Attention-Deficit/Hyperactivity Disorder Symptoms as a Function of Arousal and Executive Vigilance: Testing Halperin and Schulz’s Neurodevelopmental Model in a Sample of Community Adults

In an attempt to account for the developmental trajectory of the disorder, Halperin and Schulz (2006) elaborated an influential neurodevelopmental model of ADHD. This model postulates a double dissociation in which the onset of ADHD in childhood is due to subcortical brain dysfunctions that remain stable over the lifespan, while the reduction of symptom severity with age is dependent on the development of the prefrontal cortex. In this sense, executive functions, mediated by the prefrontal cortex, are not central to the early emergence of ADHD but influence its developmental course into adulthood. Of note, the model considers the phenomenon of late-onset ADHD. This condition refers to ADHD cases where symptoms were not substantially present in childhood, representing around half of all adults with the disorder (Asherson & Agnew-Blais, 2019). According to the model, late-onset ADHD would be the product of either early or late lesions in the prefrontal cortex subserving executive processes. Therefore, neurocognitive targets for ADHD interventions should differ as a function of individuals’ age (children vs. adults) and time of symptom onset (early vs. late).

The neurodevelopmental model generates testable predictions in measures of cognitive tasks. First, the subcortical brain dysfunctions should be observable in measures with minimal or no executive load, such as basic reaction time (RT) or RT variability in simple RT tasks. These measures should distinguish ADHD children from controls across the lifespan, regardless of adolescent or adult status. Second, measures of executive function, such as target identification or inhibition tasks, should be dimensionally related to the severity of ADHD symptomatology, particularly in adulthood.

Initial support for the neurodevelopmental model came from a prospective longitudinal study of 98 children with ADHD who were reassessed after 10 years (Halperin et al., 2008). Compared to matched controls, the adolescents diagnosed with ADHD in childhood showed deficient response variability, irrespective of their current clinical status. However, only adolescents with persistent—but not remitting—ADHD exhibited poorer inhibition and working memory than controls. Despite the promising findings of this study, subsequent research has yielded mixed results (Cheung et al., 2016; Coghill et al., 2014; Coll-Martín et al., 2021; Gmehlin et al., 2016; Leenders et al., 2021; McAuley et al., 2014; Michelini et al., 2016; Rommel et al., 2015). Moreover, a systematic review found no support for the model, as the neurocognitive deficits observed in individuals with ADHD across development were similar for high- and low-executive-load measures (van Lieshout et al., 2013). Further challenging the model, a more recent review (Franke et al., 2018) suggested that low-level cognitive processes, rather than high-level ones, are associated with remission/recovery of the disorder.

Although the current literature is not very encouraging with the neurodevelopmental model, some methodological issues are worth considering. First, most of these studies consisted of case–control comparisons that were underpowered to detect a range of effects that could be considered theoretically relevant (e.g., 0.35 ≥ Cohen’s d_s ≥ 0.2). In contrast, population-based or community samples can better fit the dimensional nature of ADHD (Coghill & Sonuga-Barke, 2012; Hilger et al., 2020) and are more efficient in collecting well-powered samples. Second, studies that included low-executive-load measures rarely used tasks with no executive component (for exceptions, see Coll-Martín et al., 2021; Gmehlin et al., 2016). In this sense, even RTs or omission errors in tasks such as the continuous performance test (CPT) are influenced by the executive processes and the response criterion involved in the task demands. The next section will address this issue.

Impaired vigilance is central to the phenotypic and neurocognitive characterization of ADHD (Huang-Pollock et al., 2012; Wilding, 2005; Willcutt et al., 2005). The construct is generally defined as the attentional capacity to maintain performance over time. Given the diversity of terms and measures linked to vigilance, some authors deem it as a nonunitary concept (Langner & Eickhoff, 2013; Luna et al., 2018; Sturm et al., 1999). In this vein, Luna et al. distinguish two components of vigilance: executive vigilance (EV) and arousal vigilance (AV).

EV is the ability to detect infrequent but critical signals among nonsignal stimuli. It is measured with tasks derived from the CPT paradigm such as the AX-CPT (Rosvold et al., 1956) or the Test of Variables of Attention (TOVA; Greenberg & Waldman, 1993). These types of tasks seem to involve executive mechanisms of sustained attention for stimuli discrimination and goal-oriented response selection. Notably, response-inhibition CPTs (e.g., Conners’s CPT [Conners, 2000], Sustained Attention to Response Task [SART; Robertson et al., 1997]), the most common CPT variant in ADHD research, also require motor suppression of the preponderant response in the presence of the target stimulus. Key measures of EV are hits (inverse of omission errors) and false alarms, both tending to decrease with time on task (for a discussion about false alarms, see Thomson et al., 2016).

AV is the capacity to sustain a rapid reactivity to any environmental stimulus—without implementing any control over the selection of the response executed. This form of vigilance is measured in simple RT tasks, such as the Psychomotor Vigilance Test (PVT; Lim & Dinges, 2008) and the WAFA test of the Vienna Test System (Schuhfried, 2013). These tasks seem to record an arousal component of vigilance, a mechanism that could be more related to physiological levels of excitability. The main indices of AV are the mean and variability of the RT and the attentional lapses. Contrary to EV, this vigilance decrement manifests as an increase in the measures during the task.

In order to evaluate both vigilance components simultaneously, Luna et al. (2018) designed the Attentional Networks Test for Interaction and Vigilance—Executive and Arousal Components (ANTI-Vea). This task is based on the Attention Networks Test for Interaction (ANTI; Callejas et al., 2004), which combines an Eriksen flanker paradigm with spatial cues and warning signals to assess the three attentional networks (Posner & Petersen, 1990). In ANTI trials, which are the bulk of the ANTI-Vea (60% of trials), participants must respond to the direction pointed by a central arrow, flanked by either congruent or incongruent arrows. To assess EV, in a small percentage of trials (20%) the central arrow appears vertically displaced for participants to detect and respond to it, thereby suppressing their preponderant response to ANTI trials. To measure AV, in another small percentage of trials (20%), a salient stimulus (i.e., a red down counter) is displayed for participants to stop it as fast as possible. Of note, the length of the task (~33 min) successfully induces a decrement in all vigilance measures (Luna et al., 2018).

Several studies have found that EV and AV are dissociable in the ANTI-Vea, both as a result of experimental manipulations (Feltmate et al., 2020; Hemmerich et al., 2023; Sanchis et al., 2020) and in relation to individual differences (Cásedas et al., 2022; Román-Caballero et al., 2021). In the context of ADHD, Coll-Martín et al. (2021) conducted a study in which the ANTI-Vea¹ was administered to 113 university undergraduates in a laboratory setting. They assessed ADHD symptom severity retrospectively in childhood and concurrently in adulthood through self-reports. In line with the neurodevelopmental model, the authors found that symptoms in childhood correlated with a higher increase in lapses (AV), while adult symptoms were associated with a greater decrease in hits during the task (EV). Although promising, this pattern of results was not predicted and came from a study with several measures and multiple contrasts per theoretical hypothesis, thereby arising in a context of discovery and exploration. To be derived from confirmatory testing, the dissociation observed should be formally replicated in a larger sample and with stricter control of false positive rates.

This study aimed to test the neuropsychological predictions of Halperin and Schulz’s (2006) neurodevelopmental model of ADHD from a dimensional framework. For this purpose, a community sample of university students was assessed for childhood and adult symptoms of ADHD and performed the online version of the ANTI-Vea to measure AV and EV. Although having different aims, this procedure can be considered a close replication of Coll-Martín et al. (2021), with the main difference being the setting in which the ANTI-Vea was administered (but see Luna et al., 2021, for the remarkable psychometric similarities between the lab and online versions). In contrast to most previous studies, our design incorporated certain methodological considerations such as the use of a valid nonexecutive measure, which are arguably crucial to properly testing the model. Unlike Coll-Martín et al., the larger sample size and the more specific contrasts planned for this replication conferred a reasonably appropriate severity to the statistical tests of the model predictions.

Figure 1 illustrates all the hypotheses tested in the present research. According to the neurodevelopmental model and the results of the original study (Coll-Martín et al., 2021), we had two general predictions: 1) Childhood ADHD symptoms would associate with higher lapses (AV), and 2) adult ADHD symptoms would associate with lower hits (EV) after accounting for symptoms in childhood (i.e., after controlling for baseline to focus on later development). For each of these predictions, we preregistered primary (H1 and H2) and secondary (Sc1 and Sc2) hypotheses (https://osf.io/tkdq7). The distinction between primary and secondary was based on the type of measure selected to operationalize vigilance for each hypothesis.

For primary hypotheses, the operationalization of AV and EV relied on measures of vigilance decrement, as its relationship to ADHD symptoms in the original study (Coll-Martín et al., 2021) yielded the neurodevelopmental dissociation that our study aimed to replicate. Of note, vigilance decrement has been considered the core feature of the construct (Huang-Pollock et al., 2012; Tucha et al., 2017). For secondary hypotheses, we based on overall vigilance scores of AV and EV, since they are the measures commonly employed in ADHD research and usually report high reliability (Coll-Martín et al., 2021, 2023; Huang-Pollock et al., 2012). In parallel, we also preregistered the set of opposite hypotheses (H1′, H2′, Sc1′, and Sc2′), that is, those predictions relating each vigilance component to ADHD symptom severity in the age period opposite to that established in the neurodevelopmental model.² Support for the opposite hypotheses would generally be interpreted as evidence against the model.

Besides the preregistered hypotheses, which focused on the severity of the tests (Mayo, 2018), the present study also analyzed the robustness of their findings (Nosek et al., 2021) in relation to the predicted dissociation. To do so, we conducted multiverse analyses (Steegen et al., 2016) for each of the hypotheses tested and explored the generalizability of the neurodevelopmental model’s predictions to the sets of arousal and executive measures of the ANTI-Vea. While our multiverse analyses mainly focused on the fragility of the results across variations in data processing and statistical modeling, our exploratory analyses examined the sensitivity of the findings to broader and more diverse operationalizations of arousal and executive processes.

In order to report the severity of the tests transparently, the study design and analysis plan were publicly preregistered with the Preregistration for Quantitative Research in Psychology Template (PRP-QUANT; Bosnjak et al., 2022) at https://osf.io/tkdq7 (for a list of deviations from the preregistration, see Table S1). The preregistered hypotheses in that document are denoted with the same label as in the present paper.

Figure 2 illustrates the sample selection process (for a more detailed description of the exclusion criteria, see Table S2). In Phase 1, we collected a total of 2,003 responses to a 15-min online survey via LimeSurvey (https://www.limesurvey.org). Participants were recruited through advertisements on the virtual distribution list of our university. In exchange for completing the survey, participants had the opportunity to win one of two prizes of €200 in a raffle. The survey included three self-reports of ADHD symptoms (see the Instruments section) as well as questions about sociodemographic, clinical, and other psychological variables (see the preregistration document for further details). The survey randomized the appearance order of the two ADHD self-reports that were employed in the analyses of the preregistered hypotheses. After exclusions of survey respondents, there were 1,540 eligible participants (72.7% women, 26.1% men, 1.23% nonbinary; 18–35 years, M = 22.5, SD = 3.7) at the end of Phase 1.

Eligible participants for Phase 2 were randomly invited to perform the online version of the cognitive task (i.e., the ANTI-Vea). Invitations were sent via email until approximately 300 responses were collected. We stratified by sex to ensure a representative distribution of the population in that factor.

A total of 318 participants (50% women, 50% men) enrolled in Phase 2. They received the link to the web-based version of the ANTI-Vea with the following instructions before starting the task: (a) sit in a comfortable place without distractions; (b) keep entertainment devices (i.e., television, radio, mobile phone, etc.) out of reach; (c) set the computer’s sound level at 75% and do not minimize the automatic full-screen mode of the task; (d) if applicable, wear glasses or contact lenses; and (e) if necessary, solve any particular issue before starting so that the task can be completed without any breaks. Upon completion of the ANTI-Vea, participants filled out a brief set of questions about the task (i.e., questions related to the task experience, which were not part of this study, and control items) and were reassessed for adult ADHD symptoms³ with the same self-reports as in Phase 1. The duration of Phase 2 was about 1 hr with 40 min dedicated to performing the task. Participants were compensated €6 for their voluntary participation.

The entire data collection took place fully online from June 2021 to June 2022. All participants completed an informed consent form at the beginning of the study. The research project was approved by our institutional ethics committee.

The Barkley Adult ADHD Rating Scale-IV (BAARS-IV; Barkley, 2011) includes two self-reports to assess ADHD symptoms: retrospectively in childhood (cBAARS-IV) and concurrently in adulthood (aBAARS-IV). Each scale is composed of 18 items, nine for inattention (e.g., “Difficulty sustaining my attention in tasks for fun activities”) and nine for hyperactivity–impulsivity (e.g., “Shift around excessively or feel restless or hemmed in”), in a Likert format ranged from 1 (never or rarely) to 4 (very often). While the cBAARS-IV refers to behaviors between 5 and 12 years of age, the aBAARS-IV refers to the last 6 months. Since the items are based on the Diagnostic and Statistical Manual of Mental Disorders (4th ed.; DSM-IV; APA, 1994), we used the Spanish version of the manual for the translation (APA, 1994/1995), as we did in the previous study (Coll-Martín et al., 2021). In our final sample, Cronbach’s alpha reliability scores were .90 and .86 for cBAARS-IV and aBAARS-IV, respectively, close to the .95 and .92 of the original BAARS-IV (Barkley, 2011). Barkley proposed the 95th percentile of these scales as a cut-off to identify individuals at high risk of ADHD.

The Adult ADHD Self-Report Screening Scale for DSM-5 (ASRS-5; Ustun et al., 2017) assesses the adult-specific presentation of ADHD symptoms based on DSM-5 conceptualization (APA, 2013). It includes six items (e.g., “How often do you put things off until the last minute?”) on a 5-point Likert scale (0 = never to 4 = very often). As in the aBAARS-IV, the questions in this scale refer to behaviors that have occurred over the last 6 months. We used the Spanish version of the ASRS-5 that was administered in the original study (Coll-Martín et al., 2021). Cronbach’s alpha reliability scores in our final sample were .60, which is close to the .64 from our previous study. Ustun et al. established a threshold of 14 points in the ASRS-5 as the preferred bound for screening purposes.

The ANTI-Vea (Coll-Martín et al., 2023) is a cognitive task that provides measures of the functioning of the three attentional networks (alerting, orienting, and executive control), along with measures of AV and EV (overall performance and decrement across time on task). The simultaneous assessment of these processes confers control of the order effect bias. We used the online version of the ANTI-Vea, which is freely available in multiple languages at https://anti-vea.ugr.es (Coll-Martín et al., 2023), and whose psychometric properties do not differ substantially from the lab version (Luna et al., 2021; see also Coll-Martín et al., 2023). The ANTI-Vea comprises three types of trials: ANTI (60%), EV (20%), and AV (20%). The stimuli sequence and correct responses for each type of trial are depicted in Figure 3 (for a more detailed description of the task with audiovisual material, see the Method section of the ANTI-Vea website at https://anti-vea.ugr.es/method.html).

The ANTI-Vea starts with a practice phase, in which instructions and feedback are given so that the participant can gradually familiarize themselves with each type of trial. The participant is required to respond as quickly and accurately as possible while keeping their eyes on the fixation point until the finalization of the task. After the practice phase, the participant performs the task itself, consisting of seamless blocks of 80 trials each (48 ANTI, 16 EV, and 16 AV). The presentation of the trials is pseudorandomized, ensuring that the conditions and types of trials are consistent across blocks. In the present study, we administered the standard format of six blocks (480 trials in total for the whole task).

Regarding our study hypotheses, we focused on two types of ANTI-Vea measures. For AV, we assessed lapses, defined as trials with an excessively slow RT (i.e., RT > 600 ms) or no response to the red down counter. For EV, we measured hits, defined as trials in which the infrequent displacement of the central arrow is correctly detected. For each of the two types of measures, we considered both the overall task performance (secondary measures) and the slope of the vigilance decrement across the six blocks of the task (primary measures). This decrement manifests as an increase in the lapse rate and a decrease in the hit rate across blocks. The internal consistency scores of the four indices were estimated using a permutation-based split-half approach (Parsons et al., 2019) with 10,000 random splits. The Spearman-Brown (SB) corrected coefficients in our final sample (see Table 2) were arguably close to the corresponding .96, .94, and .78 obtained in the original study (Coll-Martín et al., 2021) for lapse overall, hit overall, and lapse slope, respectively. The exception was the slope of hits, for which the reliability score in our sample (r_SB = .61) was substantially higher than in the original one (r_SB = .27).

As preregistered, data were preprocessed and analyzed based on the original study (Coll-Martín et al., 2021). The entire workflow has been run and documented in three reproducible R scripts (Version 4.2.1). We used the tidyverse collection of R packages (Wickham et al., 2019) for most of the data treatment and output visualization.

For the selection of ADHD symptom measures, we considered the period elapsed from the initial survey until the completion of the task. If this time interval was 90 days or less (44.2% of participants from the final sample), we used the questionnaires from Phase 1. Conversely, if the period exceeded 90 days (53.8% of participants from the final sample), we used the adult ADHD self-reports completed in Phase 2.⁴ The score for each ADHD symptom scale (i.e., the cBAARS-IV, the aBAARS-IV, and the ASRS-5) was the sum of its items. Applying the procedure of the original study (Coll-Martín et al., 2021), the distribution of the cBAARS-IV and the aBAARS-IV scores were compared with a simulated normative sample.

Figure 2 shows participants excluded from the ANTI-Vea analyses. Compared to the final sample, the level of ADHD symptoms among the participants excluded due to poor performance (i.e., more than 25% errors in ANTI trials) was slightly lower for symptoms in childhood (d_cBAARS-IV = –0.16) and negligibly to slightly higher for adult symptoms (d_aBAARS-IV = 0.03; d_ASRS-5 = 0.22). We calculated the percentage of lapses in AV trials and the percentage of hits in EV trials for each participant. Of the total number of lapses, 87.2% were due to excessively slow RT, while the remaining 12.8% were due to no response. For both lapse and hit percentages, vigilance decrement was calculated by estimating the linear slope of the scores across the six blocks of the task. We used the plyr R package (Wickham, 2022) to compute the reliability of EV and AV measures.

As for the operationalization of our preregistered hypothesis (Figure 1), a greater vigilance decrement across time on task (primary vigilance measures) manifests as a more positive slope in lapses and a more negative slope in hits. Therefore, we hypothesized a positive correlation between the slope of lapses and childhood symptom severity measured with the cBAARS-IV (H1), as well as a negative correlation between the slope of hits and adult symptom severity measured with the ASRS-5 (H2). Similarly, poorer overall vigilance (secondary vigilance measures) is reflected as more lapses and fewer hits. Hence, we hypothesized a positive correlation between the overall percentage of lapses and the cBAARS-IV (Sc1), as well as a negative correlation between the overall percentage of hits and the ASRS-5 (Sc2). To focus more closely on the association with late-developing symptoms, which arguably fits better with predictions from the neurodevelopmental model, H2 and Sc2 were partial correlations controlled for the cBAARS-IV. This operationalization procedure was also applied to the opposite hypotheses (H1′, H2′, Sc1′, and Sc2′).

Since bivariate normality was violated for H1 and H2 (both ps < .001), we tested all the preregistered hypotheses using Kendall’s rank-order correlation coefficients, computed with the correlation R package (Makowski et al., 2022). We conducted one-tailed tests at a significance level of α = .05. For secondary hypotheses (Sc1 and Sc2) and their opposite counterparts (Sc1′ and Sc2′), the threshold was adjusted for multiple comparisons to α = .025, corresponding to a Bonferroni correction for two statistical tests (i.e., primary and secondary hypotheses). Moreover, we relied on the one-tailed limit of the 95% confidence interval (CI), corrected for multiple comparisons when applicable, to test whether the effect size observed for each correlation was significantly inferior to the effect sizes of interest⁵ (for the effects of interest of each hypothesis, see Table 1 and Table S2).

We analyzed the robustness of the results by performing a multiverse analysis (Steegen et al., 2016) for each of the hypotheses above. To generate the multiverse of analyses, we identified five decision points in the analytical process with more than one apparently reasonable choice, the first being the preregistered option: participants with more than 90 days from survey to task (n = 6; retained vs. excluded), poor task performers (n = 14; excluded vs. retained), task threshold to compute the vigilance indices (absolute vs. relative), correlation coefficient (Kendall vs. Spearman vs. Pearson) and type of correlation (zero-order vs. partial⁶). Furthermore, we incorporated the aBAARS-IV as a secondary measure of symptoms in adulthood (preregistered as Ss2). This provided 48 valid specifications to analyze the correlation between each of the four vigilance indices and each of the three ADHD measures (i.e., 576 total estimates).

Finally, we examined the relationships between ADHD symptoms and the different ANTI-Vea measures of arousal or executive processes. To do so, we computed all core task indices for these neurocognitive domains. They included eight measures of arousal: lapses in AV trials (slope and overall), mean RT in AV trials (slope and overall), standard deviation of the RT in AV trials (slope and overall), and the alerting index—no tone minus tone conditions— in ANTI trials with no cue (RT and percentage of errors). They also included six executive measures: hits in EV trials (slope and overall), false alarms in EV trials (slope and overall), and the congruency index—incongruent minus congruent conditions—in ANTI trials (RT and percentage of errors). The set of 14 ANTI-Vea indices was used to predict each of the three measures of ADHD symptoms (i.e., the cBAARS-IV, the ASRS-5, and the aBAARS-IV) through multiple linear regression models. To select the predictors of each model, we employed a bidirectional stepwise regression method aimed at minimizing the Akaike information criterion (AIC).

For the sample size justification, we followed Lakens’s (2022) guidelines (see Text S1 for a more detailed report before data analysis). The sample size collected in the first and second phases of the study (see Figure 1) was determined by the financial resources provided by our funders for participant payment. Our final sample consisted of 292 university students (49.0% women, 51.0% men; 18–30 years, M = 21.7, SD = 2.7; see Figure 4A). Most of the participants were Spanish (95.2%) and current university students (92.1%).⁷ Five individuals (1.7%) reported a prior diagnosis of ADHD.

To estimate the statistical power of our preregistered hypotheses (see Text S2 for a detailed report of the final procedure), we relied on two types of effect sizes: the smallest effect size of interest (SESOI) and the expected effect size. As preregistered, the true SESOI was set at ρ = .20 or ρ = –.20 (depending on the hypothesis direction) and the expected effect sizes were based on the original study (Coll-Martín et al., 2021) after applying Perugini et al.’s (2014) correction for replications. Considering the effect size attenuation due to measurement error, which is particularly noticeable in cognitive tasks (Parsons et al., 2019), our power analyses were based on the effect size observed when accounting for the reliability of the measures involved in each correlation. The results of the power analyses for the predictions derived from Halperin and Schulz’s (2006) neurodevelopmental model are shown in Table 1 (for the power analyses of the opposite hypotheses, see Table S3). Overall, our sample size enabled our preregistered hypotheses to achieve an arguably acceptable statistical power for at least one effect of interest.

Figure 4B–D shows the distribution of ADHD symptoms in the three scales, and Figure 4B–C also compares it to an estimated normative sample (for a detailed procedure and statistical report, see Text S3). Taken together, ADHD symptoms in our sample of young university students were higher than in the general population of young adults. Even so, the spread and variability within each scale did not seem to differ from those observed in the estimated normative sample.

Unsurprisingly, the cBAARS-IV (M = 31.37, SD = 9.82), the aBAARS-IV (M = 31.16, SD = 8.14), and the ASRS-5 (M = 8.74, SD = 3.66) showed significant positive correlations among them, with effect sizes that are considered large in the field (Gignac & Szodorai, 2016). Concretely, for the cBAARS-IV with the aBAARS, r(290) = .53, p < .001, for the cBAARS-IV with the ASRS-5, r(290) = .34, p < .001, and for the aBAARS-IV with the ASRS-5, r(290) = .75, p < .001. As expected, the correlation between the two measures of symptoms in adulthood was higher than those between these measures and the one of symptoms in childhood (both ps < .001). Additionally, the correlation with childhood symptoms was stronger when the adult symptom items also corresponded to DSM criteria, as in the case of the aBAARS-IV, than when they were intended to reflect the adult-specific presentation of ADHD symptoms, as with the ASRS-5 (z = 5.11, p < .001).

Table 2 shows the means and standard deviations of the percentages of lapses and hits for both the vigilance decrement and the overall performance. In line with the original study (Coll-Martín et al., 2021), we found a positive linear slope of lapses, t(291) = 7.18, d_z = 0.42, and a negative linear slope of hits, t(291) = –9.83, d_z = –0.58, across time on task. Specifically, the lapses increased from 7.96% in Block 1 to 13.38% in Block 6, while the hits decreased from 78.94% in Block 1 to 68.90% in Block 6. Interestingly, the correlation between lapses and hits was not significant for either the overall scores (r = –.09, p = .10) or the slopes of vigilance decrement (r = –.11, p = .07), revealing a statistical dissociation between both vigilance components.

Table 2 shows the results of the preregistered hypotheses. For primary vigilance measures (i.e., vigilance decrement), none of the statistical hypotheses—namely H1, H2, H1′, and H2′—were statistically significant (all ps > .194). Even more so, each of these four correlation coefficients was significantly smaller in size than at least one of the two effects of interest (all ps < .05). Specifically, all correlation sizes, except that of H2, were lower than their corresponding SESOI. Moreover, the coefficients from H1 and H2 were smaller than their expected effect sizes.

Regarding the secondary vigilance measures (i.e., overall vigilance), the expected positive correlation between the cBAARS-IV and the percentage of lapses (Sc1) was only significant before correcting for multiple comparisons (τ = .07, p_uncorrected = .047, p_corrected = .094). In addition, the hypothesized negative correlation between the ASRS-5 and the percentage of hits (Sc2) was not significant (τ = –.04, p_uncorrected = .128). As expected, the correlation between the ASRS-5 and the percentage of lapses (Sc1′) was not significant (τ = –.00, p_uncorrected = .523), and its effect size was smaller than the SESOI (τ_SESOI = .10; p_corrected < .05). Surprisingly, the negative correlation between the cBAARS-IV and the percentage of hits (Sc2′) was significant (τ = –.09, p_corrected = .033).

The distribution of the estimates within the four multiverses is illustrated in Figure 5 (for a more comprehensive report of the method and results of these analyses, see Table S5). Figure 5A–B shows that the lack of significant findings in the preregistered H1 and H2 was robust across the different analytical options, including the use of a secondary measure of ADHD symptoms in adulthood (i.e., the aBAARS-IV). Regarding their opposite counterparts, we found that 20 out of the 96 analytical scenarios in H1′ (20.8%) were statistically significant. A closer inspection revealed that most of these significant findings (90.0%) computed the lapse index based on a relative threshold (i.e., trials with an RT higher than the participant’s mean plus one standard deviation or trials with no response). This type of lapse index computation exhibited very low reliability scores, albeit probably biased upwards when poor performers were retained for the analyses (see Figure S1). On the contrary, for the analytical combinations that employed an absolute threshold to compute lapses, the rate of significant findings was below chance (2.1%). As for H2′, the null results were consistent across all analytical choices.

In contrast to the primary hypotheses, secondary hypotheses (Figure 5C–D) showed a rather mixed pattern of statistical significances across analytical scenarios. Sc1 yielded 8.3% significant contrasts, which doubled to 16.7% when only analyses with an absolute threshold for lapses were considered. Regarding Sc2, the percentage of statistically significant scenarios dramatically varied from 16.7% in the ASRS-5 to 95.8% in the aBAARS-IV. This difference between adult measures of ADHD symptoms was also manifested in Sc1′, where 20.8% and 54.2% analytical combinations were significant for the ASRS-5 and the aBAARS-IV, respectively. Sc2′, the only statistical hypothesis whose preregistered contrast was significant, yielded 35.4% scenarios with positive findings. Since most of our analytical options did not used the linear regression model, the robustness of the secondary hypotheses could not be inferentially evaluated with known available approaches (e.g., Simonsohn et al., 2020).

Table S6 shows descriptive statistics and correlations among the ANTI-Vea core indices that measure arousal or executive processes. Regarding their association with ADHD symptoms, the three multiple regression analyses (Table 3) showed that three, six, and five ANTI-Vea indices were selected to predict the scores of the cBAARS-IV, the ASRS-5, and the aBAARS-IV, respectively. None of the models presented outliers or relevant multicollinearity issues. Crucially, both arousal and executive task indices uniquely predicted ADHD symptoms across the three models.

Are alterations in AV associated with higher ADHD symptoms in childhood (H1), while deficits in EV are related to late-developing ADHD symptoms in adulthood (H2)? This dissociation, predicted by Halperin and Schulz’s (2006) neurodevelopmental model, was supported by Coll-Martín et al. (2021). Here we designed a preregistered, closed replication of that study specifically aimed at testing the model. Based on a dimensional framework of ADHD, we assessed retrospective and current self-reported symptoms in a community sample of university students who performed an online cognitive task: the ANTI-Vea. The final sample size (N = 292) allowed our preregistered hypotheses to achieve an arguably acceptable statistical power for the effects of interest. Despite this, all these hypotheses failed to replicate the previous findings: Indeed, only the unexpected (i.e., opposite to the neurodevelopmental model) negative correlation between EV and symptoms in childhood was significant. Although multiverse and exploratory analyses yielded some significant results, neither did they support the dissociation pattern proposed by the neurodevelopmental model.

Our unsuccessful replication of Coll-Martín et al.’s (2021) findings was rather clear and robust. Not only were our correlation coefficients for H1 and H2 not significantly different from zero, but they also were significantly smaller in size than an attenuated estimation of the effect sizes in the original study. Even testing H2 as a simple correlation, as conducted in the original study, was not significant. The minimal conceptual differences between the original study and this replication—namely, the task setting (lab vs. online) and the composition of the university sample (mostly women from the Psychology degree vs. a sex-balanced sample from the entire student community)—were unlikely to account for the huge discrepancies in the results.⁸ Instead, it is far more plausible that the findings of H1 and H2 in the original study were false positives, given the exploratory context in which this pattern was discovered. Beyond the developmental dissociation that our study aimed to replicate, the complete pattern of significant correlations between ADHD symptoms and core ANTI-Vea indices in the present study differs substantially from that found in the original one (Table S7). Taken together, this highlights the importance of close replications with adequate control of statistical error rates to probe stability and improve confidence in original findings (Simons, 2014; Tunç & Tunç, 2023).

In contrast to Coll-Martín et al. (2021), our study failed to support the neurodevelopmental model. Notably, unlike most previous empirical research, our study selected a clear and valid nonexecutive measure to test the first prediction of the model. Specifically, the AV indices from the ANTI-Vea are thought to reflect the noradrenergic mechanisms of the hindbrain mediating arousal (Luna et al., 2018). According to the neurodevelopmental model, alterations in this subcortical system are a potential cause of ADHD and remain stable throughout the lifetime. The conceptual and empirical dissociation of AV from EV in the ANTI-Vea—as evidenced by the low correlation found in our data—supports the task as a useful tool in the neurocognitive research on ADHD.

Another measurement issue, in this case related to the second prediction of the neurodevelopmental model, was the use of two conceptually different scales of ADHD symptoms in adulthood. Our primary measure was the ASRS-5, which assesses the adult presentation of ADHD, including adult-specific items (e.g., frequency of difficulties in unwinding and relaxing when having time to oneself). In addition, we administered the aBAARS-IV as a secondary measure, which consists of traditional DSM items with identical description and wording (e.g., frequency of fidgeting with hands or feet or squirming in seat) as its childhood counterpart (i.e., the cBAARS-IV). While the debate between using adult-specific versus DSM-based items for ADHD diagnosis has been addressed elsewhere (e.g., Sibley et al., 2012), differences in the sensitivity of these measures to neurocognitive variables have not been formally studied. Although descriptively, our multiverse and exploratory analyses suggest that DMS-based items could be somewhat more sensitive to attentional functioning than adult-specific items, especially considering that the former measure is potentially more affected when controlling for childhood symptoms. Alternatively, the higher reliability and skewness of the aBAARS-IV compared to the ASRS-5 could explain such differences in the size of the correlations.

We used measures of vigilance decrement for our primary hypotheses (H1 and H2) and their opposite counterparts (H1′ and H2′). Compared to overall vigilance scores, the decline in performance over time has been less studied in ADHD, despite being considered by some as the defining feature of sustained attention (Huang-Pollock et al., 2012; Tucha et al., 2017). Surprisingly, our four correlation coefficients were significantly smaller in size than at least one of their corresponding effects of interest. Furthermore, the lack of a correlation statistically different from zero in the expected direction was robust across the sets of reasonable analytical options, as shown in the multiverse analyses. Only one exploratory analysis in our stepwise multiple linear regression, namely the positive association between the slope of false alarms and childhood symptoms, was significant.

Considering that our analyses had an acceptable statistical power, our lack of significant results for the relationship between vigilance decrement and ADHD symptoms was theoretically unexpected. However, they align with other clinical and community studies that used moderate to large samples (i.e., N > 150; Aduen et al., 2020; Huang-Pollock et al., 2012, 2020). In contrast, one study comparing ADHD children with controls (n > 200 per group) found that the former had a higher vigilance decrement in response speed and consistency (each d_s > 0.3; Weyandt et al., 2017). While this discrepancy needs to be solved, one possibility might be that vigilance decrement is not substantially related to ADHD symptomatology. On the contrary, momentary attentional fluctuations, the other component of vigilance (Esterman & Rothlein, 2019), could be behind the impaired sustained attention in ADHD. In our study, preliminary support for this idea comes from the association between a higher rate of lapses and higher ADHD symptom severity in 30.1% of the multiverse scenarios with an absolute lapse threshold as well as in the exploratory multiple regression models.

Looking at the big picture, the most consistent pattern across our results was the failure to support the neurodevelopmental dissociation proposed by Halperin and Schulz’s (2006) model. In our preregistered analyses, we found that the rate of hits (EV) negatively correlated with symptoms in childhood but not in adulthood, which is the opposite of the theoretical prediction. This lack of support for the neurodevelopmental model was robust to our multiverse and exploratory analyses. The former either showed null results for virtually all analytical options (primary hypotheses) or yielded mixed findings that were arguably similar for predictions derived from the neurodevelopmental model and for their opposite counterparts (secondary hypotheses). In the same vein, our exploratory regression analyses were far from suggesting any dissociation: Both arousal and executive processes independently predicted ADHD symptoms in childhood and adulthood. Although contrary to Coll-Martín et al. (2021), this lack of support for the model is in line with several studies (Arora et al., 2020; Coghill et al., 2014; Gmehlin et al., 2016; McAuley et al., 2014; van Lieshout et al., 2013).

Going further, other neurocognitive phenomena relevant to the neurodevelopmental model are arguably better explained by alternative theoretical accounts. For example, the neurodevelopmental model posits the prefrontal cortex subserving executive functions as the only area responsible for changes in ADHD symptoms across the lifespan. This contrasts with research suggesting that executive functions are implemented in distinct neural networks involving multiple brain regions (e.g., Dosenbach et al., 2008). Furthermore, the neurodevelopmental model proposal of brain lesions or insults as causes or developmental moderators of ADHD is inconsistent with the current neurodiversity model, which considers ADHD as the extreme expression of a temperamental trait (Sonuga-Barke & Kostyrka-Allchorne, 2022; Sonuga-Barke & Thapar, 2021). The latter model can provide a more parsimonious explanation for why brain stimulation through modern neurotherapeutics or computer-based cognitive training has shown, at best, limited effects on core ADHD symptoms (Rubia, 2022; Westwood et al., 2023). In addition, the heterogeneity of neurocognitive alterations associated with ADHD symptoms within each stage of development and its change trajectories is better accounted for by models of multiple developmental pathways (Nigg et al., 2005; Sonuga-Barke, 2005).

Taken together, it can be stated that Halperin and Schulz’s (2006) neurodevelopmental model in its current form has a substantially lower explanatory capacity than other alternative theories in the field. More broadly, assuming that there is no developmental dissociation in the neurocognitive processes underlying ADHD symptoms has important implications. If late-developing symptoms are pathophysiologically similar to symptoms in childhood, then late-onset ADHD cases—or a substantial proportion of them—would be close to child-onset ADHD. This runs counter to the idea that late-onset ADHD reflects an underlying distinct condition that is less neuropsychologically impaired than conventional ADHD (Moffitt et al., 2015; Sonuga-Barke et al., 2023).

We have identified three main limitations in the design of our study. The first regards the generalization of our findings. Although our community sample was sex balanced, it was mainly made up of university students, which is not fully representative of the young or general adult population. Indeed, ADHD symptoms in childhood have been associated with lower educational attainment in adulthood (Galéra et al., 2012; Pingault et al., 2011). Despite this sampling bias, our statistical analyses did not suggest that ADHD symptom scores were lower or more homogeneous in our sample than in a representative community sample. In any case, replication of our findings with a more representative sample is warranted. Additionally, although the dimensional framework assumes that the neurocognitive mechanisms of ADHD symptoms remain constant throughout a continuous trait, the empirical extension of our results to case–control studies is crucial to strengthen their implications in clinical practice.

Second, our assessment of ADHD symptoms in childhood and adulthood relied solely on self-reports, with the measurement of childhood symptoms being retrospective. Although self-reported measures of ADHD symptoms may capture a broader dimension, they seem less sensitive than parent reports to some neurocognitive measures (Du Rietz et al., 2016; Riglin et al., 2022). In addition, the validity of retrospective reports of ADHD symptoms has been questioned due to potential recall bias, with longitudinal studies finding a modest correlation between prospective and retrospective parent ratings of symptoms in childhood (e.g., r = .39; von Wirth et al., 2021). However, given the fluctuating trajectory of ADHD symptoms throughout development (Sibley et al., 2022; Stern et al., 2020), part of the mismatch between both measures may be better explained by differences in the time span assessed by each symptom scale (e.g., last 6 months vs. whole childhood) rather than recall bias. Of note, Lundervold et al. (2021) found a 7-year test–retest reliability score of .89 for a retrospective self-report measure of ADHD symptoms in childhood. Looking on the bright side, our study design held constant the rater and the time of assessment, thereby controlling for biases related to these factors. In any case, the integration of distinct but complementary assessment methods is fundamental to advancing the understanding of the neurocognitive processes associated with ADHD symptoms across the lifespan.

The third limitation concerns the statistical conclusion validity of our preregistered hypotheses. Our sample size, which was relatively large in this literature, enabled these hypothesis tests to achieve a level of statistical power that can be deemed acceptable. However, this sample would have been underpowered to test differences between correlations (Diedenhofen & Musch, 2015; Hittner et al., 2003),⁹ arguably the most pertinent analysis for the predictions of the neurodevelopmental model in our design. In any case, based on descriptive comparisons within our sample, none of the differences in the correlation coefficients between the hypotheses of the neurodevelopmental model and their opposite counterparts (e.g., τ in H1 minus τ in H1′; see Table 2) exceeded the size of any SESOI (see Table 1). Despite this constraint that precluded inferential comparison of correlation differences, our study employed a thoughtful power analysis for single nonparametric correlations that accounted for random measurement error to estimate the effects of interest. This approach not only allowed for transparent and accurate reporting of statistical power for each preregistered statistical hypothesis but also enabled all our primary hypotheses to be statistically conclusive—by comparing each CI limit with the effects of interest. Although these practices are essential to evaluate the informational value of an empirical study, they are barely implemented (Lakens, 2022; Parsons et al., 2019).

Assuming a dimensional framework, our unsuccessful preregistered close replication of Coll-Martín et al. (2021) did not support Halperin and Schulz’s (2006) neurodevelopmental model of ADHD: Only one unpredicted correlation was significant. Neither were our exploratory findings in line with the developmental dissociation hypothesized by the model: Both arousal and executive task indices uniquely predicted ADHD symptom severity in childhood and its development into adulthood. On this basis, ADHD symptoms seem to share underlying neurocognitive alterations across the lifespan, at least in terms of vigilance. Beyond going against Halperin and Schulz’s model, this idea challenges the notion of late-onset ADHD as a distinct condition with less or no neuropsychological impairment (Moffitt et al., 2015). Consistent with this, longitudinal studies found no different alterations between child- and late-onset ADHD in a variety of neurocognitive functions (Cooper et al., 2018; Ilbegi et al., 2021; Riglin et al., 2022; but see Moffit et al., 2015, for initial positive findings).

The failure to observe distinct or lower neurocognitive impairments associated with late-developing ADHD symptoms could be interpreted as late-onset ADHD being the same neurodevelopmental disorder as child-onset ADHD, emerging later due to moderating factors (e.g., IQ, supportive family) delaying its manifestation (Faraone & Biederman, 2016; Kosaka et al., 2019). Alternatively, ADHD could be conceptualized as a non-neurodevelopmental, general mental health disorder (Rohde, 2023), with either early (e.g., risky genes or prenatal environments) or late (e.g., impaired brain development or stress) etiological factors converging into identical neurocognitive alterations leading to the emergence of ADHD symptoms at any age.

In either scenario, the lack of support for the neurodevelopmental model or any other account that predicts a distinct neurocognitive profile in individuals with late-onset ADHD entails implications for clinical practice. In this sense, translational interventions for ADHD symptoms (e.g., cognitive training, neurofeedback, brain stimulation) should aim to target the same underlying neurocognitive alterations regardless of individuals’ age and time of symptom onset. This is consistent with Asherson and Agnew-Blais’ (2019) recommendation to manage adult ADHD as usual, irrespective of the reported age of onset. Future studies need to include complementary assessment methods of ADHD symptoms, clinical groups, and other neurocognitive domains to strengthen and extend our tentative recommendations.

Contributed to conception and design: TC-M, HC-D, JL

Contributed to acquisition of data: TC-M, JL

Contributed to analysis and interpretation of data: TC-M, HC-D, JL

Drafted and/or revised the article: TC-M, HC-D, JL

Approved the submitted version for publication: TC-M, HC-D, JL

We thank Rocío Martínez Caballero for her generous availability and invaluable support with the development of the online version of the task.

Our work was supported by a predoctoral fellowship (FPU17/06169) awarded to TC-M from the Spanish Ministry of Education, Culture, and Sport; a research project grant (PID2019-104239GB-I00) from the Spanish Ministry of Science, Innovation and University (State Research Agency/10.13039/501100011033) awarded to HC-D; and research project grants by Spanish Ministry of Economy, Industry and Competitiveness, MCIN/ AEI /10.13039/501100011033 (PID2020.114790GB.I00), and PY20_00693, funded by the Consejería de Universidad, Investigación e Innovación, Junta de Andalucía, and by FEDER A way of doing Europe, awarded to JL. The funders had no role in any stage of the development and publication of this work. This paper is part of the doctoral dissertation of the first author under the supervision of the last two authors.

We have no potential conflict of interests to disclose.

The design and analysis plan of this study were preregistered in the Open Science Framework on July 10, 2022 using the Preregistration for Quantitative Research in Psychology Template (PRP-QUANT; Bosnjak et al., 2022): https://osf.io/tkdq7. Data and material, including R scripts, are publicly available at https://osf.io/vqgms. The online cognitive task employed in this study (ANTI-Vea) is freely available in multiple languages at https://anti-vea.ugr.es.

Correspondence concerning this article should be addressed to Tao Coll Martín, Departamento de Metodología de las Ciencias del Comportamiento, Facultad de Psicología, Universidad de Granada, Campus Universitario de Cartuja s/n, 18071 Granada, Spain. Email: [email protected]