Attention deficit hyperactivity disorder (ADHD) is a neurodevelopment disorder affecting 5%–10% of children and 2.5%–5% of adults (
1). Stimulant medications, particularly methylphenidate, are the first-line recommended treatment for ADHD and have proven to be effective in reducing ADHD core symptoms in clinical trials and meta-analyses. Given lower effect sizes, nonstimulant medications (e.g., atomoxetine) are second-line treatment and are primarily prescribed to individuals with poor stimulant treatment response or tolerance (
2–
4). Despite the high efficacy of stimulants, many patients discontinue treatment or switch to nonstimulant ADHD medications, with the most common reasons reported being poor treatment response and adverse effects (
5–
7). Because stimulant treatment has been associated with positive effects on important functional outcomes (
8), it is of clinical importance to identify genetic, clinical, and sociodemographic factors that influence stimulant treatment initiation, discontinuation, and switch to nonstimulant medications in ADHD.
Genetics likely contribute to stimulant treatment response and risk of adverse treatment effects, yet few genetic variants have been robustly linked to stimulant treatment outcomes in ADHD. A meta-analysis of candidate gene studies reported replicated associations of methylphenidate efficacy with variants in the
ADRA2A,
COMT,
SLC6A2,
SLC6A3, and
DRD4 genes (
9). However, candidate gene studies are known to be problematic, with many identified variants failing to replicate in genome-wide association studies (GWASs) (
10,
11). Two small GWASs (N<200) have been conducted on methylphenidate response (
12,
13), without any genome-wide significant hits identified, likely because of the limited sample sizes. Utilizing genetic data linked to individual-level electronic health records (EHRs), including prescriptions, is a promising avenue to obtain larger, representative samples for pharmacogenomic research. However, treatment outcomes are rarely reported in EHRs and must be approximated from prescriptions (
14,
15). Numerous pharmacoepidemiological studies have used discontinuation and switch to nonstimulants, defined from prescriptions, as proxies of suboptimal treatment outcomes in ADHD (
16–
23). Nevertheless, we are not aware of any studies investigating the genetic contributions to these stimulant treatment outcomes, through GWAS, single-nucleotide polymorphism (SNP) heritability (h
2SNP), or polygenic risk score (PRS) analyses.
PRSs, which capture the weighted sum of an individual’s phenotype-associated risk alleles, may be a useful component in predicting treatment outcomes (
14,
15). A recent study of 214 ADHD patients found that higher ADHD PRS was associated with symptom improvement following stimulant treatment (
24). Moreover, ADHD is genetically correlated with other psychiatric disorders, including autism spectrum disorder (ASD), depression, bipolar disorder, and schizophrenia (
11,
25). Thus, it can be hypothesized that higher PRSs for these disorders in individuals with ADHD may also influence stimulant treatment outcomes, for example, through increased risk of adverse treatment effects. Nevertheless, given the low predictive ability of current psychiatric PRSs, identification of patients at high risk of suboptimal stimulant treatment outcomes will likely depend on combining genetic, clinical, and sociodemographic information (
14). Psychiatric comorbidities can affect stimulant treatment outcomes in ADHD (
2,
3); for example, comorbid ASD has been linked to more adverse treatment effects (
26), and substance misuse to lower stimulant efficacy (
27). There are also concerns that stimulants may induce or exacerbate tics, compulsive behaviors, anxiety, depression, or psychotic symptoms (
3,
28,
29). Hence, individuals with ADHD and these comorbidities may be less likely to receive stimulants and more likely to discontinue or switch treatment. Age, sex, parental psychiatric illness, and socioeconomic status have also been linked to ADHD treatment initiation, discontinuation, and switch, albeit with mixed findings (
5,
7,
23,
30–
32).
Results
Among 9,133 individuals with ADHD, 29% were female, and the median age at first ADHD diagnosis was 12 years (interquartile range, 8–17). Baseline descriptive statistics are presented in Table S2 in the online supplement. Within 2 years of ADHD diagnosis, 7,427 individuals (81%) had initiated stimulant treatment. Among stimulant initiators, 3,370 (45%) had discontinued stimulants and 1,137 (15%) switched to nonstimulants within 2 years of stimulant initiation.
Associations With PRSs and Clinical and Sociodemographic Factors
PRS associations expressed by standard deviations (
Table 1) showed that ADHD and ASD PRSs were not associated with any stimulant treatment outcome. PRSs for bipolar disorder (hazard ratio=1.05, 95% CI=1.02, 1.09) and schizophrenia (hazard ratio=1.07, 95% CI=1.03, 1.11) were associated with discontinuation. PRSs for depression (hazard ratio=1.06, 95% CI=0.99, 1.13), bipolar disorder (hazard ratio=1.05, 95% CI=0.99, 1.12), and schizophrenia (hazard ratio=1.07, 95% CI=1.00, 1.13) were marginally but not significantly associated with switch to nonstimulants. PRS associations in the fully adjusted multivariate Cox models (see Table S3 in the
online supplement) were nearly identical to the main results.
PRS associations across quintiles (
Figure 2; see also Table S4 in the
online supplement) showed that individuals with ADHD in the highest PRS quintile for bipolar disorder (hazard ratio=1.21, 95% CI=1.09, 1.35) and schizophrenia (hazard ratio=1.24, 95% CI=1.11, 1.39) had higher rates of discontinuation compared with those in the lowest quintile. Further, those in the highest PRS quintile for depression (hazard ratio=1.26, 95% CI=1.05, 1.52) and bipolar disorder (hazard ratio=1.25, 95% CI=1.04, 1.51) had higher rates of switch. For schizophrenia PRS, rates were elevated across the second to the fifth quintiles but only significantly so at the fourth quintile (hazard ratio=1.33, 95% CI=1.10, 1.61). Confidence intervals for the remaining estimates were near to or included 1, and overlapped across quintiles.
Associations with clinical and sociodemographic factors are presented in
Table 1 and
Figure 3. Sex was not associated with initiation or discontinuation; however, females had higher rates of switch (hazard ratio=1.17, 95% CI=1.03, 1.33). Being diagnosed with ADHD at age 13 or older was associated with higher rates of initiation (hazard ratio=1.27, 95% CI=1.17, 1.38), discontinuation (hazard ratio=2.01, 95% CI 1.77–2.27), and switch (hazard ratio=1.91, 95% CI=1.53, 2.39) compared with diagnosis before age 13. Low maternal education was marginally associated with lower rates of initiation (hazard ratio=0.94, 95% CI=0.90, 0.99) and low paternal income with higher rates of discontinuation (hazard ratio=1.11, 95% CI=1.03, 1.20).
Comorbid OCD, anxiety, bipolar disorder, and substance use disorder were associated with lower rates of initiation (hazard ratio range, 0.59–0.88). Further, ASD, anxiety, bipolar disorder, and substance use disorder were associated with higher rates of discontinuation (hazard ratio range, 1.18–1.46), and oppositional defiant/conduct disorder, tics, anxiety, and substance use disorder with higher rates of switch (hazard ratio range, 1.41–2.08), compared with individuals with ADHD without these comorbidities. Estimates stratified by age at first ADHD diagnosis (see Figure S2 and Table S5 in the online supplement) showed that comorbid ASD was associated with discontinuation only in children (hazard ratio=1.41, 95% CI=1.24, 1.60). Similarly, comorbid depression was associated with discontinuation (hazard ratio=3.05, 95% CI=1.76, 5.28) and switch (hazard ratio=3.43, 95% CI=1.53, 7.71) in children but not in adolescents and adults. Associations of comorbid bipolar disorder and substance use disorder with stimulant treatment outcomes in the main analysis were driven by individuals diagnosed with ADHD during adolescence or adulthood, as there were too few comorbid cases (N<10) to estimate hazard ratios in children. Confidence intervals were overlapping for the remaining age-stratified estimates.
Supplementary analyses (see Table S6 in the online supplement) showed that individuals with ADHD who initiated treatment with a nonstimulant medication (N=568) were more likely to be female (odds ratio=0.77, 95% CI=0.64, 0.93 [male as reference]), to be diagnosed with ADHD at age 13 or older (odds ratio=6.12, 95% CI=4.20, 9.00), and to have comorbid ASD, tics, anxiety, or substance use disorder (odds ratio range, 1.47–3.56) compared with stimulant initiators.
GWAS and h2SNP
h
2SNP on the observed scale was estimated to be 0.08 (SE=0.06) for initiation, 0.13 (SE=0.08) for discontinuation, and 0.09 (SE=0.11) for switch. N and liability-scale converted estimates are provided in Table S7 in the
online supplement. No genome-wide significant hits were detected for initiation and discontinuation (p>5 × 10
−8) (see Figures S3 and S4 in the
online supplement). For switch, one locus on chromosome 16q23.3 (p<4.7 × 10
−8, lead SNP rs58543609, CHR16:82376003 [GRCh37]) reached genome-wide significance (
Figure 4). Using FUMA (see supplementary note 3 in the
online supplement), AC024590.1 and RN7SKP190 were identified as the most proximal genes of the rs58543609 locus. Expression quantitative trait loci annotation identified one putatively associated gene,
MPHOSPH6. A lookup in the GWAS Catalog (
47) identified no prior reports of the lead SNP, although a locus scan (CHR16:82315555–82397332) revealed suggestive associations in prior GWASs of cognitive performance, seasonal depression, face morphology, and bone mineral density. PheWAS implemented in the GWAS ATLAS (
48) revealed two putative associations with a measure of weekly alcohol intake and cerebellar volume. Associations for the 18 independent genomic loci reaching suggestive genome-wide significance (p<10
−5) and FUMA results are presented in Tables S8 and S9 in the
online supplement.
Discussion
This is, to our knowledge, to first study thus far utilizing genetic and prescription data to investigate stimulant treatment initiation, discontinuation, and switch to nonstimulant medication in ADHD. We present novel findings suggesting that individuals with ADHD and higher polygenic liability for bipolar disorder, schizophrenia, and possibly depression may be at increased risk of stimulant treatment discontinuation and switch. We also identified several clinical factors contributing to stimulant treatment outcomes, including delayed ADHD diagnosis and certain psychiatric comorbidities. We also present the first GWASs and h2SNP estimates of stimulant treatment outcomes in ADHD defined from prescription data, identifying one putative locus associated with switch to nonstimulants.
We found that a majority (81%) of individuals with ADHD initiated stimulant treatment, yet within 2 years, nearly half (45%) had discontinued treatment and 15% had switched to nonstimulants. Similar rates of discontinuation have previously been reported in Denmark, Sweden, the United States, South Korea, and Taiwan (
19–
23), highlighting that prescription databases provide relatively consistent estimates of ADHD treatment patterns and that discontinuation of medication is an important issue in the management of ADHD globally.
Individuals with ADHD in the higher PRS quintiles for bipolar disorder and schizophrenia had 17%–25% higher rates of discontinuation and switch compared with those in the lowest PRS quintiles. We also found that higher polygenic liability for depression may increase the rate of switch. These novel findings require replication, as some confidence intervals included 1, and the pattern of associations was not entirely consistent with a dose-response relationship across all PRS quintiles. Nevertheless, it can be hypothesized that elevated genetic liability for mood and psychotic disorders in individuals with ADHD may increase the risk of adverse effects of stimulant treatment (e.g., mood destabilization or psychotic symptoms) (
3,
29). In support of this, we also found that comorbid depression in children with ADHD, and comorbid bipolar disorder in adolescents and adults, was associated with higher rates of discontinuation. Another possibility is that prodromal mood or psychotic symptoms may, in some cases, be misdiagnosed as ADHD, potentially rendering stimulant treatment inappropriate (
49). Nevertheless, given the high validity of ADHD diagnoses in the Danish registers (
50) and evidence of familial and genetic overlap between ADHD and mood and psychotic disorders (
11,
51), this is unlikely to fully explain our results. Regardless, these findings underscore the importance of screening for family history and symptoms of psychosis, mania, and depression prior to initiating stimulants and suggest that PRSs may complement such screening in the future, assuming that their predictive validity can be substantially improved (
14). This is especially notable as we found that PRS associations, although modest in effect size compared with psychiatric comorbidities, were largely unchanged after adjusting for comorbidities and sociodemographic factors. This suggests that higher PRSs for mood and psychotic disorders in ADHD may act as an independent risk factor for stimulant treatment discontinuation and switch, which is measurable prior to the (potential) onset of such disorders and in patients with symptoms below the diagnostic threshold.
We found limited evidence for the contribution of ADHD and ASD PRSs to stimulant treatment outcomes, possibly because of the lower power in the discovery GWASs of these disorders or the challenge of using PRS for ADHD to predict secondary associations in a sample selected for ADHD (i.e., “index event bias”) (
52). It is also possible that risk variants for ADHD are not necessarily the same variants influencing treatment outcomes in ADHD. Interestingly, a recent study (
53) found no correlation between 23 genes identified as targets of ADHD medications and ADHD GWAS summary statistics. Instead, associations were observed for GWAS of alcohol consumption, schizophrenia, and depression with the
DRD2,
CYP2D6, and
CHRM2 genes. This, together with our findings, suggests that the effectiveness and/or tolerance of ADHD medications may be mediated by pathways other than those underlying ADHD, and that genetic liability for psychiatric disorders emerging in adolescence and adulthood may be of importance for stimulant treatment outcomes (
53).
Beyond bipolar disorder and depression (discussed above), several psychiatric comorbidities stood out as risk factors for discontinuation and switch. Comorbid OCD and anxiety increased the risk of discontinuation and switch, and ASD increased the risk of discontinuation, although the latter was only significant in children. Although evidence is low, these comorbidities have historically been suggested to increase the risk of adverse effects of stimulant treatment (
2,
3). Our data may either support the comorbidities’ association with adverse effects or merely confirm that this clinical perception is common. Among individuals with ADHD diagnosed in adolescence or adulthood, substance use disorder increased the risk of both discontinuation and switch, while comorbid oppositional defiant or conduct disorder was associated only with switch. The former may relate to evidence of lower stimulant efficacy in ADHD with comorbid substance misuse and clinical concerns of misuse or diversion (
27). Finally, comorbid tic disorder increased the risk of switch to nonstimulants across ages, and the risk of discontinuation in children, which could reflect the fact that stimulant treatment can, although rarely, exacerbate or induce tics (
2,
3). Our findings emphasize the need for a broad and age-sensitive clinical assessment in ADHD prior to treatment initiation, as well as close monitoring for psychiatric problems emerging during treatment, as comorbidities may negatively influence stimulant tolerance and efficacy. Older age at first ADHD diagnosis was also associated with elevated risk of discontinuation and switch, and treatment initiation with nonstimulants was more common in those diagnosed with ADHD after childhood. Similar findings have been reported in previous studies (
7,
16,
20,
23), emphasizing that adolescence and early adulthood are high-risk periods for ADHD treatment dropout. Our results suggest that efforts are needed to improve treatment continuity in these age groups.
Finally, we evaluated the feasibility of using EHRs in ADHD pharmacogenomics. GWASs of stimulant initiation and discontinuation did not identify any genome-wide significant loci, and the common variant contribution estimated by h
2SNP was not significantly different from zero for any stimulant treatment outcomes. These null findings are unsurprising given our limited sample size and the challenge of defining treatment outcomes that are more proximal to the underlying genetics (
54,
55). Nevertheless, it can be noted that despite large standard errors, our h
2SNP point estimates are in line with heritabilities reported for other treatment response traits in psychiatry (e.g., antidepressant treatment response [
56]). Moreover, (potentially) low h
2SNP does not imply that genetic factors are unimportant, as the power to detect genetic variants for a given phenotype depends both on its genomic architecture and GWAS sample size (
57). We did identify one genome-wide significant locus associated with switch to nonstimulants, on chromosome 16q23.3. The locus and most proximal gene (AC024590.1) have not been identified in prior GWASs, but the RN7SKP190 and cadherin-13 (CDH13) gene are located within 250 kb of the locus and have been associated with theoretically relevant traits, such as body mass index, blood pressure, and educational attainment (
47). However, the potential molecular consequences of the rs58543609 locus for switch require further investigation. Based on the modest h
2SNP point estimates and GWAS without any major gene effects, our results suggest that stimulant treatment outcomes (as defined here) likely have a complex genetic architecture and that larger samples will be needed for gene discovery. Given the continuous integration of genetics in large EHR databases globally, and rapid methods development for approximating treatment outcomes (
15), we believe our findings illustrate the potential of using prescription data to advance pharmacogenomics in ADHD.
Strengths and Limitations
Our study has several strengths, including the representativeness of the iPSYCH2012 sample and access to longitudinal register data. There are also important limitations. First, pharmacoepidemiological studies rely on modeling treatment outcome proxies from prescriptions, and we do not know how well these proxies map onto actual medication consumption. Further, the reasons for discontinuation or switch are not recorded in the Danish National Prescription Registry. Second, our study was underpowered for GWASs and h
2SNP estimation, limiting our ability to draw firm conclusions regarding the common variant contribution to stimulant treatment outcomes. Third, iPSYCH2012 only includes ADHD cases of the combined subtype (F90.0), meaning that the results may not apply in the broader ADHD case group. Fourth, we did not have data on several factors shown to be important for ADHD treatment adherence, such as perceived stigma and patients’ and parents’ attitudes toward treatment (
5,
7). Finally, the reliance on a highly homogeneous Danish population sample may limit the findings’ generalizability in more diverse populations.