Attention-deficit/hyperactivity disorder (ADHD) is characterised by inattention and/or hyperactivity-impulsivity levels that interfere with functioning. Symptoms have been associated with a range of impairments including higher levels of substance use, criminality, co-occurring mental health issues, and social difficulties (e.g., [9]. Research has highlighted considerable heterogeneity in the lifespan course of ADHD symptoms [6]. Accordingly, researchers have begun to organise this variation into developmental ‘subtypes’ of ADHD, such as ‘early-onset persisting’, ‘early-onset remitting’ and ‘late-onset’, reflecting the primary ways in which symptoms are assumed to present over time [6]. However, less is known about whether such groups can be differentiated on the basis of clinically meaningful outcomes, which may suggest a benefit of diagnostic specifiers for ‘developmental subtype’. For example, demonstrating that some ADHD symptom trajectories are associated with greater or different patterns of impairment compared to others (e.g., higher rates of criminality, co-occurring conditions etc.), would suggest the need for tailored intervention strategies that target the particular needs of each trajectory group.

Analysing outcomes of different trajectories is also important for establishing whether there are impairments (relative to those who never show elevated ADHD symptoms) that outlast clinically significant symptoms for those who remit, implying that this group may require continued support beyond symptom remission. Likewise, given the application of age of onset cut-offs for ADHD, it is important to establish whether those who do not meet age cut-offs may nevertheless show impairments and benefit from intervention, despite not showing a ‘classical’ ADHD symptom trajectory. The extent to which this trajectory predicts outcomes associated with ADHD can inform current debates about the clinical validity of a ‘late onset’ category [12].

ADHD symptom trajectories and their outcomes have been studied predominantly through the a-priori classifications. For example, studies have classified symptoms as persisting or remitting depending on whether symptoms are clinically significant at both an early and later age (persisting) or at just an earlier age (remitting) and have tended to find poorer outcomes for symptom-persisting compared to symptom-remitting individuals (e.g., [2]. Studies have also compared early- versus late-onset ADHD based on whether symptoms first appear after versus before the age of onset in diagnostic criteria for ADHD (i.e., after age 12, previously 7). These studies have yielded somewhat mixed findings, however, most have found comparable levels of co-occurring mental health issues, delinquency, social difficulties, and tobacco, alcohol, and illicit drug use/misuse among early and late-onset subtypes (e.g., [1, 7, 10, 15]. However, these previous studies have not modelled the full variation that exists in ADHD symptom trajectories. For example, defining late-onset as age 12 or above (as is encoded in clinical diagnostic criteria) may be considered arbitrary because the emergence of clinically relevant symptoms can occur across a wide range of ages, possibly up to and including adulthood (e.g., [3]. Similarly, rather than there being a single point in time at which symptoms remit/appear, evidence points to continual fluctuations in symptoms over time for many [39].

To better reflect individual and developmental variation and thus better detect differences in outcomes, longitudinal studies drawing on data-driven techniques such as latent class growth analysis or growth mixture modelling can be employed. Such methods model linear and non-linear changes in symptoms over a developmental period to identify trajectory groups that optimally reflect patterns of symptom variation in a particular sample (e.g., [16, 18, 19, 21,22,23,24,25,26,27, 31, 32, 36, 41]. Whilst varying in terms of their samples, measurement methods, and developmental periods covered, certain commonalities have surfaced across the findings of such studies. Using these approaches, trajectories that could be mapped approximately to the early-onset persisting, early-onset remitting, and late-onset groups that are typically specified in studies using a-priori classification often emerge, however, with a more detailed picture of how symptoms develop over time.

ADHD symptom trajectories emerging from trajectory analyses can also be compared with respect to various outcome variables. Sasser et al. [36] used trajectory analysis with parent-reported ADHD symptom data across ages 8–18 and found that three trajectory groups emerged, labelled ‘low’ (consistently low symptom levels across time), ‘declining’ (symptoms that remitted over time) and ‘high’ (symptoms that persisted over time). Those in the high trajectory group had elevated rates of parent-reported antisocial behaviour and school dropout, but similar levels of unemployment and juvenile arrests compared to the declining group. Tandon et al. [41] also identified ‘low’, ‘remitting’ and ‘high’ trajectory groups when analysing ADHD symptoms across ages 9–21. On a range of psychiatric disorders (including major depressive and oppositional defiant disorder), the high group was found to have the most co-occurring issues, followed by the declining, and then the low group. However, rates of other disorders such as generalised anxiety disorder, alcohol and cannabis use disorder did not differ between any of the three groups. Murray et al. [18] found that their ‘late-onset’ group (characterised by rising symptoms across ages 7–15) and ‘persistent’ group (persistently high symptoms across ages 7–15) were similar to each other across most outcomes including comparably high rates of delinquency, internalising problems, violent ideations, and cigarette smoking in comparison to ‘unaffected’ individuals. However, consistent with some of the aforementioned a-priori studies, they also noted some poorer outcomes for early-onset individuals such as higher levels of reactive aggression compared to late-onset individuals, thus leading authors to conclude that late-onset may represent a milder, though still impaired, subtype of ADHD. This study provided some initial evidence on how ADHD symptom trajectories link to outcomes but the sample was considerably smaller than some other datasets that have relevant symptom trajectory data and though community-ascertained, was not nationally representative.

Trajectory analysis studies to date have thus started to suggest the possibility of differential impairments between groups with different developmental trajectories of ADHD symptoms. However, with such studies remaining relatively scarce and with the importance of such work for informing diagnostic and treatment procedures, further work is necessary to ensure the replicability of these preliminary findings. As such, in a large UK-representative sample, we sought to investigate different developmental trajectories of ADHD. In this, we build on a previous trajectory analysis study examining trajectories in the same sample [21]. However, in that study trajectories were only estimated up to the age of 14 based on data availability. As adolescence is a time of rapid and marked change, with specific implications for ADHD-related traits such as sensation-seeking and self-regulation [22,23,24,25,26,27, 37], it is important to build on these earlier analyses to examine how trajectories extend up to the end of middle adolescence. For example, it is unclear if ‘adolescent-onset’ trajectories might represent temporary versus sustained increases in symptoms, whether remitting trajectories tend to show an accelerating, decelerating, or stabilising trajectory towards the end of middle adolescence,or whether there may be entirely new trajectory classes emerging (e.g., with an onset later in adolescence) when a more extended developmental period is considered. Further, whilst these earlier analyses sought to validate and provide a broader characterisation of trajectory-based distinctions by examining early-life predictors and possible etiological of correlates of trajectory analysis membership, they did not examine outcomes of following particular trajectories. This is arguably of greater immediate clinical relevance than identifying early-life predictors of class membership as they can inform the provision of tailored preventive interventions and support to mitigate anticipated challenges.

In this study we, therefore, extend earlier analyses to examine developmental trajectories, now using data from ages 3–17 and also examine the links between these trajectories and outcomes at age 17. We examine whether age 3–17 trajectories are associated with differing levels of impairment on multiple outcome variables that have previously been associated with ADHD symptoms, assessed when participants were aged 17. Outcomes included substance use, dimensions of mental health (self-esteem, psychological distress, and well-being), peer victimisation, and delinquency. Though specific hypotheses are difficult to define prior to selecting a trajectory model, based on trends from previous research, we hypothesised that the pattern across outcome variables would be: (1) those with persistently high ADHD symptom levels will have the most impairment compared to relevant other groups whilst unaffected individuals will have the least impairment, (2) those with a late-onset of symptoms will have fewer impairments than those with an early-onset persistent but more impairments than unaffected individuals, and (3) those whose symptoms remit will have fewer impairments than those whose symptoms persist but more impairments than unaffected individuals.

Participants (N = 10,262) were from the Millennium Cohort Study (MCS; [8] who provided data across sweeps 2–7 of the study (average age at each sweep was 3, 5, 7, 11, 14 and 17 years, with information about the age distributions at each sweep provided in Supplementary Materials Table S1). MCS has been tracking the development, family, and wider social lives of individuals born in the United Kingdom (UK) between 2000 and 2002. Participants were sampled using a stratified, clustered random sampling design in which individuals were clustered geographically and disproportionately stratified so as to over-sample residents of the three smaller countries of the UK (Scotland, Wales and Northern Ireland), disadvantaged areas and ethnic minorities. As such, sampling weights were used in all analyses to adjust for the effects of attrition and non-random sampling, thus ensuring results were UK-representative. For further details, the MCS is fully documented and freely accessible at: https://ukdataservice.ac.uk/. Written/verbal informed consent was obtained from all parents/participants where required. Ethical approval for the current secondary data analysis was granted by the University of Edinburgh School of Philosophy, Psychology and Language Sciences Ethics Committee.

ADHD symptoms were measured using the Strengths and Difficulties Questionnaire (SDQ; [11]. The SDQ is one of the most widely used and well-validated behavioural screening instruments for children and adolescents (Kersten et al., 2016). It has shown good psychometric properties in the current sample, including a high degree of gender and developmental invariance [25]. The hyperactivity/inattention subscale has shown high correlations with ADHD diagnosis [32]. It includes five items asking parents about their child’s behaviour during the last six months with reference to the following behaviours: ‘restless, overactive, cannot stay still for long’,‘constantly fidgeting or squirming’,‘easily distracted, concentration wanders’,‘thinks things out before acting’; and ‘sees tasks through to the end, good attention span’. At age 3, the item ‘thinks things out before acting’ was replaced with ‘can stop and think things out before acting’ to improve its age-appropriateness. Responses were recorded on a 3-point scale including not true (0), somewhat true (1), and certainly true (2). Positively worded items were reverse-coded, and responses were summed to produce an overall hyperactive/inattentive score with higher scores indicating greater hyperactivity/inattentiveness (possible range = 0–10). In a similar sample to the current study, Riglin et al. [32] found an optimal cut-off for identifying clinically significant symptoms (assessed via a DSM-IV diagnostic interview) to be a score of 7 or more on the SDQ subscale (specificity = 90%, sensitivity = 86%, area under the curve = 0.88), with 6 representing a borderline score.

All outcomes were assessed when participants were aged 17 via online self-report questionnaires and face-to-face interview. Measures are described below with further details in Supplemental Materials. Outcome variables were selected based on the availability of measures of concepts that have been associated with ADHD symptoms in previous research.

Peer victimisation was assessed as the number of victimisations the participant had experienced in the last 12 months. Six items (α = 0.70) measured whether participants had experienced a range of verbal, physical, emotional, and online abusive behaviours (e.g., ‘Has anyone called you names?’, ‘Has anyone been physically violent towards you?’). Participants responded ‘no’ (0) or ‘yes’ (1) to each item and responses were summed to create a total victimisation score for each participant, with higher scores reflecting a greater number of victimisation behaviours experienced (range = 0–6).

To assess alcohol consumption, participants were asked ‘How many times have you had an alcoholic drink in the last 12 months?’ with responses recorded on a 7-point scale (Never = 0, 1–2 times = 1, 3–5 times = 2, 6–9 times = 3, 10–19 times = 4, 20–39 times = 5, 40 or more times = 6). One item assessed participants’ cannabis use: ‘In the past year how many times have you taken cannabis?’ with responses recorded on a 5-point scale (Not taken in the last year = 0, 1–2 times = 1, 3–4 times = 2, 5–10 times = 3, More than 10 times = 4).

Anxiety and depressive symptoms over the last 30 days were assessed via the 6-item Kessler Psychological Distress Scale [13], α = 0.86). Responses were on a 5-point scale ranging from none of the time (1) to all of the time (5) and summed to create a total composite score (range = 6–30). Mental wellbeing over the past two weeks was measured via the 7-item Short Warwick-Edinburgh Mental Wellbeing Scale [40], α = 0.83). Responses were on a 5-point scale from none of the time (1) to all of the time (5). Positively worded items were reverse-coded, and responses summed to create a total composite score (range = 7–35). Global self-esteem was measured using five items from the Rosenberg Self-esteem Scale [33], α = 0.91). Responses were on a 4-point scale from strongly disagree (1) to strongly agree (4). Positively worded items were reverse-coded, and responses summed to create a total composite score (range = 5–20). Higher scores on each of the three scales indicated greater impairment.

Delinquency was measured via nine items (α = 0.65) assessing whether participants had engaged in various delinquent behaviours during the past 12 months including theft, vandalism, breaking and entering, arson and online hacking (e.g., ‘Have you taken something from a shop without paying for it?’, ‘Have you deliberately set fire to something that you shouldn’t have?’). Participants responded ‘no’ (0) or ‘yes’ (1) to each item and responses were summed to create a total delinquent score for each participant, with higher scores indicating greater delinquency (range = 0–9). This ‘variety index’ score method of measuring delinquency is recommended (as opposed to summing the frequency of individual behaviours) because it avoids scores being disproportionately influenced by non-serious but frequent delinquent acts.

Latent class growth analysis models were fit with increasing numbers of classes until a stopping point was reached, defined by a non-significant Lo-Mendell-Rubin (LMR) adjusted test. Akaike’s information criterion (AIC), Bayesian information criterion (BIC), and sample size adjusted BIC (saBIC) were used to help with model selection in cases where the adjusted LMR test yielded ambiguous results. It is known that no single class enumeration will consistently select the ‘correct’ number of classes and as such class enumeration indices are best used to inform the numbers of classes alongside substantive and pragmatic criteria such as the interpretability of classes [30, 43]. Moreover, when conceptualizing the latent class models as a means of providing a convenient but defensible discretization of an underlying continuous distribution, as we do in the current context, there is no ‘correct’ number of classes to detect, only an optimal number for summarising variation in a parsimonious manner [28].

Growth models with intercept, linear slope, and quadratic slope factors included were fit based on previous research suggesting that ADHD symptom trajectories tend to be curvilinear [18, 22]. Time was scaled by fixing the slope factor loadings proportional to the distance between waves with age 3 loadings fixed to 0 (baseline) and age 17 loadings fixed to 1. Factor variances and covariances were fixed to 0 within classes, implying that all trajectory variation is due to the underlying latent categorical variable. This operationalises the assumption of the latent classes as convenient summaries of an underlying continuous distribution rather than necessarily reflecting true typologies (see [29] for a discussion). This can be contrasted to a growth mixture modelling (GMM) approach which has been interpreted conceptualising groups in terms of subpopulations (see e.g., [29]. Given this conceptualisation, GMM thus allows variation around an average growth curve within each subpopulation. As a consequence of allowing this within-class variation, it typically models the same data using fewer sub-groups.

Following the selection of an optimal latent class growth analysis model, age 17 outcomes were compared across classes, using the three-step method described in Asparouhov and Muthén [4] to correct for classification uncertainty. This method can be vulnerable to changes in the nature of the classes with the inclusion of outcomes in the model,however, this is checked and flagged by the analysis programme when it arises. If this occurred for a given outcome we used the BCH method discussed in Asparouhov and Muthén [4]. The BCH method involves fitting a multi-group model treating the class membership as known and weighting observations based on weights derived from their classification probabilities. Most likely class membership and classification probabilities are obtained from the latent class growth analysis model estimated in the first step. All outcomes were treated as continuous as they had a minimum of five response options. All analyses were conducted in Mplus 8.4 using (robust) pseudo-maximum likelihood estimation that adjusts for the complex sampling design of MCS [27]. Missing data were dealt with using attrition weights provided by MCS. These up-weight respondents with a low probability of responding and down-weight those with a high probability to correct for non-random drop-out. This provides unbiased parameter estimates under a ‘missing at random’ (MAR) assumption in Rubin’s [34] terminology.