DSH indicates deliberate self-harm; RD, risk difference. The dotted line indicates the strength of confounding implied if true RD is 2 attempts per 1000. The shaded area indicates the strength of confounding implied if the true risk difference is null or protective.
eTable 1. Propensity Score Model Odds Ratios, Ages 10-24 Years
eTable 2. Propensity Score Model Odds Ratios, Ages 25-64 Years
Customize your JAMA Network experience by selecting one or more topics from the list below.
Miller M, Swanson SA, Azrael D, Pate V, Stürmer T. Antidepressant Dose, Age, and the Risk of Deliberate Self-harm. JAMA Intern Med. 2014;174(6):899–909. doi:10.1001/jamainternmed.2014.1053
Copyright 2014 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.
A comprehensive meta-analysis of randomized trial data suggests that suicidal behavior is twice as likely when children and young adults are randomized to antidepressants compared with when they are randomized to placebo. Drug-related risk was not elevated for adults older than 24 years. To our knowledge, no study to date has examined whether the risk of suicidal behavior is related to antidepressant dose, and if so, whether risk depends on a patient’s age.
To assess the risk of deliberate self-harm by antidepressant dose, by age group.
Design, Setting, and Participants
This was a propensity score–matched cohort study using population-based health care utilization data from 162 625 US residents with depression ages 10 to 64 years who initiated antidepressant therapy with selective serotonin reuptake inhibitors at modal or at higher than modal doses from January 1, 1998, through December 31, 2010.
Main Outcomes and Measures
International Classification of Diseases, Ninth Revision (ICD-9) external cause of injury codes E950.x-E958.x (deliberate self-harm).
The rate of deliberate self-harm among children and adults 24 years of age or younger who initiated high-dose therapy was approximately twice as high as among matched patients initiating modal-dose therapy (hazard ratio [HR], 2.2 [95% CI, 1.6-3.0]), corresponding to approximately 1 additional event for every 150 such patients treated with high-dose (instead of modal-dose) therapy. For adults 25 to 64 years of age, the absolute risk of suicidal behavior was far lower and the effective risk difference null (HR, 1.2 [95% CI, 0.8-1.9]).
Conclusions and Relevance
Children and young adults initiating therapy with antidepressants at high-therapeutic (rather than modal-therapeutic) doses seem to be at heightened risk of deliberate self-harm. Considered in light of recent meta-analyses concluding that the efficacy of antidepressant therapy for youth seems to be modest, and separate evidence that antidepressant dose is generally unrelated to therapeutic efficacy, our findings offer clinicians an additional incentive to avoid initiating pharmacotherapy at high-therapeutic doses and to closely monitor patients starting antidepressants, especially youth, for several months.
Quiz Ref IDThe US Food and Drug Administration’s (FDA) meta-analysis of antidepressant trials found that children randomized to receive antidepressants had twice the rate of suicidal ideation and behavior compared with children who received placebo.1 Meta-analysis of adult placebo-controlled trials found that participants 18 to 24 years of age randomized to receive antidepressants were at elevated risk of suicidal thoughts and behavior, those 25 to 64 years of age were at equal risk, and those 65 years or older were at lower risk.2
Nonrandomized studies can address some of the limitations of existing randomized antidepressant trials, including their short duration, the small number of suicide-related events observed, the homogeneity of participants, and the different antidepressant types and doses administered across trials. Nonrandomized studies, however, require careful consideration of confounding, especially with respect to the indication for using antidepressants in the first place.3-8 To minimize such confounding, rigorous observational studies have focused on treatment initiation9 and avoided nonuser comparison groups.10,11 In addition, because suicide attempts may lead clinicians to prescribe antidepressants, careful research has eschewed before vs after study designs and instead focused on whether deliberate self-harm (DSH) differs across antidepressant classes and agents. In general, these studies have reported either no evidence of differential risk across class or agent, or small and inconsistent differences.12-19
Patients exposed to higher doses of antidepressants tend to experience more frequent and severe adverse effects, including putatively suicidogenic ones, such as akathisia,20-27 compared with patients prescribed lower doses.28,29 Despite this dose-related phenomenon and scant evidence that higher doses are more effective in alleviating depressive symptoms,28,29 neither the FDA meta-analyses nor any observational study to date has examined whether the risk of suicidal behavior is related to antidepressant dose. The current study takes up this question among a cohort of initiators of antidepressant therapy and addresses as well whether dose-related risk is modified by a patient’s age.
The current cohort study involved 162 625 patients 10 to 64 years of age with a depression diagnosis who initiated therapy with selective serotonin reuptake inhibitors (SSRIs) from January 1, 1998, through December 31, 2010. Initiation was defined as filling an SSRI antidepressant prescription without evidence of prescriptions fills for any class of antidepressants in the preceding 12 months. Analyses focus on the first treatment episode. Eligibility required evidence of depression as indicated by an International Classification of Diseases, Ninth Revision (ICD-9) code for depression recorded during the 12 months prior to antidepressant initiation (Table 1 and Table 2). To allow uniform assessment and selection of all patients, participants were required to be actively enrolled in a contributing health plan for the 15 months prior to initiation (ie, 12 months for baseline covariate assessment + 2 months [ie, a 60-day grace period] + 30 days [ie, the usual days’ supply]). The cohort study was based on observational health care utilization data. Informed consent was not obtained for persons in the data set. The study was exempted by the Harvard School of Public Health institutional review board.
The PharMetrics Claims Database used in this study comprises commercial health plan information obtained from managed care plans throughout the United States. The database includes medical and pharmaceutical claims for over 61 million unique patients from over 98 health plans, and includes International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) inpatient and outpatient diagnoses, Current Procedural Terminology 4 and Healthcare Common Procedure Coding System procedure codes, and both retail and mail-order records of all reimbursed dispensed prescriptions. Our analyses focus on persons younger than 65 years, the age group for which our data are nationally representative of the commercially insured and for which we had sufficient sample size.
We restricted analyses to new initiators of therapy with antidepressants, rather than prevalent users, because such a design allows us to detect adverse events that follow soon after therapy with a drug is started, assess risks over time, and control for selection bias with baseline patient characteristics that are not influenced by effects of antidepressant treatment. Incident user designs also mitigate potential bias owing to a patient’s drug exposure history influencing current treatment assignment. To further minimize potential confounding by the class of antidepressants prescribed and the use of unusual agents, analyses were restricted to patients initiating therapy with the most commonly prescribed SSRIs (citalopram hydrobromide, sertraline hydrochloride, and fluoxetine hydrochloride), which together constitute 67% of all SSRI therapy initiated.
Patients were assigned to 1 of 3 empirically derived dose categories (modal dose, higher than modal, lower than modal) based on the distribution of doses prescribed among antidepressant initiators. In the 10- to 24-year cohort there were 32 504 in the modal dose category, 7117 in the higher than modal dose category, and 14 542 in the below modal dose category; in the 25- to 64-year-old cohort there were 99 316 in the modal dose category, 23 668 in the higher than modal dose categoy, and 20 065 in the lower than modal dose category. The modal daily dose for citalopram hydrobromide, sertraline hydrochloride, and fluoxetine hydrochloride were, respectively, 20 mg/d, 50 mg/d, and 20 mg/d. Participants who received doses below our empirically derived modal dose often received doses considered below the minimal effective dose for depression30-32; to minimize potential confounding, analyses were restricted to participants who received modal dose or doses higher than the modal dose. Patients receiving index doses in excess of the recommended maximum therapeutic dose constituted less than 2% of all patients (112 in the 10- to 24-year-cohort and 677 in the 25- to 64-year-old cohort) and, for similar reasons, were excluded from analysis; maximum daily doses for citalopram hydrobromide, sertraline hydrochloride, and fluoxetine hydrochloride were, respectively, 40 mg/d, 200 mg/d, and 80 mg/d. Patients initiating therapy with more than 1 drug or with more than 1 dosing regimen were also excluded from analyses.
Exposure status was assigned based on the initiated medication. Study follow-up began on the day after initiation of the first antidepressant therapy. When a dispensing occurred before the previous prescription should have run out, use of the new prescription was assumed to have begun the day after the end of the old prescription. Primary analyses used a 60-day grace period (≤60 days beyond the provided days’ supply can elapse before censoring).
Patients were censored the date they switched agents, added other antidepressant agents, changed dose, 360 days after the index date, ended enrollment in their health insurance plan, engaged in DSH, or the end of the study period, whichever came first.
The first occurrence of DSH was our outcome of interest, and it was defined as a medical claim with an ICD-9 external cause of injury code (e-code) (E950.x-E958.x).
Baseline patient characteristics included age, sex, medical comorbidities, treatment history, and concomitant medication use in the 12 months prior to initiation of antidepressant therapy. Psychiatric risk factors included the number of acute psychiatric hospitalizations, the number of acute hospitalizations for substance abuse, psychiatric comorbidity, and prior DSH. A hierarchy of depression was constructed as a function of the proximity of the most recent depression diagnosis to antidepressant initiation and whether the diagnosis was rendered during an inpatient stay or an outpatient visit.
Patients were divided into 2 age groups guided by the age-related risk of suicidal behavior identified in the FDA’s meta-analyses (ages 10-24 years vs 25-64 years). We estimated propensity scores for treatment initiation of high vs modal dose for each age group separately based on the patient characteristics described in the previous subsection (age is modeled as a continuous variable within each cohort). Up to 2 patients receiving a modal dose were matched to every patient receiving a high dose using an adaptation of a published propensity score algorithm.33 Thus, the research question we addressed is what would have happened, with respect to future DSH, if people who initiated high-dose therapy had instead initiated with modal-dose therapy.
Crude rates of DSH were calculated over the 1-year exposure period (ie, subject to censoring as described herein). Crude rates were also reported for 3 time periods after initiating therapy: days 1 to 30, days 31-90, and days 91-360. Exact methods were used to calculate 95% confidence intervals. Modified Poisson regression using an identity link was used to estimate the 90-day risk differences; Cox models were used to estimate hazard ratios (HRs).
Sensitivity analyses examined how robust our findings were to a range of grace periods. Additional subgroup analyses restricted participants to those without a prior suicide attempt and to those who had not received antidepressants in the 3 years prior to their index date. In mid-2005 a new diagnostic billing code for suicidal ideation became available. We used this historical development to assess the extent to which our primary analyses may have been confounded by this well-established predictor of DSH by examining the extent to which matched cohorts based on propensity score algorithms that did not incorporate suicidal ideation achieved balance on suicidal ideation and the extent to which adjusting for suicidal ideation as a covariate in Cox models attenuated the risk of suicide attempts associated with higher dose.
Bias analyses assessed the strength of residual confounding that would need to be present to fully explain associations found in our primary analyses.34,35 Specifically, we estimated the strength of a single dichotomous unmeasured confounder that would be necessary to nullify the estimated 90-day risk difference. Bias in the risk difference is dependent on 2 factors: how predictive the hypothetical confounder is of attempts, and how imbalanced the confounder is across high- vs modal-dose treatment groups (both assessed on the additive scale). A priori, we expected depression severity (eg, prior inpatient stays or psychiatric comorbidities) and a history of suicidality (eg, prior ideation or attempts) to be the strongest potential confounders. Accordingly, we present the magnitude of confounding introduced by these measured confounders in the prematched cohort along with the bias analyses.
Baseline patient characteristics in age-group–stratified, propensity score–matched cohorts of high- vs modal-dose users were well balanced across dose categories (Table 1 and Table 2). Quiz Ref IDFor example, the age and sex distributions among high- vs modal-dose initiators were almost identical, and the distribution of our constructed tiers of depression severity differed little across dose categories. Suicidal ideation was also closely balanced across dose cohorts within each age stratum even though it did not contribute to the matching algorithm. Approximately half of all patients (45%) with a depression diagnosis who initiated high-dose SSRIs therapy (30 785) filled prescriptions written by internists or general practitioners (13 859) (Table 1 and Table 2). Among patients younger than 25 years, internists and/or general practitioners wrote one-third (33%) of all initial high-dose prescriptions; psychiatrists and/or psychologists wrote 30% of all such prescriptions (Table 1). Among patients 25 to 64 years of age, internists and/or general practitioners wrote 49% of all such high-dose prescriptions; psychiatrists and/or psychologists wrote 21% of such prescriptions (Table 2). Roughly one-third (32%) of all high-dose antidepressant initiators received prescription written by other types of physicians (31% among patients aged 25 to 64 years [Table 2], 37% among patients aged ≤24 years [Table 1]; 10% of the latter physicians were pediatricians [data not shown]).
In our matched cohorts, 142 participants ages 10 to 24 years (68 modal-dose initiators, 74 high-dose initiators) engaged in DSH within 1 year of treatment initiation; the corresponding rates of DSH for the modal- vs high-dose initiators were 14.7 (95% CI, 11.5-18.5) and 31.5 (95% CI, 24.9-39.3) events per 1000 person-years, respectively. For participants ages 25 to 64 years, there were 81 such acts (49 in modal-dose initiators, 32 in high-dose initiators), corresponding to rates of 2.8 (95% CI, 2.1-3.6) and 3.2 (95% CI, 2.3-4.5) events per 1000 person-years for the modal- vs high-dose participants, respectively (Table 3). Although the hazards were proportional throughout the 1-year follow-up period, most of the events occurred in the first 3 months after initiation (Table 3, Figure 1).
Propensity score–matched analyses produced HRs that were substantially higher in the 10- to 24-year-old cohorts than in the older cohorts (Table 4): among 10- to 24-year-olds the rate of DSH among high-dose participants was approximately double that among modal-dose participants (HR, 2.2; 95% CI, 1.6-3.0); for participants 25 to 64 years of age, the HR was considerably lower (HR, 1.2; 95% CI, 0.8-1.9). The age group by dose interaction achieved statistical significance (P = .04) (data not shown).
Findings for the young cohort were robust to several different model specifications, including analyses that varied the grace period from 7 to 360 days, excluded participants without a DSH history prior to their index prescription. and among patients who were treatment naïve for at least 3 years (Table 4). The DSH HRs for high vs modal dose among those initiating therapy in 2006 or later were found to be consistent with those found for the entire study period. Moreover, DSH HRs among this population were virtually identical regardless of whether Cox models adjusted for suicidal ideation (because suicidal ideation, while strongly predictive of the outcome, was not strongly related to dose) (Table 4).
In our primary analysis of the 10- to 24-year-old cohort, for every 1000 patients initiating high-dose therapy there were approximately 7 (7.3) more DSH events over the first 90 days of treatment among high-dose initiators compared with modal-dose initiators (95% CI, 3.6-11.1) (Table 4). The corresponding number needed to harm was 136. For the older cohort, the risk difference was effectively zero. Not all acts of DSH are e-coded in claims data (eg, some overdose by drug events are known to occur but lack e-codes, leading to underestimates of both intentional and unintentional overdose event rates). Although this is likely to be nondifferential with respect to our exposure of interest and therefore is unlikely to bias estimates appreciably because event rates are underestimated, the number needed to harm we derived is likely a conservative estimate (ie, an upper bound on the true number needed to harm).
Figure 2 depicts how strong a risk factor and how imbalanced a hypothetical unmeasured confounder would need to be in order to nullify our 90-day risk difference in the younger cohort. As can be seen, a putative confounder that would nullify our findings (or even reduce the risk difference to what we considered a clinically meaningful difference of 2 events per 1000 patients) would need to be both far more predictive of subsequent DSH than our most highly predictive covariates (history of DSH, suicidal ideation, substance abuse, or inpatient hospitalization for depression) and also an order of magnitude more imbalanced across dose levels than our most imbalanced prematched covariate. It is important to note that such an unmeasured confounder would need to be this strong in the matched cohort; that is, it would need to be very strongly associated with dose and subsequent DSH even after adjustment for all the measured confounders accounted for in the propensity score–matching process.
To our knowledge, the current investigation is the first prospective cohort study to examine the relation between dose of antidepressants and the risk of DSH. Quiz Ref IDWe found that the rate of DSH for children and young adults was approximately twice as large among patients initiating high-dose therapy compared with those initiating modal-dose therapy. Given the high baseline rates of DSH among these patients, we expect approximately 1 additional DSH event for every 136 patients 10 to 24 years of age who are treated with high-dose therapy (instead of modal-dose therapy). For the older cohorts, the absolute risk of DSH was far lower, and the difference in risk between the cohorts was effectively null.
Several possible mechanisms linking antidepressant use to suicidal behavior have been suggested,20,23-26,36-39 including an early energizing effect that allows patients with depression to act on suicidal impulses, suicidogenic adverse effects (eg, akathisia, insomnia, panic attacks), episode-shifting effects (from depressive to manic episodes), and paradoxical worsening of depression. Although our study does not address the mechanisms whereby higher doses might lead to higher suicide risk, if depression-independent suicidogenic effects increase with dose, as has been observed for akathesia,28 but antidepressive effects are insensitive to doses within a broad therapeutic range, as seems to be the case,29,40-45 higher doses might produce a net tendency toward suicidal behavior.
To the extent that depression-independent suicidogenic effects of antidepressants exist, older adults may be less susceptible, on balance, if the antidepressive efficacy of antidepressants is superior for older adults compared with children and younger adults.46,47 Our finding that dose-related suicide risk seems to be more pronounced among children and young adults might also reflect an age-related susceptibility to suicidogenic effects of antidepressants independent of depression severity, as was observed in randomized trials with placebo controls.1
The elevated risk of DSH we observe among youth receiving therapy with high-dose antidepressants compared with those receiving therapy with a modal dose might also be due to more frequent and severe drug discontinuation syndromes among patients receiving high-dose therapy.48 Although we censor at known discontinuation of therapy, it is still possible that differential nonadherence and/or differential severity of withdrawal reactions due to nonadherence contributed to our findings. This form of nondifferential adherence would, however, bias findings to the null, as would poorer adherence that was related to untoward adverse effects, which in general tend to be more common among high-dose users, suggesting that our estimate of the risk of DSH associated with high-dose therapy is conservative. The robustness of our findings to grace periods as disparate as 1 week to 1 year also militates against withdrawal reactions playing a major role. Finally, the half-lives of our SSRIs are relatively long (range, 16-35 hours),49 making severe withdrawal reactions less likely.
To examine the extent to which confounding not explicitly modeled in our propensity score adjustment may account for our results, we applied our primary matching algorithm, which did not include a covariate for baseline suicidal ideation, to data from 2006 through 2010. Baseline suicidal ideation, while a potent predictor of subsequent DSH (as expected), was not associated with dose (even across unmatched cohorts), suggesting that other potential but unaccounted-for risk factors for DSH might also be reasonably balanced across our cohorts defined by dose. Although it is still possible that unmeasured confounding accounts for the dose-response relationship we observed, it is not obvious what other factors might have led to meaningful confounding of our results. Indeed, such an unmeasured confounder would have to possess a very strong association with both dose and suicidal behavior—and also remain largely uncorrelated with risk factors we explicitly accounted for in analyses. Quiz Ref IDEstimates from our bias analysis suggest that any such unmeasured confounder would need to be both more predictive of subsequent DSH than the most highly predictive risk factor in our data set (history of DSH) and also an order of magnitude more imbalanced across dose levels than our most imbalanced covariate.
When interpreting findings from the current study, one should bear in mind several additional caveats. First, as in all analyses relying on claims databases, we have limited ability to adjust for the severity of psychiatric illness. We do, however, use propensity score techniques to adjust for psychiatric comorbidity and comedication and for a proxy of depression severity involving whether a patient’s depression diagnosis occurred during an inpatient admission for depression, whether the diagnosis was a primary or secondary diagnosis, and whether the diagnosis occurred within the month prior to their index date or more remotely. Propensity scores offer an advantage in studies of rare outcomes (eg, DSH) because propensity scores model the relation of covariates and their interactions with the drug exposure (which is relatively frequent) and not directly with the study outcome (which is often rare), thereby mitigating the risk of overfitting in a traditional outcome model.50,51Quiz Ref IDAs is the case for all observational studies, however, our ability to adjust fully for underlying suicide risk at baseline depends on our ability to accurately classify baseline confounders—and is compromised to the extent that measurement and reporting of conditions coded on insurance claims are misclassified.52 Second, we used administrative data and therefore did not measure antidepressant adherence directly. Using automated prescription data may, however, more accurately measure use than studies that rely on data from self-report surveys.53-55 A related point is that we define drug exposure in our primary analysis in a way that seeks to capture how patients fill their medications (ie, analyses are “as treated”) but in so doing admit possible selection bias owing to censoring.56 Nevertheless, our findings were robust to analyses in which exposure was defined using “first treatment carried forward,” which is not subject to immeasurable time bias or other selection biases due to censoring, but rather likely bias findings toward the null (especially over extended follow-up periods). Finally, it should be noted that although our study provides strong evidence against initiating adolescents and young adults with depression using high-dose antidepressant therapy, it does not address whether initiating antidepressant therapy with the most commonly prescribed (modal) dose increases or decreases the risk of DSH relative to no pharmacological treatment, or, for that matter, if our findings apply to patients with other indications. While the question “Does prescribing antidepressants increase or decrease suicide risk?” is a question of great clinical importance (and controversy), we decided against using untreated patients as the reference group to minimize the potential for confounding by indication.
In our study, approximately half of all patients initiating high-dose antidepressant therapy filled prescriptions written by internists or general practitioners. This statistic, coupled with the acknowledgment that treatment decisions should be made on the basis of expected benefits and harms, underscores the relevance of our findings to clinicians caring for patients in both specialty and nonspecialty settings. Considered in light of recent meta-analyses concluding that the efficacy of antidepressant therapy for youth seems to be modest,46,47 and separate evidence that dose is generally unrelated to the therapeutic efficacy of antidepressants,29,40-45 our findings offer clinicians an additional incentive to avoid initiating pharmacotherapy at high-therapeutic doses and to monitor all patients starting antidepressants, especially youth, for several months and regardless of history of DSH.
Corresponding Author: Matthew Miller, MD, ScD, Department of Health Policy and Management, Harvard School of Public Health, 677 Huntington Ave, Boston, MA 02115 (firstname.lastname@example.org).
Accepted for Publication: November 18, 2013.
Published Online: April 28, 2014. doi:10.1001/jamainternmed.2014.1053.
Author Contributions: Dr Miller had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: Miller, Azrael, Pate, Stürmer.
Acquisition, analysis, or interpretation of data: All authors.
Drafting of the manuscript: Miller, Swanson, Azrael.
Critical revision of the manuscript for important intellectual content: Azrael, Pate, Stürmer.
Statistical analysis: Swanson, Azrael, Pate, Stürmer.
Obtained funding: Miller.
Administrative, technical, or material support: Miller, Pate.
Study supervision: Miller, Stürmer.
Conflict of Interest Disclosures: Dr Stürmer does not accept personal compensation of any kind from any pharmaceutical company, although he receives salary support from the Center for Pharmacoepidemiology and from unrestricted research grants from pharmaceutical companies (GlaxoSmithKline, Merck, Sanofi) to the department of epidemiology, University of North Carolina at Chapel Hill. No other disclosures are reported.
Funding/Support: Drs Miller, Azrael, Pate, and Stürmer received support for this work from an investigator-initiated research grant (RO1MH085021) from the National Institute of Mental Health (principal investigator, Dr Miller ). Dr Stürmer receives investigator-initiated research funding and support as principal investigator (grant R01 AG023178) from the National Institute on Aging at the National Institutes of Health. He also receives research funding as principal investigator of the UNC-DEcIDE Center from the Agency for Healthcare Research and Quality.
Role of the Sponsor: The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.