|Home | About | Journals | Submit | Contact Us | Français|
Antidepressant effects on increased suicidality in children have raised public concern in recent years. Approved in 2002 for attention-deficit/hyperactivity disorder treatment, the selective noradrenalin-reuptake-inhibitor atomoxetine was initially investigated for the treatment of depression. In post-hoc analyses of clinical trial data, atomoxetine has been associated with an increased risk of suicidal ideation in children and adolescents. We analyzed whether the observed increased risk of suicidal ideation in clinical trials translates into an increased risk of suicidal events in pediatric patients treated with atomoxetine compared with stimulants in 26 Medicaid programs.
Employing a retrospective cohort design, we used propensity score–adjusted Cox proportional hazard models to evaluate the risk of suicide and suicide attempt in pediatric patients initiating treatment with atomoxetine compared with stimulants from 2002 to 2006.
The first-line treatment cohort included 279315 patients. During the first year of follow-up, the adjusted hazard ratio for current atomoxetine use compared with current stimulant use was 0.95 (95% CI 0.47–1.92, P = .88). The second-line treatment cohort included 220215 patients. During the first year of follow-up, the adjusted hazard ratio for current atomoxetine use compared with current stimulant use was 0.71 (95% CI 0.30–1.67, P = .43).
First- and second-line treatment of youths age 5 to 18 with atomoxetine compared with stimulants was not significantly associated with an increased risk of suicidal events. The low incidence of suicide and suicide attempt resulted in wide confidence intervals and did not allow stratified analysis of high-risk groups or assessment of suicidal risk associated with long-term use of atomoxetine.
Antidepressant and atomoxetine effects on increased suicidality in children have raised public concern in recent years resulting in boxed warnings. However, this association is based on clinical trial data.
This study analyzed if the observed increased risk of suicidal ideation in clinical trials translates into an increased risk of suicidal events in youths aged 5 to 18 treated with atomoxetine compared with central nervous system stimulants in 26 Medicaid programs.
Concern about an association between antidepressant use and increased suicidality peaked with the results of a meta-analysis in 2003, which concluded nearly twice the rate of suicidality in antidepressant users compared with placebo.1 These findings resulted in boxed warnings in the United States and Europe, indicating an increased risk of suicidal thinking and behavior in children and adolescents treated with antidepressant medications.2,3
Atomoxetine was approved in 2002 as a novel mechanism of action, nonstimulant and noncontrolled substance alternative for attention-deficit/hyperactivity disorder (ADHD) treatment. Although not approved by the Food and Drug Administration for depression, the selective norepinephrine reuptake inhibitor atomoxetine and was originally developed as an antidepressant.4,5 Approximately 1 year after the boxed warning for antidepressants in 2004, the Food and Drug Administration and the European Medicines Agency directed the manufacturer of atomoxetine to include a boxed warning regarding an increased risk of suicidal ideation in children and adolescents treated for ADHD.3,6 For clinical context, the EU summary of product characteristics for methylphenidate lists “suicidal tendencies” as a contraindication, whereas US labels do not.
The boxed warning decision for atomoxetine was based on a meta-analysis including 14 trials with 2208 patients (1357 atomoxetine/851 placebo). The analysis showed a statistically significant higher risk of suicidal ideation in the atomoxetine treatment arm with 5 cases paralleled by none in the placebo group. One patient attempted suicide during atomoxetine treatment compared with none on placebo. All cases of suicidal events occurred in children younger than 12 years and within 32 days of treatment initiation. Although the age range of study subjects was 6 to 17.9 years, the mean age was 10.5 years (SD ± 2.4), indicating a population predominantly comprised younger children.7 Similarly, the follow-up time ranged from 6 to 18 weeks but was skewed toward shorter follow-up periods. Two additional meta-analyses were published that were likewise compromised by sample size, resulting in limited inferences for rare events.8,9 Also, there is some evidence of increased rates of suicide in nontrial populations.10
In summary, available evidence lacks inferences for nonclinical trial populations, older adolescents, risk after 3 months of treatment, and, importantly, whether suicidal ideation indeed manifests in risk of suicide.7
ADHD is the most common mental health disorder in children and adolescents, with ~2.7 million youths receiving pharmacotherapy for treatment of ADHD in the United States.11–13 Although central nervous system stimulants are the principal and most common pharmacotherapy, an estimated 15% of youths with ADHD received atomoxetine in 2003.12,14 The objective of this study was to evaluate whether atomoxetine is associated with an increased risk of suicide attempt and suicide in patients newly treated with atomoxetine when compared with use of stimulants.
The study cohort was assembled from Medicaid Analytic eXtract (MAX) data, consisting of administrative health care claims, obtained from the 26 US states with the largest pediatric populations eligible for Medicaid fee-for-service benefits between 1999 and 2006. MAX data, made available by the Centers for Medicare and Medicaid Services, provide details on Medicaid eligibility, demographic information, diagnoses and procedures associated with in- and outpatient visits, as well as medications reimbursed by Medicaid.
In this cohort study, subjects entered the cohort at the first dispensed prescription (index date) for atomoxetine or stimulants. Employing a new user design, the index date had to be preceded by a minimum of 6 months of continuous Medicaid eligibility (baseline period) with at least 1 diagnosis of a mental health disorder commonly treated with atomoxetine or stimulants. Included disorders defined by International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes included besides ADHD, adjustment reaction, conduct disorder or mixed emotional disturbances of childhood or adolescence, and unspecified emotional disturbance of childhood or adolescence (Supplemental Table 5).15 All subjects had to be at least 5 and not >18 years of age at the index date. Subjects were excluded if they had drug claims for pemoline or methamphetamine because of low utilization or monoamine oxidase inhibitors because they are contraindicated during treatment with stimulants. Finally, we excluded subjects with severe or terminal diseases that alter the baseline risk for suicidality and that were generally rare, including any diagnosis of HIV/AIDS, malignant neoplasm, organ transplant, dialysis dependency, pervasive developmental disorders, or severe or profound mental retardation (Supplemental Table 6).16
Although only stimulants were considered first-line therapy of ADHD during the study period, we observed about half of all atomoxetine initiations in pharmacotherapy-naive patients. Furthermore, we observed significant differences in the baseline characteristics of patients in which atomoxetine was introduced as first- versus second-line treatment. Therefore, we established 2 subcohorts to evaluate suicidal risk separately for first- and second-line atomoxetine treatment. The first subcohort included subjects initiating treatment for the first time with either atomoxetine or stimulant between 2003 and 2006 (first-line treatment cohort) following a minimum of 180 days of continuous Medicaid eligibility. For the second subcohort, we matched second-line atomoxetine initiators (who either switched to or added atomoxetine after initial treatment with stimulants) by the number of months since stimulant initiation, to patients exposed to stimulants at the same number of months since stimulant treatment initiation in a 1 to 3 ratio (second-line treatment cohort). The index date for the second-line treatment cohort was the date of matching and also required a 180-day baseline period of continuous Medicaid eligibility immediately before the matching date.
Subjects were followed until the end of Medicaid eligibility, their 19th birthday, death, a hospitalization of >30 days, or pregnancy, whichever occurred first.
The primary study end point was a composite including completed suicide and suicide attempt requiring hospitalization or an emergency department visit. To identify suicides, we linked subjects identified in MAX to the Social Security Agency Death Master File. All deaths obtained through this linkage or flagged in the MAX eligibility files were then verified with the National Death Index. We defined completed suicides based on the ICD, Tenth Revision (ICD-10) codes X60–X84 on the National Death Index death certificate.17
Suicide attempts were identified from billing records for emergency department visits or hospitalizations with ICD-9-CM codes for external cause of injury E950.x–E959.x involving deliberate self-harm.18 Previous research has shown adequate sensitivity and specificity >90% as well as positive predictive values >85% for these end points.19–22
Periods of atomoxetine or stimulant exposure, including any dose or dosage form of atomoxetine, methylphenidate, and mixed amphetamine salts, were defined on the basis of pharmacy dispensing claims. We defined begin of atomoxetine and stimulant exposure based on the filling date on pharmacy claims for respective prescriptions. The end date for each prescription fill was calculated from the recorded dispensed days’ supply plus a grace period of 25% to incorporate residual supply as a result of drug holidays (eg, days without school). Because many states restrict dispensing amounts of controlled substances, the majority (>85%) of prescription fills for stimulants but also for atomoxetine involved a 30-day supply.23
If active prescriptions for both medications were present, exposure was defined as current atomoxetine exposure and flagged as dual therapy. Periods after current use were defined as former use.
We ascertained potential confounding variables from the 6-month baseline period preceding the index date, including age and calendar year, gender, race/ethnicity, state of residence, reasons for Medicaid eligibility (eg foster care), the number of hospitalizations for mental and nonmental diagnoses, and diagnoses of other psychiatric disorders such as substance use disorder, anxiety, bipolar disorder, schizophrenia, depression, or oppositional defiant disorder (Supplemental Table 7).24 Because Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, criteria to measure ADHD severity are not reflected in ICD-9-CM coding, we only distinguished ADHD subtypes with regard to presence or absence of hyperactivity. We also incorporated measures of the total number of distinct psychiatric disorder diagnoses and psychotropic drug classes during baseline as an indicator of mental illness severity. Variables capturing exposure to other psychotropic drugs during baseline period included antidepressants, anticonvulsants, antipsychotics, anxiolytics, α-agonists, lithium, and opioid analgesics (Supplemental Table 8).
Finally, we captured any suicide attempt and in/outpatient visits involving suicidal ideation (ICD-9-CM V62.84, 300.9) during baseline.18
For each subcohort, we used logistic regression models to calculate exposure propensity scores to estimate the likelihood to receive atomoxetine conditional on baseline covariates.25,26 The propensity score is a common method to control for confounding in observational research with the advantage to summarize numerous covariates as a single composite score, especially when the number of observed covariates was large and the number of observed outcomes was small.25–29
Participants were then weighted by the inverse of their propensity score to evaluate the level of balance achieved between exposed and unexposed groups across all baseline covariates.
We fitted 2 separate Cox proportional hazard models comparing new users of atomoxetine versus new use of stimulants (first-line user cohort) and patients starting atomoxetine treatment after initial stimulant treatment to patients who continued use of stimulants (second-line user cohort). Models were adjusted for the propensity score, as well as time-varying age and presence of dual therapy.
For computational efficiency, we segmented follow-up time after the index date into 15-day increments with exposure status determined based on the majority of days assigned to current or former use of atomoxetine or stimulants.
The cohort included 297315 patients initiating ADHD treatment (first-line) with atomoxetine (56012, 20.1%) or stimulants (223303, 79.9%) and accrued 428272 person-years of follow-up (Fig 1). Patients treated with stimulants contributed 190026 person-years (44.3%) and 144144 person-years (33.7%) of current and former stimulant exposure, whereas patients initiating atomoxetine treatment contributed 46929 person-years (11.0%) and 47173 person-years (11.0%) of current and former atomoxetine exposure, respectively. The most common reasons for censoring were end of Medicaid eligibility (94.9%) and hospitalization >30 days (3.5%). A total of 92 (0.03%) children and adolescents died during follow-up of causes other than the study end point. In general, covariates were similarly distributed among the 2 treatment groups with age and calendar year showing the greatest imbalance. Inverse weighting of subjects by their propensity score established balanced groups with <0.5% absolute difference (Table 1).
We observed a total of 140 suicidal events (suicide or suicide attempt). The majority (60%) of suicide attempts occurred in girls, and the majority of suicides (89%) was in boys. The average age at a suicidal event was 15.5 years (SD ± 2.7) and occurred after a mean of 1.12 years (median 0.93 years) after the index date.
We observed 50 suicidal events during current stimulant exposure (26.3 per 100000 person-years), 47 during former stimulant use (32.6 per 100000 person-years), 18 during current atomoxetine use (38.4 per 100000 person-years) and 25 suicidal events during former atomoxetine use (53.0 per 100000 person-years).
During the first year of follow-up, the unadjusted hazard ratio (HR) for current atomoxetine use compared with current stimulant use was 1.51 (95% confidence interval [CI] 0.77–2.95, P = .23) (Table 2). The fully adjusted HR for current atomoxetine use compared with current stimulant use was 0.95 (95% CI 0.47–1.92, P = .88). Varying follow-up times showed no appreciable effect on risk estimates.
The second-line treatment cohort included after matching 220215 patients who were initially treated with stimulants, contributing a combined 300772 person-years of follow-up. Of those, 56948 (25.9%) subsequently initiated atomoxetine contributing 37948 person-years (12.6%) and 43344 person-years (14.2%) of current and former atomoxetine exposure, respectively. A total of 163267 patients (74.1%) still exposed to stimulants were matched to these second-line atomoxetine initiators by month since initial stimulant treatment start, contributing 142015 person-years (47.2%) and 78054 person-years (26.0%) of current and former stimulant exposure, respectively (Fig 1). The most common reason for censoring were end of Medicaid eligibility (94.2%) and hospitalization >30 days (3.9%). A total of 29499 (12.2%) patients were excluded.
In general, covariates were similarly distributed among the 2 treatment groups except of the calendar year of their entry into the study. Inverse weighting of subjects by their propensity score established balanced groups with <0.5% absolute difference (Table 3).
We observed 46 suicidal events during current stimulant treatment (32.4 per 100000 person-years), 17 during former stimulant use (21.8 per 100000 person-years), 11 during current atomoxetine use (29.0 per 100000 person-years) and 16 suicidal events during former atomoxetine use (37.4 per 100000 person-years) (Table 4). The majority of suicide attempts occurred in girls (60%), whereas the majority of suicides were in boys (73%). The average age at a suicidal event was 14.7 years (SD ± 2.4) and occurred after a mean of 0.98 years (median 0.6 years) after the index date.
During the first year of follow-up, the unadjusted HR for current atomoxetine use compared with current stimulant use was 0.88 (95% CI 0.40–1.91, P = .74). When adjusted for the propensity score, age and dual therapy, the HR for current atomoxetine use was 0.71 (95% CI 0.30–1.67, P = .43 (Table 4). Varying follow-up times showed no appreciable effect on HRs.
Our study did not observe a statistically or clinically meaningful increase in the risk of suicidal events (suicide or suicide attempt) associated with first- or second-line treatment of youth age 5 to 18 with atomoxetine compared with stimulants. All point estimates were close to 1, consistent over varying periods of follow-up time, and consistent among current and former use, indicating no excess immediate or residual suicidal risk. Adjustments for age, dual therapy, and propensity score usually decreased HRs, suggesting that atomoxetine was not channeled toward patients at lower suicide risk.
It is noteworthy that most of the study time was before the possible risk of suicidal ideation was communicated. Our drug utilization analyses (data not shown) as well as other published data show a steep increase in atomoxetine utilization early after approval in 2002 followed by a gradual decline starting in 2004.32 We also found that atomoxetine users were older, twice as likely to have substance use disorder, had more oppositional defiant disorder, and distinctly more depression than stimulant users, all significant risk factors for suicide (Table 1). Thus, restricting of our study cohort to the early years of atomoxetine use alleviate concerns that that atomoxetine may have been channeled toward patients less risk for suicide.
Finally, unadjusted HRs suggested that atomoxetine users were at higher risk for suicide or suicide attempt, an association that vanished if adjusted for our measured confounders. Any residual (unmeasured confounder) that could mask an elevated suicidal risk of atomoxetine would need to have the opposite association than the confounding effects of age, substance use disorder, oppositional defiant disorder, and depression.
Importantly, because CIs of all HRs were wide, our study cannot exclude an excess risk of atomoxetine smaller than 40% to 70% (depending on follow-up time). However, considering the baseline incidence rate (during stimulant use) of 30 suicidal events per 100000 person-years, the resulting increase in the absolute risk would be small. Even if general concerns about bias are considered, the observed incidence rates provide assurance regarding a limited potential for clinically significant risk differences.
Current treatment recommendations emphasize the evidence that supports the efficacy of stimulants in the treatment of ADHD but indicate that atomoxetine and α-agonists may offer viable alternatives.33 Guidelines also point to the varying side effect profiles and make special note of treatment of adolescents in light of concerns about diversion and substance abuse associated with stimulants. Our real-world findings should be integrated in treatment decisions that weigh stimulant and atomoxetine effectiveness against their respective side effect profiles, especially considering the demonstrated risk for injury and potential self-harm associated with untreated ADHD itself.34,35
Interestingly, former use periods of both atomoxetine and stimulants showed consistent trends toward a reduced risk compared with current use periods. One possible explanation is residual confounding, which might be more pronounced in drug user to nonuser than in head-to-head comparisons. For example, patients who discontinue treatment altogether may have dissipating ADHD or comorbidity severity, resulting in reduced suicidal risk. Alternatively, ADHD treatment in adolescents and young adults could be associated with substance use problems, which in turn present a critical risk factor for suicide.
Major strengths of our study are its large population, its new-user design and balanced treatment groups resulting from propensity score adjustment.36–38 Our study included data from 4 years and 26 US state Medicaid programs. Although mostly Caucasian (60%), our study population allowed good representation of Hispanic, African American, and vulnerable pediatric populations with complex psychiatric needs. We established balanced treatment groups by restriction to patients with indication for stimulant or atomoxetine treatment and with comprehensive coverage of health care services, permitting a broad selection of covariates for propensity score adjustment. We further stratified our analysis to patients who initiated treatment with either stimulant or atomoxetine versus those who switched or added atomoxetine after stimulant treatment had been initiated. Because this second-line treatment group had a longer history of mental health problems and related treatment and an increased suicidal risk, stratification improved our ability to establish balance between treatment groups. Follow-up time was sufficient to cover the time to development of suicidal events as suggested by clinical trial data.
Our study is based on health care claims data, intended for reimbursement and not as electronic medical records. Therefore, codes used for billing practices might not accurately reflect clinical diagnoses, and pharmacy claims, representing dispensed prescriptions, do not necessarily imply drug utilization. However, Medicaid pharmacy data have been validated to define psychotropic drug exposure with a positive and negative predictive value >85%.39 Some important covariates influencing treatment decisions and/or suicidality risk such as disease severity, violence, or other stressful life events are not captured in administrative health data. The following sections discuss the effect of potential biases on our results.
First, our ability to detect suicidal events was dependent on clinician diagnoses, and reported incidence estimates might be underestimated. Consistency of our rates with previous estimates provides some assurance about the sensitivity of our method.24,40,41 Gender distribution and the distribution of methods to commit/attempt suicide were also consistent with previous research.24 Importantly, reduced sensitivity of our outcome ascertainment would have resulted in reduced statistical power but not systematically biased HRs.
Alternatively, publicized safety concerns might have resulted in increased clinician awareness, resulting in increased suicidal diagnoses in atomoxetine users. We minimized diagnostic bias by including only events with significant harm rather than investigating suicidal ideation and by excluding self-harm with ambiguous intent.16,42–44 Importantly, if such bias were present, it would alter risk estimates toward an increased risk during atomoxetine exposure, which we did not observe.
Second, it is conceivable that suicidal patients were channeled toward atomoxetine because of lesser concerns about substance abuse. We minimized confounding by using an active comparator and by requiring a mental health diagnosis that is consistent with indications for both atomoxetine and stimulants.26 Sensitivity analyses requiring at least 2 diagnoses to determine presence of mental disorders or excluding high-risk patients with history of suicidal ideation or suicide attempt showed no appreciable effects on HRs. Of note, because our confounding adjustment alleviated an initially increased risk of atomoxetine, it is unlikely that adjustment for residual confounders would have an opposite effect.
Finally, in addition to the low incidence of suicidal events, limited follow-up resulted in wide confidence of risk comparisons. Statistical power also limited our ability to stratify analyses to high-risk groups or long-term users of atomoxetine. However, the overall small suicidal event rates indicate a small absolute risk increase potential in typical clinical practice.
First- and second-line treatment of youth aged 5 to 18 with atomoxetine compared with stimulant treatment was not associated with an increased risk of suicide attempts requiring medical attention and suicides in current practice. Limited utilization periods of atomoxetine and low incidence of suicidal events resulted in limited statistical power, which did not allow stratified analysis of high-risk groups or assessment of suicidal risk associated with long-term use. However, the observed low suicidal event numbers indicate a small absolute risk increase potential in typical clinical practice.
This study was approved by the University of Florida Institutional Review and Privacy Boards (269-2012), National Death Index (2012-0044), and the Centers for Medicare and Medicaid Services Privacy Board (DUA 23778), including a waiver for the Health Insurance Portability and Accountability Act authorization. Data were obtained under data users agreements with the Centers for Medicare and Medicaid and the National Death Index and are not available for sharing.
Dr Linden conceptualized as well as designed the study and conducted all data analyses; had full access to all the data in the study; takes responsibility for the integrity of the data and the accuracy of the data analysis; affirms that the manuscript is an honest, accurate, and transparent account of the study being reported and no important aspects of the study have been omitted; finally he interpreted the results and drafted the manuscript; Drs Bussing, Gerhard, Segal, and Shuster conceptualized and designed the study and interpreted the results and provided critical revisions of the manuscript for important intellectual content; Mr Kubilis conducted the data acquisition and management, conceptualized and designed the study, interpreted the results, and provided critical revisions of the manuscript for important intellectual content; Dr Winterstein organized the data acquisition, conceptualized and designed the study, participated in all data analyses, interpreted the results, and drafted the manuscript; all authors approved the final manuscript.
FINANCIAL DISCLOSURE: The authors have indicated they have no financial relationships relevant to this article to disclose.
FUNDING: The Centers for Medicaid and Medicare Services granted free data reuse for doctoral research. This study was partially funded by a grant from the Agency for Healthcare Research and Quality (R01-HS0185606). Dr Linden’s current affiliation is Boehringer Ingelheim GmbH. Boehringer Ingelheim GmbH did not contribute any direct or indirect financing to this study and did not influence the content of the publication in any aspect at any time. At the time of study conduct, Dr Linden was in the Department of Pharmaceutical Outcomes and Policy at the University of Florida. Dr Linden was partially supported by the Mary Kay Owens Healthcare Innovation and the DuBow Family fellowships. Dr Bussing has received past research support from Otsuka and Pfizer and served as consultant to Pfizer; her contribution was not supported by any grant mechanism. Dr Gerhard’s contribution was partially supported by grant HS 106097 from the Agency for Healthcare Research and Quality (Centers for Education and Research on Therapeutics). Dr Shuster’s contribution was partially supported by grant 1UL1TR000064 from the National Center for Translational Sciences, National Institutes of Health. None of these entities had influence on the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.
POTENTIAL CONFLICT OF INTEREST: The authors have indicated they have no potential conflicts of interest to disclose.