Conclusions Drug treatment for ADHD was associated with beneficial effects in reducing the risks of suicidal behaviours, substance misuse, transport accidents, and criminality but not accidental injuries when considering first event rate. The risk reductions were more pronounced for recurrent events, with reduced rates for all five outcomes. This target trial emulation study using national register data provides evidence that is representative of patients in routine clinical settings.
Results Of 148 581 individuals with ADHD (median age 17.4 years; 41.3% female), 84 282 (56.7%) started drug treatment for ADHD, with methylphenidate being the most commonly prescribed at initiation (74 515; 88.4%). Drug treatment for ADHD was associated with reduced rates of the first occurrence of suicidal behaviours (weighted incidence rates 14.5 per 1000 person years in the initiation group versus 16.9 in the non-initiation group; adjusted incidence rate ratio 0.83, 95% confidence interval 0.78 to 0.88), substance misuse (58.7 v 69.1 per 1000 person years; 0.85, 0.83 to 0.87), transport accidents (24.0 v 27.5 per 1000 person years; 0.88, 0.82 to 0.94), and criminality (65.1 v 76.1 per 1000 person years; 0.87, 0.83 to 0.90), whereas the reduction was not statistically significant for accidental injuries (88.5 v 90.1 per 1000 person years; incidence rate ratio 0.98, 0.96 to 1.01). The reduced rates were more pronounced among individuals with previous events, with incidence rate ratios ranging from 0.79 (0.72 to 0.86) for suicidal behaviours to 0.97 (0.93 to 1.00) for accidental injuries. For recurrent events, drug treatment for ADHD was significantly associated with reduced rates of all five outcomes, with incidence rate ratios of 0.85 (0.77 to 0.93) for suicidal behaviours, 0.75 (0.72 to 0.78) for substance misuse, 0.96 (0.92 to 0.99) for accidental injuries, 0.84 (0.76 to 0.91) for transport accidents, and 0.75 (0.71 to 0.79) for criminality.
To overcome these limitations, this study for the first time applied the target trial emulation framework to examine the effects of drug treatment for ADHD on five critical outcomes—suicidal behaviours, substance misuse, accidental injuries, transport accidents, and criminality. This approach enhances causal inference by mimicking the design principles of a randomised controlled trial within an observational context and provides estimates of treatment effects for the entire ADHD population from routine practice. Leveraging Swedish national registers, we examined both first and recurrent events, reflecting the recurrent nature of these outcomes. The selection of outcomes was made in consultation with people with lived experience, aligning with the practical needs of those affected by ADHD.
Randomised controlled trials have shown the beneficial effects of drug treatment for ADHD in alleviating core symptoms. 8 However, evidence from randomised controlled trials remains limited or inconclusive for broader and important clinical outcomes such as suicidal behaviours and substance use disorder. 9 10 11 12 Moreover, randomised controlled trials often exclude a substantial population of patients seen in clinical practice—around half of those receiving drugs for ADHD, 13 thereby limiting the generalisability to the entire ADHD population. In this context, pharmacoepidemiological studies using routinely collected data offer opportunities to assess the benefits and risks of ADHD drug treatment on broader outcomes. 14 15 In particular, studies using within individual designs have linked use of drugs for ADHD to reduced risks of suicidal behaviours, 16 17 18 substance misuse, 19 20 accidental injuries, 21 transport accidents, 22 23 and criminality. 24 Although effectively controlled for time invariant confounders, these studies remain susceptible to time varying confounding and carryover effects, 25 and their reliance on treated patients who have experienced the outcomes of interest limits both the generalisability and comparability to trial findings. Thus, rigorous population based studies using routine clinical data, designed to ensure representativeness and comparability to trials, are needed.
Attention deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder, affecting approximately 5% of children and 2.5% of adults worldwide. 1 2 3 Although typically diagnosed in childhood, its impairing symptoms often persist into adulthood. 4 Beyond core symptoms, ADHD is linked to a range of adverse functional outcomes, including increased risks of suicidal behaviours, substance misuse, accidental injuries, transport accidents, and criminality. 1 5 Treatment for ADHD includes drug, non-drug, and combined approaches. Although non-drug treatment is often recommended for younger children or milder cases, drug treatment (including stimulants and non-stimulants) is commonly used in the management of school aged and older individuals with ADHD. Prescriptions of drugs for ADHD have risen markedly in recent years worldwide, sparking intense debate on their effectiveness and safety. 6 7
As this is a register based study, we had no direct contact with patients or participants at any stage. However, public discourse, media coverage, and interactions with individuals affected by ADHD show that many patients and care givers lack awareness of the risks and benefits of ADHD drug treatment, leading to uncertainty in treatment decisions. This knowledge gap served as a key motivation for our research. We discussed the aim and design of this study with representatives of people with lived experience of ADHD from ADHD Europe, the largest association of people with lived experience of ADHD in Europe. The board of ADHD Europe noted the importance of this research and the need for evidence from routine clinical settings. Their feedback guided the selection of outcomes and informed the interpretation of the findings.
We did subgroup analyses based on sex, age (children and youths (<25 years), adults (≥25 years)), and people with and without a history of events. To test the robustness of our findings, we further did the following sensitivity analyses. Firstly, we extended the grace period to six months after diagnosis to account for potential variations in clinical practice and patients’ adherence. Secondly, we allowed drug switches during follow-up by not censoring individuals who switched between ADHD drugs. This approach enabled us to estimate the causal contrast between starting drug treatment for ADHD within three months after diagnosis and sustaining any ADHD drug treatment (that is, allow switching between ADHD drugs) versus not starting drug treatment for ADHD during the follow-up. Thirdly, we applied negative outcome control to assess potential biases and residual confounding. 50 We used type 1 diabetes as a negative outcome given that previous studies did not find any significant effect of ADHD drug treatment on glycaemic management for type 1 diabetes. 51
In secondary analyses, we examined the association between drug treatment for ADHD and recurrent events of the five outcomes. To minimise misclassification of recurring treatment visits as outcome events, we allowed a maximum of one event a month. In these recurrent event analyses, follow-up was not censored at the occurrence of the outcomes, allowing us to study the rates of recurrent events over time while otherwise applying the same criteria for determining the end of follow-up as in the main analysis. To compare the effects of stimulant and non-stimulant drugs, we emulated a head-to-head trial comparing the effects of starting stimulants (methylphenidate, amphetamine, dexamphetamine, and lisdexamfetamine) with starting non-stimulants (atomoxetine and guanfacine) on the outcomes of interest. Follow-up began at the initiation of ADHD drug treatment and ended according to the same criteria used in the main analysis. We used SAS 9.4 and R version 4.4.0 for all analyses and defined statistical significance as a two tailed P value of ≤0.05.
To assess covariate balance at the end of the grace period (three months after ADHD diagnosis), we calculated standardised mean differences, with a difference <0.10 indicating sufficient balance. 48 We fitted separate models for the five outcomes of interest by using weighted pooled logistic regression, regressing the outcome on treatment and time, which approximates the incidence rate ratio. 49 We applied non-parametric bootstrapping with 500 full re-samples of individuals from the cohort to calculate the 95% confidence intervals.
The two treatment strategies considered in our main analysis were starting drug treatment for ADHD within three months of diagnosis and remaining on the prescribed therapy (initiation group) versus not starting drug treatment for ADHD during the follow-up (non-initiation group). To estimate the average treatment effect of sustained ADHD drug treatment on five outcomes over the two year period for the entire study population, we applied a three step approach—cloning, censoring, and inverse probability weighting—designed to emulate the key features of randomised controlled trials and eliminate immortal time bias (supplementary figure B). 44 45 Firstly, in the cloning step, we created a dataset with two identical copies (clones) of each eligible individual at baseline. One clone was assigned to the treatment strategy of starting ADHD drug treatment within three months of diagnosis and remaining on treatment, and the other one was assigned to the strategy of not starting ADHD drug treatment during the follow-up. This step ensured alignment of treatment assignment with the start of follow-up and eliminated baseline confounding. 45 46 Secondly, in the censoring step, we assessed whether each clone adhered to the assigned treatment strategy at monthly intervals and censored them when they deviated from the assigned treatment strategy. Clones in the initiation group were censored if they had not started treatment by the end of grace period or discontinued/switched drug treatment after the grace period. Clones in the non-initiation group were censored on receipt of any ADHD drug treatment. Thirdly, in the weighting step, we applied pooled logistic regression models to calculate time varying inverse probability of censoring weights. These models included time and all time fixed and time varying covariates described above, to account for potential selection bias induced by the artificial censoring in the second step. 47 Weights were truncated at the 99.5th centile to reduce the influence of extreme values (see supplementary methods for details).
... continue reading