰. For all participants, IQ<70 (determined by the Wechsler Abbreviated Scale of Intelligence, WASI ²¹), history of head trauma, neurological disorder, pervasive developmental disorder, medical illness preventing study participation, cardinal bipolar symptoms, post-traumatic stress disorder, schizophrenia spectrum disorders, current major depressive disorder or substance abuse within three months were exclusionary for study participation.

Of 221 participants, 34 youth were excluded from analyses: 32 for task performance (see Methods for criteria), one child was excluded as the DDM failed to produce valid parameter estimates and one child was identified as a high leverage value in the multilevel model (Cook’s distance >2.5). Characteristics of the remaining 187 participants are presented in Table 1. Age and sex did not significantly differ between diagnostic groups. Significant differences in IQ emerged across groups.

Table 1.

Sample Characteristics by Diagnostic Group

	HV (N=47)	ADHD (N=44)	DMDD- ADHD (N=19)	DMDD+AD HD (N=77)	Stat	p	es	Post hoc contrasts a
Age (years)	13.61 (2.55)	12.95 (2.59)	12.95 (2.91)	12.88 (2.43)	F=0.90	.44	η2=0.01
Sex (female)	20 (42.55%)	12 (27.27%)	10 (52.63%)	21 (27.27%)	χ2=6.93	.07	V=0.11
IQ *	113.09 (14.74)	114.59 (11.73)	113.74 (9.77)	108.01 (13.22)	F=3.09	.03*	η2=0.05	ADHD> DMDD+ADHD
Accuracy	0.97 (0.03)	0.96 (0.04)	0.97 (0.04)	0.96 (0.04)	F=1.06	.37	η2=0.02
MeanRT	511.61 (77.33)	521.15 (76.01)	530.38 (72.56)	528.02 (72.79)	F=0.55	.65	η2=0.01

Note. ADHD = attention dysregulation/hyperactivity disorder; DMDD = disruptive mood dysregulation disorder; HV = healthy volunteers.

^aPost hoc contrasts are pairwise comparisons using t-test with pooled SD, Holm’s method for p value adjustment

^*p<.05.

For a subsample of n=87 youth (30 youth with ADHD only, 7 youth with DMDD only, 38 youth with DMDD+ADHD and 12 youth without psychopathology), a parent completed the Affective Reactivity Index (ARI ²²) and the Conners’ Parent Rating Scale ²³; (CPRS – hyperactive and inattentive subscale, raw scores), assessing irritability and ADHD severity, respectively.

Task

Youth were presented with a modified version of the Eriksen Flanker task that used arrows rather than letter stimuli ^24,25. Five stimuli were arrayed horizontally with a central target (a left or right pointing arrow) and two distractors on each side. Conditions were: 1) congruent: distractors were identical to the target; 2) incongruent: distractor arrows were pointed opposite to the target; and 3) neutral: distractors were squares. Each trial started with a fixation cross (500 ms) followed by stimuli and response window (1000 ms) and then a blank screen (1500 ms). Figure S1, available online, depicts the stimuli and trial sequence. The task was presented as one continuous block of 130 trials, taking 6.5 minutes to complete. Condition order was pseudo-randomly determined to maintain approximately the same frequency for each condition. Each participant experienced the same order of trial conditions.

The task was presented via E-Prime 1.1 Build 1.1.4.1 or E-Prime 1.2 Build 1.2.1.847 on a laptop computer running Windows XP version 2002 or Windows 7 Professional. Stimulus arrays were 5.93x1.22 cm presented on a laptop placed at a viewing distance of approximately 60 cm. Participants responded to the target’s direction with their corresponding index finger, placed on the laptop’s “a” (left) and “l” (right) keys. Children completed a practice task of 30 trials and were trained to 83% (25/30 correct responses) accuracy before proceeding to the task.

Statistical analyses

Data preparation

Task data were prepared for analyses by removing non-physiologic anticipation responses (RT<150ms; 0.2% of all trials) and trials without responses. Participants who performed at or below chance (≤50% accuracy in any condition) or were too poorly engaged in the task (an overall non-response rate ≥15%) were excluded from further analyses (n=32). All analysis was performed in R ²⁶ Version 3.5 using the “lme4” package for mixed effects model analyses ²⁷, “sjPlot” for creating tables ²⁸.

Drift Diffusion Model

As a speeded, binary choice reaction time paradigm generally completed with high accuracy and speed, the Flanker task is fit for the application of the Drift Diffusion Model. Previous work has successfully used drift diffusion modeling in the analysis of Flanker task data ¹⁸. Diffusion parameters were estimated for each subject from all non-missing trials, including both correct and error responses, using the full distribution of reaction times. Diffusion parameters were estimated from the trial-level data for each participant using the fast-dm modeling program Version 30.2 ^29,30 Drift rate (v) was allowed to vary across condition. Boundary separation (a) and non-decision time (T) were held constant across conditions. Bias (zr) was set to the neutral value of 0.5. The maximum likelihood approach was used to estimate parameters and assess fit.

To examine model performance, we compared simulated data to the empirical data. We used the construct-samples tool in fast-dm to simulate datasets from participants’ specific parameter sets. For each participant, one dataset was simulated with n=1000 trials for each condition and q-p plots for simulated and empirical data were compared. Simulated and empirical data were each pooled across subjects by averaging the quantiles of individual reaction time distributions. Quantile probability plots ¹⁶ are a standard method to represent the distribution of reaction time by accuracy for all conditions in a task. See Figure 1 for a quantile probability plot for five reaction time quantiles (0.1, 0.3, 0.5, 0.7 and 0.9) for the empirical and simulated data for all youth. Given the small number of errors, medians were plotted rather than quantiles ¹⁸. The comparison of simulated and empirical data indicates the model does a good job accounting for accuracy, mean RT, and RT variability across all three conditions.

Figure 1. Quantile Probability Plot for Pooled Data From all Youth.

Note: For each condition, the 0.1, 0.3, 0.5, 0.7, and 0.9 quantiles of reaction time are averaged across participants or simulations of their parameter set. Correct responses are on the right of each panel and represent almost all responses (average accuracy >92% by condition). Overall, the model reasonably represents behavior, especially in the neutral and congruent conditions. In the incongruent condition, the model underestimates the fastest 30% of reaction times, especially the fastest 10%. This is expected as attention dynamics that represent early interference effects on the drift rate are not modeled by a constant drift rate. ¹⁸

A linear mixed-effects model was used to examine the DDM drift rate (v) parameter of interest, testing the fully interactive effects of diagnosis (DMDD and ADHD) and task condition (three level repeated measure), with age and IQ as covariates. Participant was a random effect. In a subsample of youth (n=87), we used similar mixed-effects models to examine associations between continuous measures of irritability and ADHD (hyperactivity and inattentiveness tested separately) and task condition predicting drift rate, with age and IQ as covariates. Given a floor effect in ARI scores, the measure was converted to a binary factor representing low versus high irritability symptoms by median split. As including sex in the models did not affect results, the models are presented without this additional covariate. The fixed effects of differences across factor levels were tested by ANOVA, with marginal sum of squares and Satterthwaite correction, reported as F tests for the highest order term. For post hoc analysis, regression tables of parameter estimates, and factor contrasts are reported, estimates of effects are taken at the reference levels for all other factors.

Supplement 1, available online, includes analyses examining reaction time variability (intra-subject variability of reaction time measured as the coefficient of variation [CoV], the standard deviation/mean RT for correct responses by condition). We also include analyses exploring the relationship between CoV and drift rate for each condition using Pearson correlations (Supplement 2 and Table S1, available online).

Results

Diffusion Model Parameters

As expected, response time variability as measured by CoV was associated with drift rate for each task condition, all rs(185)≥-.50, ps<.001. See Supplement 1 and 2 and Table S1, available online, for an analysis of CoV and correlations between drift rate and CoV by condition.

Diagnostic Associations

For the categorical analysis of the DDM drift rate parameter (v), a significant interaction between ADHD, DMDD, and task condition emerged (F(2,366)=3.47, p=0.03; see Table 2 for full model, and Figure 2 for their effects). Differences resulting in the three-way interaction are indicated by the contrasts in Table 2. In the absence of ADHD, there were slower drift rates for those with DMDD relative to those without psychopathology, in the neutral and congruent relative to the incongruent condition (neutral: b=−.92, p=.008, 95%CI[−1.60, −.25]; congruent: b=−.85, p=.014, 95%CI[−1.53, −.17]; Table 2). In parallel, in the absence of DMDD, ADHD was associated with a decrease in drift rate in the neutral condition (relative to the difference in the incongruent condition, b=−.62, p=.021, 95%CI[−1.14, −.10], Table 2). Note, having both ADHD and DMDD did not result in a simple addition of the effects of each diagnosis on the drift rate difference between the neutral and incongruent conditions (b=1.07, p=.012, 95%CI[.24, 1.89], Figure 2). The sign is opposite and the magnitude similar to the effect of having each diagnosis alone in the same condition contrast, meaning that youth with both diagnoses are similar to those with one diagnosis or the other. A post hoc model (Supplement 3, available online) supports the interpretation that a change in drift rate between conflict to non-conflict conditions is greater for HVs than for youth with any diagnosis (Diagnosis x Conflict F(1,372)=9.01, p=.003, marginal delta v = 0.56, Figure S2, available online).

Table 2.

Mixed Effects Model for Drift Rate predicted by Diagnostic Classification

	v
Predictors	Estimates	CI	p
(Intercept)	3.01	2.47 – 3.55	<.001
ADHD [Present-Absent]	0.31	−0.47 – 1.08	.438
DMDD [Present-Absent]	0.17	−0.83 – 1.18	.733
Condition [Congruent-Incongruent]	2.31	1.95 – 2.68	<.001
Condition [Neutral-Incongruent]	2.53	2.17 – 2.89	<.001
Age	0.42	0.32 – 0.52	<.001
IQ	0.04	0.02 – 0.06	<.001
ADHD [Present-Absent] * DMDD [Present-Absent]	−0.36	−1.59 – 0.87	.566
ADHD [Present-Absent] * Condition [Congruent-Incongruent]	−0.44	−0.96 – 0.08	.098
ADHD [Present-Absent] * Condition [Neutral-Incongruent]	−0.62	−1.14 – −0.10	.021
DMDD [Present-Absent] * Condition [Congruent-Incongruent]	−0.85	−1.53 – −0.17	.014
DMDD [Present-Absent] * Condition [Neutral-Incongruent]	−0.92	−1.60 – −0.25	.008
ADHD [Present-Absent] * DMDD [Present-Absent] * Condition [Congruent-Incongruent]	0.79	−0.04 – 1.61	.062
ADHD [Present-Absent] * DMDD [Present-Absent] * Condition [Neutral-Incongruent]	1.07	0.24 – 1.89	.012
Random Effects
σ2	0.80
τ00 Subject	2.70
ICC	0.77
N Subject	187
Observations	561
Marginal R2 / Conditional R2	0.373 / 0.857

Note. Contrasts estimated at reference level for absent factors. Reference levels for factor contrasts are attention-deficit/hyperactivity disorder (ADHD) is absent, disruptive mood dysregulation disorder (DMDD) is absent, and Flanker condition is “Incongruent.” Probability values are estimated via Kenward-Roger approximation.

Figure 2. Effects of diagnostic status and task condition on drift rate v, adjusted for IQ and Age.

Note: For diagnostic codings, 1=Present and 0=Absent. Error bars reflect 95% CI. ADHD = attention dysregulation/hyperactivity disorder; DMDD = disruptive mood dysregulation disorder; HV = Healthy volunteers

Symptom Associations

Additional analyses were conducted on symptoms in the smaller subsample (n=87). For the inattention model, these analyses also revealed a three-way interaction between both symptom dimensions and task condition on drift rate (F(2,166)=3.32, p=.039; see Table 3 for full model, Figure 3). In the incongruent condition, high irritability was associated with a more pronounced decrease in drift rate with increasing levels of inattentiveness, whereas in the neutral and congruent conditions this pattern of moderation was reversed. See also Supplement 4 and Figure S3, available online, for an illustration of raw drift rate data with median split on symptoms.

Table 3.

Mixed Effects Model for Drift Rate Predicted by Symptom Dimension

	v
Predictors	Estimates	CI	p
(Intercept)	3.07	2.47 – 3.66	<.001
ARI_Factor [High-Low]	0.29	−0.64 – 1.22	.533
Inattentiveness	−0.02	−0.09 – 0.05	.489
Condition [Congruent-Incongruent]	2.14	1.74 – 2.54	<.001
Condition [Neutral-Incongruent]	2.16	1.76 – 2.56	<.001
Age	0.44	0.27 – 0.60	<.001
IQ	0.05	0.02 – 0.07	.003
ARI_Factor [High-Low] * Inattentiveness	−0.05	−0.17 – 0.06	.365
ARI_Factor [High-Low] * Condition [Congruent]	−0.54	−1.14 – 0.07	.082
ARI_Factor [High-Low] * Condition [Neutral-Incongruent]	−0.24	−0.85 – 0.36	.429
Inattentiveness * Condition [Congruent-Incongruent]	−0.05	−0.10 – −0.00	.036
Inattentiveness * Condition [Neutral-Incongruent]	−0.07	−0.11 – −0.02	.006
ARI_Factor [High-Low] * Inattentiveness * Condition [Congruent-Incongruent]	0.08	0.00 – 0.16	.043
ARI_Factor [High-Low] * Inattentiveness * Condition [Neutral-Incongruent]	0.10	0.02 – 0.18	.018
Random Effects
σ2	0.83
τ00 Subject	2.69
ICC	0.76
N Subject	87
Observations	261
Marginal R2 / Conditional R2	0.436 / 0.867

Note. Reference levels for factor contrasts low ARI score (median split), and Flanker condition is “Incongruent.” Probability values are estimated via Kenward-Roger approximation. ARI: parent-report Affective Reactivity Index.

Figure 3. Effects of Dimensionally-Assessed Irritability and Inattention and Task Condition on Drift Rate v, Adjusted for IQ and Age.

Note: Error bars reflect 95% CI.

In the hyperactivity model, there were no significant main or interactive effects of symptoms.

For all linear models, all predictors had a corrected generalized variance inflation factor <5, suggesting acceptable levels of multicollinearity for analysis.

Discussion

Here, we conduct the first examination of processing efficiency, via the DDM, in youth with severe irritability accounting for ADHD symptoms. We demonstrate common and distinct associations with drift rate, ADHD, DMDD, and their core symptoms that vary by attentional task demands. With regards to our hypothesis, we did not find that drift rate was simply reduced for those with ADHD; we found adaptation of processing efficiency between non-conflict conditions were larger in youth without psychopathology relative to those with ADHD. This was also true for those with DMDD. Analyses of continuous measures in a subsample revealed interactive effects between irritability and inattention on drift rate.

As expected, the strong correlations occurred between response variability and the drift rate parameter; lower drift rates will result in a wider distribution of reactions times ^14,31. This suggests that the increased RT variability previously noted in sustained attention tasks in youth with ADHD and DMDD ¹⁰ can be partially explained by less efficient processing. Hence, the computational modeling perspective adds to our understanding of the behavioral dysfunction by bringing interpretational clarity to decision computation impairments observed in youth with ADHD and/or DMDD. Of note, age and IQ accounted for large portions of variance in drift rate: with increasing age and IQ, RTs are usually faster and less variable. Future work could explore the gap in our understanding of developmental changes on symptom-drift rate associations.

Though much has been written about the involvement of sustained attention in response time variability in ADHD ^2,12, caution is warranted when interpreting cognitive control functions underlying individual differences in reaction time variance ³. In this study, attentional demands reflect cognitive load influences on processing efficiency ¹⁸. Notably, we find associations with diagnostic status to be modulated by attentional demands, with the most pronounced associations in switching between conflict and non-conflict demands. This may reflect a floor effect on drift rate, where high cognitive load suppresses drift rate to the extent that it obscures diagnostic associations. Despite this limitation, the data suggest symptom-related issues with effectively recruiting cognitive resources by task demands, e.g. cognitive flexibility.

The only previous fMRI study ¹⁰ to examine sustained attention in youth with ADHD and DMDD found blunting in parietal attention networks among both patients with ADHD and DMDD associated with longer trial-wise reaction times. Additionally, the study found DMDDspecific increases in pre-stimulus activation associated with longer trial-wise reaction times in several frontal and parietal regions. Converging with the current findings, this evidence suggests that cognitive control impairments are not specific to ADHD, but also link to DMDD, and perhaps to chronic irritability. Decomposing the cognitive processes of sustained attention reflected in reaction time on the neural level would be a natural next step to aid interpretation of neural findings. Previous work on DDM-measured cognitive processing inefficiency supports hypotheses of neural processing inefficiency (i.e., lower signal to noise ratio in decision-making networks ³). Empirically, efforts have been made to combine neural and computational approaches to map neural signals to specific computations captured in latent parameters ³² or even undertake neurally informed modeling ³³.

There are several limitations to consider when interpreting the findings. First, these results may only be generalized to populations able to learn and complete the Flanker task. A number of youth (~15%) were unable to train to adequate task performance and were not invited to complete the task. Second, with a correlation of r=.49, p<.001, there is significant overlap between the constructs of inattention and irritability, and issues of multicollinearity arise, biasing toward type II error. However, examinations of variance inflation factors found these to within generally acceptable limits (<5) for all terms in all models. Though fast-dm is well validated, it only allows for a constant drift rate. A drift rate that varies with time within a trial might better capture early attentional dynamics in the Flanker task, which are especially prominent in the incongruent condition ¹⁸.

This study was designed for diagnostic comparisons, the symptom examination is secondary to provide information for more future investigations. As such data was available only for a subsample for the current report. These implicate inattention and its interactive effects with irritability. Notably, the diagnostic analysis is meaningful. Diagnoses were determined by semistructured interview including multiple sources of information and arrived at by consensus among experts. They reflect differences in attention/hyperactivity and irritability on which diagnoses of ADHD and DMDD are solely based. Finally, they are highly relevant for clinical practice and directly inform clinicians on the specificity of response time variability to ADHD. Taken together, the diagnostic and symptoms analysis support future investigations to leverage dimensional data in contemporary transdiagnostic approaches.

An interesting future direction is examining cognitive control in an affective, rather than a ‘cold’ cognitive context. Cognitive control processes contribute to the regulation of affective states, i.e. emotional regulation ³⁴ Current models of chronic, severe irritability posit impairments in cognitive control as a potential mediator for experiences of frustration and behavioral manifestations of irritability, i.e., temper outbursts ³⁵. Hence, ‘cold’ cognitive control differences observed in the current study may be magnified in cognitive control tasks where goal conflicts evoke negative affect.

In summary, applying a computational modeling approach, we can map increased variability in reaction times in youth with ADHD and/or DMDD onto difficulties in basic processing of stimuli under different attentional demands. Our results suggest that attentional impairments are not specific to ADHD; rather they may represent shared psychopathology between ADHD and DMDD.