If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Address correspondence to: Djøra I. Soeteman, Center for Health Decision Science, Harvard School of Public Health, 718 Huntington Ave, 2nd Fl, Boston, MA 02115
Center for Health Decision Science, Harvard School of Public Health, Boston, MA, USADepartment of Health Policy and Management, Harvard School of Public Health, Boston, MA, USA
To quantify the trade-offs of alternative strategies in treating pediatric major depressive disorder with respect to the clinical benefit and risk of fatal and nonfatal suicidal behavior over a 5-year time horizon.
Methods
We developed a disease simulation model integrating epidemiological and clinical data from the literature to simulate the effect of selective serotonin reuptake inhibitors (SSRIs), cognitive behavioral therapy (CBT), and a combination of both on a US pediatric population with major depressive disorder.
Results
In a cohort of 1,000,000 simulated individuals (ages 10–24 years), the use of SSRIs was associated with the highest number of suicide-related events, while CBT was associated with the lowest number. Over a 5-year period, the strategy with the highest number of symptom-free weeks depended on assumptions made regarding treatment efficacy beyond the available clinical data.
Conclusions
Considering the risk-benefit profile over a 5-year period, CBT offers a safer profile than combination treatment or SSRIs alone with respect to suicide deaths and attempts. Any additional benefits of SSRIs, either alone or in combination with CBT, must be weighed against the expected increase in suicides.
Since October 2004, antidepressants have carried a black-box warning indicating an increased risk of suicidal ideation and behavior in children and adolescents. These warnings are based on results from the US Food and Drug Administration (FDA) meta-analyses of placebo-controlled randomized controlled trials (RCTs) of antidepressants among children and adolescents, which found an approximate twofold increase in the risk of suicidal ideation and behavior among children and adolescents randomized to antidepressants (4%) compared with placebo (2%) [
]. In 2007, the warning was extended to young adults aged 18 to 24 years, based on a reexamination of age-related trend data from the FDA meta-analysis [
A more recent meta-analysis, which included two additional RCTs, found that the risk of suicidal thinking and behavior among pediatric patients with major depressive disorder (MDD) on antidepressants was increased significantly under model specifications used in the FDA meta-analysis (i.e., fixed effects models), relative risk (RR) = 1.9 (1.2–2.9), but not under different model specifications (i.e., random effects models), RR = 1.6 (1.0–2.7) [
Clinical response and risk for reported suicidal ideation and suicide attempts in pediatric antidepressant treatment: a meta-analysis of randomized controlled trials.
]. The authors concluded that the overall benefits of antidepressant use in this population outweighed the risks. Because this study was confined to information from placebo-controlled RCTs, it was unable to examine the trade-off of risks and benefits under conditions in which the alternative to antidepressant treatment was anything other than placebo, such as cognitive behavioral therapy (CBT) and combination treatments. In addition, the meta-analysis could not address how the overall benefits and risks of antidepressant treatment extended over longer time horizons, because the included RCTs were of short duration (8–12 weeks).
Under conditions of uncertainty, decision-analytic methods have been used to extrapolate short-term effects observed in empirical studies to project longer-term health outcomes under different “what-if” scenarios. Using such an approach, the current study complements prior analyses of depression treatment for children and young adults by incorporating empirical evidence from clinical studies in a disease simulation model to compare three treatment strategies (selective serotonin reuptake inhibitors [SSRIs], CBT, and a combination of SSRIs and CBT) in terms of symptom-free weeks and suicidality over a 5-year time horizon. The main focus of the study was to explicitly describe the trade-off in risks and benefits associated with these three competing strategies and to examine the influence of plausible yet empirically uncertain assumptions regarding suicide attempt risks and patients' response to treatment.
Methods
Model
We developed a discrete event simulation (DES) model (TreeAge Pro 2008, Williamstown, MA) integrating empirical data on the natural history of depression and clinical effects of treatments from the published literature to simulate a real-world population of children and young adults with a clinical diagnosis of MDD. We assumed a 5-year analytic time horizon; although MDD events manifest over a longer time horizon, we elected a conservative time frame of 5 years given the short duration of the clinical trial data on clinical effects and risks of treatments. Unlike many decision-analytic models that follow an entire cohort simultaneously over time (e.g., Markov), the DES approach simulates individual patients one at a time, allowing for consideration of important patient-specific sociodemographic characteristics and individual variation of disease history [
]. Importantly, the DES method relaxes the fixed time assumption that is characteristic of Markov models, allowing for reflection of time to event that may vary from patient to patient, either stochastically or because of an individual's history. The model is then used to estimate the impact of different interventions on the average patient population.
The schematic of the DES model is shown in Figure 1. The competing events that are possible during the simulation include the following: 1) occurrence of a major depressive episode, 2) full symptom relief, 3) partial symptom relief, 4) suicide attempt, 5) suicide death, and 6) death from other causes (age- and sex-specific). At the start of the analysis, an individual with a major depressive episode enters the model and is assigned a starting age (ranging from 10 to 24 years) and sex from a distribution. The major depressive episode can result in one of four events: full symptom relief, partial symptom relief, suicide attempt, or death. The order and timing of events are governed by time-to-event data from survival distributions for each possible event [
The use of modelling to evaluate new drugs for patients with a chronic condition: the case of antibodies against tumour necrosis factor in rheumatoid arthritis.
]; once a new event occurs, the time elapsed is noted and the age of the patient is updated. If the patient experiences full symptom relief (i.e., Children's Depression Rating Scale-Revised score of 28 or less), the competing events that may occur are recurrent depressive episode, suicide attempt, or death. Partial symptom relief is experienced by those who respond to treatment but fail to remit in a given MDD episode (i.e., an improvement score of 1 [very much improved] or 2 [much improved] on the Clinical Global Impressions-Improvement and a Children's Depression Rating Scale-Revised score of 29 or more) [
]; during partial symptom relief, the competing events that are possible include relapse into a depressive episode, suicide attempt, and death. A suicide attempt can either be a fatal event or a nonfatal event. In the event of a nonfatal suicide attempt, we assumed that patients returned to the health status they were experiencing at the time of the attempt.
To achieve convergence in model outcomes, a population of 1,000,000 patients was simulated for each treatment strategy, and the history of events for each patient was tracked over a 5-year period and aggregated. Model outcomes included symptom-free weeks and numbers of both fatal and nonfatal suicidal acts. The symptom-free weeks were estimated as the duration of time spent with no symptoms (i.e., in full symptom relief).
Data and assumptions
Natural history
The age and sex distribution shown in Table 1 was based on data from the Sequenced Treatment Alternatives to Relieve Depression study, which reported the age of onset of the study participants' first major depressive episode, ranging from childhood onset (ages <12 years) to late adult onset (ages ≥60 years) [
]. According to evidence in the literature that women are twice as likely to be depressed compared with men and that this sex gap emerges by age 14 years, we applied a male/female ratio in depression of 1:1 for children aged 10 to 13 years and a ratio of 1:2 for children and young adults aged 14 to 24 years [
We used survival distributions conditional on an individual's current age and past events to inform time to events (Table 2). Natural history data on episode duration were obtained from a Canadian study of individuals (ages 15 years or older) experiencing a first depressive episode [
]. Time to recurrence of an episode for individuals experiencing full symptom relief was influenced by two factors: 1) duration of time without symptoms, such that the risk of recurrence decreased with longer duration of full symptom relief, and 2) number of previous depressive episodes, such that the risk of recurrence increased (by 16%) with each successive episode [
]. For those individuals experiencing partial symptom relief, we calculated an increased risk of relapse of 48%, based on the probability of relapse in the presence of residual symptoms versus the absence of residual symptoms in the placebo arm of an RCT [
General population data on fatal and nonfatal suicidal acts, stratified by age and sex, were obtained from the Centers for Disease Control and Prevention [
] for years 2001 to 2006. A recent study found that the risk of suicide attempts varied depending on the level of depression, with the likelihood being nearly eightfold higher during major depressive episodes and fourfold higher during partial remission compared with that during full remission [
]. We therefore assumed hazard ratios of suicide attempts of 7.74 during an episode and 4.20 during partial symptom relief in the base-case analysis but also explored the impact of varying these assumptions. Each patient was subject to background, all-cause mortality based on the 2004 US life tables [
We evaluated three treatment strategies on the basis of clinical guidelines for treating children and adolescents with MDD in the United States: SSRIs, CBT, and combined SSRIs and CBT [
US Preventive Services Task Force Screening and treatment for major depressive disorder in children and adolescents: US Preventive Services Task Force recommendation statement.
]. Treatment efficacy was modeled as reductions in time to full symptom relief and partial symptom relief from a depressive episode and was, for the first 36 weeks after initiating therapy, based on remission rates reported from the Treatment for Adolescents with Depression Study (TADS) [
]. The proportion of patients who had not achieved remission at the end of the 36 weeks of the TADS was 45% for patients treated with SSRIs, 36% for those treated with CBT, and 40% for those treated with combination treatment. We assumed that the rates of remission remained constant over time for each treatment cohort and estimated the points in time at which all patients were remitted by exponential extrapolation of the existing data for each of the three treatment arms (see Figure 2 in Appendix A of Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2012.03.1390). We calculated the parameter of the exponential model by using the probabilities observed at 36 weeks (panel A) but also explored the impact of an alternative approach (panel B). These population-based distributions were used to randomly draw individually based probabilities of time to remission for each treatment arm.
Among all responders to treatment in the TADS, 44.4% remitted and 55.6% failed to remit after 12 weeks of treatment [
]. We translated this increased risk of response without remission into a 25% reduction in time to partial symptom relief compared with full symptom relief for all treatment groups, but we varied this RR in sensitivity analyses.
Treatment-related suicide attempt risks
We assumed an elevated risk of suicide attempts of 70% associated with SSRI treatment on the basis of a meta-analysis of placebo-controlled trial data submitted to the Committee on Safety in Medicines in the United Kingdom [
], in which suicide attempts in youths with depression were 70% more frequent among SSRI users than among users of placebo. This estimate is very similar to the FDA meta-analysis RR of 1.6 reported in the article by Bridge et al. [
Clinical response and risk for reported suicidal ideation and suicide attempts in pediatric antidepressant treatment: a meta-analysis of randomized controlled trials.
]. By using this information, coupled with the TADS data on the RRs of suicidal events among users of SSRIs versus CBT versus combination treatment, we generated population-based estimates of suicide attempt rates for CBT and combination treatment: 35% decrease in the risk of suicide attempts for CBT and 11% decrease in the risk of suicide attempts for combination treatment relative to placebo [
All patients begin the model in a major depressive episode and initiate one of the three treatment strategies evaluated. In modeling the course of disease of an individual patient, we established the temporal order and exact timing of events. For each competing event, a random number between 0 and 1 is drawn from a uniform distribution. The inverse of the survival distribution is applied to this random number to give the age at which the current event is terminated. Subtracting the patient's current age from this age of termination yields the time to that particular event. From the sampled times, the event with the earliest time occurs next and the other times are discarded. This process is repeated after the occurrence of each new nonfatal event (i.e., remission to full or partial symptom relief, nonfatal suicide attempt, recurrence or relapse of depressive episode). By this approach, the sequence of events experienced by the patient is randomly generated according to the distributions assigned in the model. The simulation ends when the patient dies from suicide or other causes or the total time elapsed equals or exceeds the analytic time horizon of 5 years. This process was replicated for 1,000,000 patients per treatment cohort, and the outcomes, including symptom-free weeks and both fatal and nonfatal suicide attempts, were averaged across age and sex subgroups.
First (scenario 1), we assessed the internal consistency of the model projections to the empiric data at 36 weeks. Therefore, we assumed treatment-related remission and recovery rates, as well as RRs of suicidal events, over the 36-week trial period according to the TADS, supplemented with data from the published literature regarding risks of recurrence and relapse of major depressive episodes, background mortality, and fatal or nonfatal suicidal acts.
In extrapolating beyond the time horizon of empiric data (i.e., beyond the 36 weeks of TADS data), we generated several “scenario” analyses, each reflecting different assumptions about long-term effects of treatment. In the first extrapolation (scenario 2), we assumed that the suicide attempt risks associated with the treatment strategies do not extend beyond 36 weeks and that the remission rates follow natural history trends beyond 36 weeks, independent of treatment strategy. In scenario 3, treatment-related suicide attempt risks and remission rates observed in TADS over the first 36 weeks after treatment initiation were assumed to persist unabated over the 5-year time horizon for each treatment strategy.
We also explored the impact of varying several uncertain assumptions, including 1) the RR of suicide attempts associated with SSRI use; 2) the RR of suicide attempts according to the level of depression symptoms (i.e., episode vs partial symptom relief vs full symptom relief); 3) the extrapolation approach of remission rates; and 4) the proportion of responders who fail to fully remit from a given episode of MDD (i.e., those who experience partial symptom relief).
Results
Scenario 1: Analysis with TADS and additional data sources over a 36-week period
Over a 36-week period, combination treatment (SSRIs plus CBT) provided an additional 2 to 3 symptom-free weeks per patient compared with either monotherapy (Table 3). This rank ordering of treatment strategies with respect to symptom-free weeks seems to be reflective of the estimated remission rates in the early stage of treatment from the TADS (i.e., over 18 weeks of treatment, the remission rates were the highest for combination therapy [56%], followed by SSRIs [37%] and then CBT [27%] [
]). With respect to fatal and nonfatal suicide attempts, the use of SSRIs alone was associated with the highest number of events per 100,000 patients, while CBT was associated with the lowest number of events: (e.g., an additional 854 nonfatal and 16 fatal suicide attempts per 100,000 patients treated with SSRIs and an additional 83 nonfatal and 1 fatal suicide attempts per 100,000 patients treated with combination treatment compared with CBT alone). The consistency of these results with findings from the TADS showing that suicidal events were more common in patients receiving fluoxetine therapy (14.7%) than combination therapy (8.4%) or CBT (6.3%) over 36 weeks of treatment [
] demonstrates that the model works as intended (i.e., model verification or internal validation).
Table 3Impact of different assumptions regarding data extrapolation.
Scenario
Time horizon
Strategy
Symptom-free weeks (per patient)
Fatal suicide attempts (per 100,000 patients)
Nonfatal suicide attempts (per 100,000 patients)
1
36 wk
SSRIs
8.5
26
1404
CBT
7.8
10
550
COMB
10.6
11
633
2
5 y
SSRIs
110.6
140
6159
CBT
109.8
87
3706
COMB
113.2
96
4173
3
5 y
SSRIs
104.8
175
8272
CBT
106.2
65
3111
COMB
104.6
95
4362
Note. In scenario 1, analysis is over 36 wk with no data extrapolation from the TADS; in scenario 2, analysis is over 5 y and assumes that the treatment-related suicide attempt risks and remission rates do not extend beyond 36 wk; in scenario 3, analysis is over 5 y and assumes that the treatment-related suicide attempt risks and remission rates persist beyond 36 wk.
Scenarios 2 and 3: Analyses with alternative assumptions of data extrapolation past 36-week period
With respect to symptom-free weeks over a 5-year period, the two scenarios showed different results. In scenario 2, where treatment-related suicide attempt risks and remission rates did not extend beyond 36 weeks, combination therapy remained superior to both SSRIs and CBT alone, with an additional 3 weeks, on average, spent without symptoms per patient. In scenario 3, where treatment-related suicide attempt risks and remission rates persisted beyond 36 weeks, CBT was associated with the highest number of symptom-free weeks, with an additional 1 to 2 symptom-free weeks compared with SSRIs or combination treatment. With respect to fatal and nonfatal suicide attempts, the use of SSRIs alone was associated with the highest number of events in both scenarios, while CBT was associated with the lowest number of events. For example, in scenario 2, SSRI use was associated with an additional 2453 nonfatal and 53 fatal suicide attempts per 100,000 patients and combination treatment was associated with an additional 467 nonfatal and 9 fatal suicide attempts compared with CBT treatment alone.
Varying the suicide attempt risk associated with the use of SSRIs
We repeated the analysis for treatment with SSRIs under scenario 2 by using the RR values of suicide attempts that were estimated from the FDA meta-analysis (RR = 1.6 in the analysis assuming random effects, and RR = 1.9 in the analysis assuming fixed effects) (see Appendix B of Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2012.03.1390) [
Clinical response and risk for reported suicidal ideation and suicide attempts in pediatric antidepressant treatment: a meta-analysis of randomized controlled trials.
]. As expected, the differential in suicide attempts between SSRI and the other strategies increased as the relative suicide risk increased. Our model estimated an additional 722 nonfatal and 22 fatal suicide attempts for SSRIs per 100,000 patients in the scenario expected to be least favorable to SSRIs (RR = 1.9) relative to the scenario expected to be most favorable to SSRIs (RR = 1.6). Symptom-free weeks per patient remained unchanged under both scenarios.
Varying the suicide attempt risk for the level of depression
We explored two alternative scenarios of suicide attempt risk for each level of depression, one representing the upper and the other the lower bound of the 95% confidence intervals of the hazard ratios used in the base-case analysis (see Appendix C of Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2012.03.1390) [
]. Our model estimated an additional 10,400 nonfatal and 219 fatal suicide attempts per 100,000 patients with SSRI treatment by using the elevated values of suicide risk (95% confidence interval upper bound) relative to the more conservative scenario (95% confidence interval lower bound).
Varying the extrapolation approach of remission rates
We explored an alternative extrapolation approach of remission rates under scenario 3 including the last three data points (i.e., probabilities observed at 12–36 weeks rather than probabilities in the base model based on probabilities at 18 and 36 weeks) in calculating the parameter of the exponential model (see panel B of Figure 2 in Appendix A of Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2012.03.1390). Our model estimated an increase of 6 symptom-free weeks per patient for combination treatment relative to the base-case results, while there were only marginal changes in symptom-free weeks for SSRIs and CBT alone. Under this alternative scenario, combination therapy yielded the highest symptom-free weeks compared with the other treatments.
Varying the proportion of responders who failed to fully remit (i.e., those who experience partial symptom relief upon remission)
The analysis was repeated examining the impact of varying the proportion of responders who fail to fully remit. As the proportion increased, symptom-free weeks decreased (see Appendix D of Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2012.03.1390). The effect on symptom-free weeks was most prominent for lower risk ratios. For example, when increasing from RR = 1.0 and RR = 2.0, there were an additional 31 to 32 symptom-free weeks per patient over the 5-year time horizon, while between RR = 2.0 and RR = 4.0, there were also an additional 31 to 32 symptom-free weeks.
Discussion
By using a disease simulation model to integrate data from the TADS and the published literature, we found that the use of SSRIs alone was associated with the highest number of suicide-related events over a 36-week and 5-year period, while CBT was associated with the lowest number of events. The substantially higher number of suicide deaths and nonfatal attempts associated with SSRI use compared with CBT alone was, in some scenarios, somewhat offset by a marginal increase in symptom-free weeks. These results were robust when varying uncertain assumptions including suicide attempt risk for SSRIs, suicide attempt risks for level of depression, and the proportion of responders who failed to fully remit. With respect to symptom-free weeks, the combination treatment of SSRIs and CBT was superior to either monotherapy in the short-term (scenario 1) and the long-term when treatment effects were assumed to not persist (scenario 2); however, when treatment-related suicide attempt risks and remission rates persisted beyond 36 weeks (scenario 3), CBT was marginally superior to both SSRIs and combination treatment. These findings suggest that the optimal long-term treatment strategy in terms of symptom-free weeks depends on treatment effects over time, for which little available clinical data exist.
Our model-based projections are consistent with a recent meta-analysis indicating psychotherapy to be marginally superior to second-generation antidepressants in the longer-term management of depressive symptoms [
]. In the present study, the number of attempted and completed suicides for SSRIs over a 5-year period when we assumed that treatment benefits did not persist beyond 36 weeks (i.e., scenario 2) was roughly 12.60 per 1,000 patient-years (6,299 fatal and nonfatal suicide attempts over 500,000 patient-years). This rate of suicidal acts is low in comparison to that reported in a community sample of children and adolescents during SSRI treatment (28.33 per 1,000 patient-years) [
] were, by design, new users of antidepressants followed for at most 1 year after their index prescription fill, whereas subjects in our study were followed over a 5-year time horizon with rates derived from prevalent as well as incident users. Moreover, many of our subjects were on maintenance therapy for most of the 5-year period. Because index prescriptions are often prescribed at the height of a depressive crisis, and maintenance treatment is associated with lower risk periods, it is not surprising that we observed lower rates of suicidal behavior in our cohort.
The outcomes of the model are presented as point estimates only (e.g., the number of fatal suicide attempts in SSRI-treated patients in the scenario 2 analysis is 140). Because of the analytic technique of simulating individual patients in a DES, natural variability in projected outcomes arises because of random chance. In our particular analysis, we achieved reliable results when we simulated 1,000,000 individuals in each treatment arm (i.e., variance = <0.0001 across 10 model runs of 1,000,000 patients for the different outcomes).
The major strength of this study was the use of state-of-the-art methodology to evaluate the trade-off in risks and benefits of treatment strategies for depression. Decision-analytic modeling provides a framework for informed decision making under conditions of uncertainty. Specifically, it allows for exploration of the data to expressly examine “what-if” scenarios and determination of how robust the base-case results are to parameter uncertainty and changes in model assumptions. Furthermore, models facilitate projection of results beyond the time horizon of clinical trials. Also, data from future studies can be used to update the parameters and assumptions of this existing disease simulation model.
Our analysis has a number of limitations. First, treatment dynamics were not included in the model, and as in the clinical trials, we assumed that patients did not change treatment options over the 5-year period. To reflect real-world sequences of interventions, the impact of cross-over to other treatment options should be explored. This limitation is somewhat mitigated, at least in the antidepressant arm, by accumulating evidence that antidepressant augmentation and cross-over to other pharmacologic treatment options do not produce differential outcomes [
]. Moreover, our estimates of treatment effect are based on intention-to-treat analysis of the TADS, mitigating concerns about cross-over. Second, because of limited data, simplifying assumptions were made with regard to the model structure, and some transitions were not allowed in the model. For example, we did not allow patients to move directly from partial symptom relief to full symptom relief. One consequence of this assumption is that patients in partial remission can transition only to death, suicide attempt, or relapse into a depressive episode. While in partial remission, patients remain at a higher risk of suicide attempts (hazard ratio = 4.20). Omitting this transition could therefore imply an overestimation of the time in partial symptom relief, which would result in an overestimation of suicide attempts. Because RCTs are too small to observe rare events such as death by suicide, we estimated the incidence of suicide death by applying the population-based case-fatality rate for suicide attempts among the age-/sex-matched US general population (e.g., 1 of 206 suicide attempts among 15-year-old girls are fatal) [
]. Third, we do not address how to value the risks and benefits that are explicitly estimated in the analysis. Expressing these outcomes by using measures such as quality-adjusted life-years, commonly employed in cost-effectiveness analysis, could facilitate balancing these risks and benefits by using a common metric for informed decision making.
To our knowledge, no previous study has investigated the risks and benefits of clinically relevant treatment strategies for pediatric MDD beyond the time horizon of available clinical data. By using a DES model, we leveraged multiple sources of data, demonstrated the consistency of model projections over a 36-week period compared with the TADS data, and extended the evaluation of relevant strategies to 5 years. We also provide estimates for the risk of completed suicides, which clinical trials have not been able to measure given the short study periods and small sample sizes. The limitations of our study notwithstanding, our findings suggest that CBT alone offers a safer profile with respect to death by suicide and nonfatal attempts compared with combination treatment or SSRIs alone. Any additional benefits of SSRIs, either alone or in combination with CBT, must be weighed against the expected increase in suicides.
Future research should consider the costs associated with the burden of depression, treatment interventions, and nonfatal suicide attempts to determine the cost-effectiveness of depression treatments, which may be an important consideration by policymakers in identifying the optimal approach to managing depression in children and young adults.
Source of financial support: Department of Health Policy and Management, Harvard School of Public Health, Boston, MA.
Clinical response and risk for reported suicidal ideation and suicide attempts in pediatric antidepressant treatment: a meta-analysis of randomized controlled trials.
The use of modelling to evaluate new drugs for patients with a chronic condition: the case of antibodies against tumour necrosis factor in rheumatoid arthritis.