Cost-Effective Drug Switch Options After Unsuccessful Treatment With an SSRI for Depression

The STAR*D trial design and protocol have been previously described in detail (9,12). STAR*D was a multilevel trial, and the patients could move to the next level or a 12-month naturalistic follow-up phase (wherein they continued the indicated treatment) depending on their therapeutic response and side effects. In level 2, 727 patients were randomly assigned to one of the drug switch options (bupropion, sertraline, or venlafaxine) for depression treatment.

Calculated costs were based on three components: antidepressant medications, other medications, and all health care facility utilization. The antidepressants included the assigned study medication. Other medications included antidepressants other than the assigned treatment that the patient might have used during level 2 and concomitant medications required to manage side effects of antidepressants. The concomitant medications included trazodone, anxiolytics, sedatives, laxatives and antiemetics, and sexual dysfunction medications. Utilization of all medication types was based on total dose intake, which depended on the specific dose and duration of use of each medication. For antidepressant medications, doses were obtained from the trial data. For concomitant medications, doses were not entered in the study and were determined by using the middle value of the ranges recommended in the UpToDate database (13). The duration of medication use was determined by using the recorded start and end dates. [The algorithm used to determine the recorded start and end dates is available as an online supplement to this article.]

Health care facility utilization included outpatient visits, emergency room (ER) visits, and hospitalizations that occurred during level 2. Health care facility utilization data for the course of level 2 were collected by using an interactive voice response (IVR) system. The IVR system was implemented by telephone and involved a script of recorded instructions followed by a set of questions. The patients responded to the questions by pressing the appropriate number on their phone keypad. The outpatient visits after baseline of level 2 (as specified by the protocol) were also considered. The product of the determined utilization and unit costs estimated the costs associated with health care facility utilization.

The total costs for each treatment were calculated as the sum of costs of study medication, other antidepressants, concomitant medications, and health care facility utilization during level 2. All medication unit costs were obtained from RED BOOK (14), and health care facility costs were obtained from the Physician Fee Schedule (15) or the Healthcare Cost and Utilization Project (16) databases. [A table listing the means and standard deviations of unit costs along with the information source is available in the online supplement.] To reflect the variability and skewed nature of cost data obtained from these national databases, they were modeled by using a gamma distribution (17,18). All costs in the study were discounted at a 3% rate to year 2014.

All missing cost information for health care facility utilization was handled by using multiple imputation methods. The missingness of IVR data was not significantly different among the three treatment arms; we therefore considered it reasonable to assume the missing-at-random mechanism for our analyses. The Markov chain Monte Carlo method (19) was used to impute missing patient baseline information. Also, no assumptions were made about the distribution of costs in our study sample, and nonparametric methods were used. A nonparametric propensity score multiple imputation method was used to impute missing IVR cost information (20). This method generates a propensity score to indicate the probability of IVR missingness. The observations are then grouped based on these propensity scores, and Bayesian bootstrap imputation is applied within each group. Existing IVR cost information and patient characteristics were used to impute missing IVR cost data. Costs were compared among the groups by using Kruskal-Wallis tests. If the costs among groups were significantly different, pairwise comparisons were made by using the Bonferonni adjustment. The p values for the adjustment were calculated by multiplying the number of comparisons (N=3) with the raw p value.

Response and remission were used as measures of treatment effectiveness in this study. Remission was defined as a total score of ≤5 on the QIDS–Self-Report (QIDS-SR-16) at study exit. Response was defined as a reduction of ≥50% on the QIDS-SR-16 at level-2 exit compared with scores at level-2 baseline.

The cost-effectiveness analyses were carried out from the perspective of government as the payer. Costs included the medication costs, outpatient and ER visit costs, and costs for bed days due to hospitalizations. No indirect costs were included. The cost-effectiveness analyses were conducted by using the net health benefits (NHBs) framework (21). The NHBs were calculated by using the following formula:

In this formula, µ_Ei and µ_Ci are the average effectiveness and costs, respectively, for treatment (i), and λ is the amount of money for which there is a willingness to pay per unit of effectiveness (“willingness-to-pay”). The treatment option with the highest NHB is considered to be the cost-effective option. For our analyses, we assumed the willingness-to-pay to be $30,000 for remission or response over the duration of level 2. Linear regression models were utilized to estimate the effect of treatment on NHBs (22,23). We did not adjust for the time spent during level 2 because there was no significant difference in the treatment duration among the switch therapies (9). Further, a sensitivity analysis varied the willingness-to-pay from $10,000 to $50,000 per effectiveness outcome (remission or response). Stochastic analysis was performed by using a bootstrapping method, which involved resampling costs and effects 1,000 times randomly with replacement (in a joint manner) to obtain estimates of NHBs variability for each treatment option. The bootstrapped sample was used to determine estimates of the average NHB and percentile interval of NHBs (2.5%–97.5%) in the study.

In level 2 of STAR*D, 727 patients were randomly assigned to a switch drug treatment after initial unsuccessful treatment with citalopram; 239 (33%) were randomly assigned to bupropion, 238 (33%) to sertraline, and 250 (34%) to venlafaxine. Overall, the average time (mean±SD) in level 2 was 8.9±5.03 weeks. At the end of level 2, 227 (31%) patients moved to level 3, 257 (35%) entered the follow-up phase, and 243 (33%) terminated from the study. The time spent in level 2 and in the subsequent actions was not significantly different among treatments [see online supplement].

This study demonstrated that the switch options of bupropion, sertraline, and venlafaxine in level 2 of the STAR*D project were not significantly different in terms of cost-effectiveness. The total costs among the three groups were significantly different; however, none of the pairwise differences in overall costs for the three drugs attained statistical significance. This lack of significance can be partially attributed to the large variability in costs. This study reemphasizes that overall treatment costs need to be considered rather than drug costs alone in order to estimate the costs incurred or cost-effectiveness within a clinical trial.

The literature on economic analyses of second-line treatment options for depression is sparse. Two cost analysis studies used administrative databases to compare second-line therapies for major depressive disorder (24,25). However, these were observational studies with biased medication prescribing patterns. In contrast, STAR*D was a clinical trial in which patients were assigned randomly to one of the investigated medications during level 2.

Two other studies used computational models to evaluate second-line treatment options for major depressive disorder (26,27). However, neither used generalizable cost estimates. In our analyses, we used cost estimates from national databases, such as the Healthcare Utilization Project (16) and the Physician Fee Schedule (15). To the best of our knowledge, no previous study has simultaneously compared the cost-effectiveness of the three switch options (bupropion-SR, sertraline, and venlafaxine-XR) after initial SSRI treatment while appropriately representing the uncertainty of the estimates.

Our study had some limitations. The health care services utilization information was collected by the IVR system and resulted in missing IVR data for 68.5% of participants. The IVR system has been associated with greater patient convenience, fewer transcription errors, and lower costs compared with personal interviews (28). In the STAR*D trial, the IVR responses had an intraclass correlation of .68 compared with actual medical records (29). We employed multiple imputations to minimize bias due to missingness and to appropriately represent the uncertainty of the estimates (30). The high variance of costs in our study reflects the uncertainty in the estimated costs. Also, our analyses focused on cost-effectiveness from the perspective of government as a payer of direct costs and only during level 2 of the STAR*D study. A different perspective may yield different conclusions. For example, an individual patient could be more concerned about his or her out-of-pocket expenses than about the overall costs incurred, when considered jointly with effectiveness. Willingness-to-pay is another factor of uncertainty in our analyses; there are no recommendations for using this measure when assessing effectiveness in terms of response and remission (31). To overcome this limitation, we systematically varied willingness-to-pay over a wide range of values on the basis of a discussion with researchers and experts. Overall, STAR*D included a broad population and was designed to closely reflect real practice (12); thus the cost-effectiveness results of this study are highly generalizable.

Our results show that costs of the study medications differed significantly. However, there were no significant differences in the pairwise comparisons of total costs and cost-effectiveness of the three medications. Thus we concluded that after considering costs and cost-effectiveness, there is no rationale to change the conclusions that are based on therapeutic effectiveness. Other factors, such as clinicians’ preference, family history, or treatment of most evident cardinal symptoms, could lead to a choice of one antidepressant over another. Also, given the high uncertainty and variability associated with costs and effects, it is only appropriate to consider them together by using a cost-effectiveness framework.