Hierarchical endpoints analyzed using the win‐ratio method as a practical innovation for deprescribing trials

Deprescribing, commonly defined as the systematic process of discontinuing drugs when existing/potential harms outweigh existing/potential benefits,1 has emerged as a potential intervention to improve outcomes in older adults with multimorbidity. Meta-analyses indicate that deprescribing may improve quality of life, functional status, and patient and caregiver burden.2, 3 These meta-analyses have been important to better understand the impact of deprescribing on clinical outcomes, in large part because individual randomized clinical trials (RCTs) of deprescribing have often lacked sufficient power to determine clinical benefit. Indeed, many deprescribing studies have been powered to examine whether the intervention can effectively reduce total medication count, and have used surrogate markers like potentially inappropriate medications and drug burden index. While there are larger observational studies powered for clinical outcomes such as mortality and hospitalizations, inferences from such studies are substantially limited by confounding by indication (i.e., deprescribing is often due to an intolerance and/or contraindication that may be related to the outcome of interest through a known [or unknown] causal pathway). The challenge of determining the single best primary outcome for deprescribing trials has also been well-described.4 Herein, we propose a new methodological approach for deprescribing RCTs that can facilitate study of important clinical outcomes through the use of a hierarchical composite endpoint and a statistical approach called the win-ratio method. The win-ratio approach is an analytic strategy that integrates patient-endorsed clinical priority among components of a composite hierarchical outcome. In this case, the hierarchy frequently covers multiple fundamental domains of clinical care such as (1) death, (2) morbidity (i.e., hospitalizations), and (3) patient reported outcomes (PROs).5, 6 This approach has recently been embraced by cardiology, and been used as the primary endpoint in multiple recent clinical trials including both novel drugs7, 8 and devices.9 To obtain the win-ratio, each intervention participant is compared pairwise with each control participant, and a “win,” “loss” or “tie” is determined for each pair. The ratio of the probability of a win to that of a loss for the intervention is defined as the win-ratio. The pairwise comparison follows a pre-specified hierarchy—head-to-head comparisons start with the outcome at the top of the hierarchy; if a win occurs, then this outcome contributes to the subsequent win-ratio calculation; if a tie occurs, the head-to-head comparison continues for the next outcome in the pre-specified hierarchy, and so on. The win-ratio is estimated as the number (or proportion) of wins in the intervention group compared to wins in the control group. The Finkelstein-Schoenfeld method is a variation of this win-ratio method5 that uses a non-parametric score test on sum of scores (+1 for a win, 0 for a tie, −1 for a loss) to estimate the win-ratio. A win-ratio of 1.25 means that a participant in the intervention arm is 25% more likely to win on the composite endpoint than a participant in the control arm; in other words, the likelihood that the winner is the intervention group rather than the control group (for any untied pair of patients) is 1.25. Figure 1 demonstrates an illustrative example of the win ratio method applied to a study comparing treatments A and B with a hierarchical composite endpoint of (1) time to death, (2) time to hospitalization, and (3) improvement in quality of life (QOL). Pairings are first compared for death; those who had longer time to death are determined to Win; if there is a tie (same duration to death, or neither died), the comparison moves to the next level of the hierarchy (hospitalizations). Those who had longer time to hospitalization are determined to Win; if there is a tie (same duration to hospitalization or neither experienced a hospitalization), the comparison moves to the next level of the hierarchy (change in QOL). Those who had greater improvement in QOL (depending on pre-specified threshold), are determined to Win; if there is a tie at this last outcome of the hierarchy, the pair is considered to have tied. Each pair accordingly contributes only 1 win (or 0 wins in the case of a tie) to the subsequent calculation of the win-ratio. The number of total wins versus total losses are then compared; the number of ties does not contribute to either the numerator or denominator of the win-ratio. Of note, a loss for treatment A is a win for treatment B. As shown in the figure, Treatment A was determined to win based on the death outcome in 2.2% of the pairwise comparisons, based on the hospitalization outcome in 11.5% of the pairwise comparisons, and based the QOL outcome in 36.3% of the pairwise comparisons; while Treatment B was determined to win based on the death outcome in 1.9% of the pairwise comparisons, based on the hospitalization outcome in 8.8% of the pairwise comparisons, and based the QOL outcome in 29.3% of the pairwise comparisons. Among all pairwise comparisons, 10% tied on all outcomes. The win-ratio for Intervention A is then calculated based on the Win percentage of 50% (cumulative sum of wins from each hierarchical outcome) for Intervention A divided by the Win percentage of 40% for Intervention B. This generates a Win-Ratio of 1.25, indicating that for any untied pair of patients, the likelihood that the winner is the intervention group (rather than the control group) is 1.25. Using a hierarchical composite endpoint (and the win-ratio) has several important advantages relevant to deprescribing research in older adults. First, incorporating multiple parameters into a single composite measure can improve statistical power to detect an intervention effect for even a modest study sample size. For example, recent pharmacologic studies for patients with cardiac amyloidosis were able to provide meaningful statistical inferences on clinical endpoints with a sample size of just 441 patients10—in fact, this study ultimately led the FDA-approval of the first pharmacologic therapy for transthyretin cardiac amyloidosis. Second, a hierarchical composite endpoint can account for competing risks of mortality. Mortality as a competing risk is particularly problematic for studies focused on QOL, which is an outcome of interest for many deprescribing interventions. This is because post-intervention data on QOL cannot be obtained for participants who died during the study. A hierarchical composite endpoint can incorporate QOL and mortality into a single endpoint, eliminating the concern for biased estimates that result from missing (or imputed) data for those who died. This is important and could have broader implications beyond deprescribing studies given the emergence of QOL as an important patient-centered goal of therapy. Third, a hierarchical composite endpoint can incorporate both safety and efficacy, where safety is prioritized. In many situations, the benefit of deprescribing may stem from improvements in QOL or reductions in treatment burden; many patients might prefer this, but do not want this at the expense of experiencing a clinical event like a hospitalization or death. A hierarchical endpoint that incorporates clinical events and other patient-centered outcomes can facilitate a study that prioritizes safety (i.e., survival, hospitalization), while also studying multiple other relevant outcomes (i.e., quality of life). In other words, a hierarchical composite endpoint can first prioritize “no harm,” followed by an assessment other outcomes important to patients and relevant to deprescribing. There are several potential limitations that have emerged in the literature related to use of hierarchical composite endpoints and the win-ratio.11 However, in our view, many of these limitations are less problematic for deprescribing RCTs. First, critics have been concerned about how to prioritize individual components of the hierarchical composite endpoint. For deprescribing RCTs, it is logical to prioritize safety endpoints (i.e., survival) over efficacy endpoints (i.e., measures of morbidity), staying consistent with the medical tenet of first do no harm. When using multiple efficacy endpoints, it may be prudent to solicit feedback directly from patients with the condition of interest to facilitate a patient-centered approach to developing the hierarchy. Second, critics have called attention to the fact that the timescale of expected benefit (or potential harm) may vary for the individual components of the hierarchical endpoints. For example, changes in symptoms may occur sooner than the accrual of deaths. If the study duration is relatively short, endpoints that change more quickly may impact the overall study result more than endpoints that take more time to accrue (such as mortality). For deprescribing studies, this may not necessarily be a disadvantage, since symptom burden, quality of life, and functional status are often high priority outcomes in deprescribing trials. Third, critics have shared concern about inferring benefit from positive studies when the benefit of the intervention was largely driven by less-clinically relevant outcomes. This critique is primarily focused on studies that incorporate surrogate measures like biomarkers which do not have direct clinical significance and may even diverge from the directionality of effect on clinical events. This is an important point but is less relevant for deprescribing RCTs that are focused on patient-centered outcomes. Patient centered outcomes such as QOL may be just as important as traditional clinical events from the patient perspective. The relevance of QOL, especially among older adults (the population in whom most deprescribing studies have been focused on to date) is well-illustrated by data indicating that some older adults prioritize QOL even above longevity.12-14 The use of hierarchical composite outcomes analyzed by the win-ratio method has the potential to advance the field of deprescribing by facilitating the conduct of studies focused on the impact of deprescribing on relevant clinical outcomes including patient-reported outcomes. This innovative approach is patient-centric and makes it feasible to design and conduct pragmatic deprescribing RCTs that are adequately powered to answer important clinical questions. The future of deprescribing is here—time will tell on whether this methodological innovation will be a catalyst for more Wins for the field. “The use of hierarchical composite outcomes analyzed by the win-ratio method has the potential to advance the field of deprescribing by facilitating the conduct of studies focused on the impact of deprescribing on relevant clinical outcomes including patient-reported outcomes.” Study concept and design: Parag Goyal and Andrew P. Ambrosy. Acquisition, analysis, or interpretation of data: N/A. Drafting of the manuscript: Parag Goyal and Andrew P. Ambrosy. Critical revision of the manuscript for important intellectual content: All authors. Statistical analysis: N/A. Administrative, technical, or material support: Andrew P. Ambrosy. Study supervision: Andrew P. Ambrosy. The authors have nothing to report. Dr. Goyal has received consulting fees from Agepha Pharma, Akros Pharma, Axon therapies, Bayer HealthCare Pharmaceuticals, and Sensorum Health; and has received personal fees for medicolegal consulting and expert testimony related to heart failure. Dr. Ambrosy has also received relevant research support through grants to his institution from The Permanente Medical Group, Northern California Community Benefits Programs, Garfield Memorial Fund, Abbott Laboratories, Amarin Pharma, Inc., Bayer, Cordio Medical, Edwards Lifesciences LLC, Esperion Therapeutics, Inc., Merck, and Novartis. Dr. Steinman has received royalties from UptoDate and honoraria from the American Geriatrics Society. None. Dr. Goyal is supported by National Institute on Aging (NIA) grants K76AG064428, R21AG077092, R01AG085420, and R01AG088522; Dr. Steinman is supported by NIA grants R33AG086944, K24AG049057, P30AG044281; Dr. Ambrosy is supported by National Heart, Lung, and Blood Institute grant K23HL150159 and a 2nd Century Early Faculty Independence Award from the American Heart Association.