Randomized Clinical Trials Using a Hierarchical Composite Primary End Point

Key Points

Question

How are hierarchical composite end points (HCEs) currently being used and analyzed as primary end points in randomized clinical trials across medical disciplines?

Findings

This scoping review of 92 randomized clinical trials (with 79 567 planned participants) using an HCE as a primary end point showed substantial heterogeneity and shortcomings in the planning and analysis of HCEs, as well as the reporting of the hierarchy construction, sample size assumptions, and handling of ties.

Meaning

These findings suggest that the current use of HCEs in randomized clinical trials lacks methodological consistency and transparent reporting, limiting interpretability and comparability across studies and underscoring the need for clear methodological and reporting guidelines.

This scoping review of randomized clinical trials using hierarchical composite end points examines characteristics of their use, including how they are constructed, analyzed, and reported.

Abstract

Importance

Hierarchical composite end points (HCEs) are a promising tool used in randomized clinical trials (RCTs) to integrate multiple outcomes of varying clinical relevance into a single measure.

Objective

To describe how often HCEs are used as primary outcomes in RCTs, and how they are constructed, analyzed, and reported.

Evidence Review

MEDLINE, Embase, CENTRAL, and Web of Science were searched on December 9, 2024, complemented by a forward citation search of methodological papers on HCEs. RCTs that used an HCE as their primary outcome were included, defined either by self-declaration (hierarchical composite or outcome ranking) or through use of an HCE-specific analytical approach (win ratio, win odds, probabilistic index, or generalized pairwise comparison). Pilot studies, post hoc analyses, and hierarchically tested coprimary end points were excluded. Data were independently screened and extracted in duplicate. Trial characteristics, end point composition, hierarchy justification, and analytical methods were summarized descriptively.

Findings

Among 5188 screened records, 92 RCTs were included, with 79 567 planned participants. The use of an HCE as a primary end point has increased, with 72 trials (78.3%) initiating recruitment within the past decade. Most RCTs were drug trials (43.5% [40 of 92]), in cardiology (43.5% [40 of 92]), multicenter (91.3% [84 of 92]), and non–industry sponsored (67.4% [62 of 92]). The 92 HCEs had a median (IQR) of 4 (3-5) components and the highest ranked component was usually mortality (80.4% [74 of 92]). The last hierarchical component was most often a continuous component (64.1% [59 of 92]). The majority of RCTs did not report any information on how the hierarchy was established (82.2% [60 of 73]; excluding RCTs where only information from registries was available). Generalized pairwise comparison was the most frequent analysis approach (57.3% [26 of 45]) among the 45 published RCTs, yet no standardized way for presenting results was observed.

Conclusions and Relevance

This scoping systematic review of 92 RCTs using HCEs found that their use has increased across medical fields, but their construction, analytical approaches, and reporting of results remained highly heterogeneous. By systematically mapping how HCEs are currently implemented, further review is essential for the development of much needed methodological and reporting standards.

Introduction

While randomized clinical trials (RCTs) are the criterion standard to assess the efficacy of new medical interventions, their conduct is often time consuming and expensive.^1,2,3,4,5 In cases where the events of interest are rare and when treatment effect sizes are likely to be modest, large sample sizes are required. This may not always be feasible due to limited resources.^6,7 To facilitate adequately powered RCTs capable of detecting small, but clinically important differences, many trials adopt composite end points, combining multiple individual outcomes into a single metric.⁸ This strategy increases event frequency and reduces required sample sizes. However, conventional composite end points have a significant limitation: they typically treat all component events equally, regardless of their relative clinical importance. This may obscure important changes in higher priority outcomes. For instance, when considering a composite end point of mortality and hospitalization, an intervention that substantially reduces mortality may appear to have limited benefit if survivors experience more nonfatal events such as rehospitalization.^9,10,11 Hierarchical composite end points (HCEs) are designed to address this limitation by ranking 2 or more outcome components by clinical importance (eg, mortality is ranked of higher importance than hospitalization).

The concept of generalized pairwise comparisons, which is often used to analyze HCEs, was described in detail by Finkelstein and Schoenfeld in 1999.^12,13 Building upon this, Pocock and colleagues proposed the win ratio in 2011.¹⁴ In these methods, all participants in the intervention group are compared pairwise with control participants to assess whether the intervention group has an overall better outcome. Winners and losers are determined starting from the highest-priority component, with ties carried forward to the next component (unresolved ties remain possible on the lowest priority component).¹⁵ Using this concept, multiple components can be integrated in a hierarchical outcome assessment, which might be more reflective of clinical decision-making.¹⁶

The win ratio and other related HCE analytic methods (eg, win odds, probabilistic index) have been increasingly used in cardiovascular trials¹⁷ as well as in infectious disease RCTs where HCEs are often labeled as desirability of outcome ranking (DOOR).^16,18,19 A systematic review focusing specifically on DOOR end points in infectious disease studies has been conducted,¹⁵ but found only 2 infectious disease RCTs using DOOR-based HCEs as primary outcomes, one of them a pilot trial.^20,21

Despite growing interest in HCEs, key questions remain unanswered, including how to define, justify, and analyze these end points, how prevalent these hierarchical end points are in different medical areas, what terminology is used for their description, what processes are used to justify their ranking, and how sample size calculations and results should be reported. We therefore conducted a scoping review to identify all RCTs which used an HCE as their primary end point to provide an overview of current research practice.

Methods

This scoping review is reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guideline extension for scoping reviews.^22,23 A detailed protocol was prospectively registered on Open Science Framework.²⁴ No ethical approval was obtained as this scoping review used only data from publicly available sources and did not involve the additional collection or use of primary data from human participants.

Systematic Search and Eligibility Criteria

An initial pilot search was conducted in Embase using generic terms for HCE to identify potential pertinent vocabulary. Following this, an information specialist composed the final search for RCTs (covering published studies, protocols and trial registries) using an HCE as a primary outcome. The systematic search was conducted on December 9, 2024, in Embase, Medline (Ovid), Web of Science Core Collection, and the Cochrane Central Register of Controlled Trials (CENTRAL) (detailed search strategies in eAppendix 1 in Supplement 1). Records were identified and exported to EndNote 21 (Clarivate) and deduplicated using the Automated Systematic Search Deduplicator.²⁵

References were uploaded into Covidence and screened by titles and abstracts, followed by full-text screening in duplicate by 2 independent reviewers (S.E., A.J.dJ., Y.M., M.E.A.dK., H.J., N.H.K., M.S., N.S.B., N.R., S.W.X.O., and B.S.). Discrepancies were resolved by discussion, or, if no consensus could be reached, by consulting a third reviewer (S.E., S.W.X.O., M.B., and B.S.).

Trials were included independently of the trial status (ie, we also included ongoing trials based on their trial registration or protocol) if they used an HCE as their primary outcome. HCEs were defined as end points that combine multiple distinct clinical components into a single outcome, with a specified ranked or prioritized order for evaluating these components. Because differentiating between an HCE and an end point based on an ordinal scale can be challenging, we used the following definition for operational purposes: the primary composite end point was considered hierarchical if it was either (1) explicitly self-declared as a “hierarchical” end point or as an “outcome ranking” by the authors, or (2) if it was analyzed using a method specific for an HCE (eg, win ratio, win odds, probabilistic index, generalized pairwise comparisons). Studies were excluded if they were pilot or feasibility studies, were post hoc or secondary analyses, used only hierarchical secondary end points, used hierarchically tested coprimary end points, or used a hierarchy that was primarily developed to account for missing data or hierarchical testing to account for multiple testing. No restrictions on language or publication date were applied. From the included RCTs, we extracted cited methodological papers on HCEs, deduplicated these records, and conducted forward citation searches to identify additional potentially eligible trials.

Data Extraction

Data extraction was performed in duplicate by 2 researchers (S.E., A.J.dJ., Y.M., M.E.A.dK., H.J., N.H.K., M.S., N.S.B., N.R., S.W.X.O., and B.S.) using REDCap^26,27 (agreement 88% on extracted variables). Conflicts were consolidated by a third reviewer (S.E., S.W.X.O., M.B., and B.S.) and discussed with the 2 researchers who performed extraction. We extracted detailed trial characteristics using a standardized data extraction form (eAppendix 2 in Supplement 1). Extracted information included trial identifiers, trial status, available sources of information (ie, publications, protocols, registry, other), patient population, medical field, intervention, sponsorship, planned and achieved sample size, study design, masking status of care provider, patients and outcome assessors, dates of first patient inclusion and publication of results, and details of the hierarchical composite primary end point (including number and type of components, process used to establish the hierarchy, reported strengths and limitations of HCEs, handling of continuous outcomes, tie-breaking methods, statistical analysis methods, and reporting of overall summary measures). Sample size assumptions and justification for expected improvements were also collected.

Statistical Analysis

We summarized the characteristics of the included RCTs descriptively using frequencies and percentages for categorical variables. When reporting proportions, we included only RCTs in the total denominator for which a result could have been expected (eg, when assessing reported limitations of HCE, we excluded RCTs for which we only had information from trial registries). Data cleaning, coding, and analysis were conducted in R version 4.5.1. (R Foundation for Statistical Computing).

Results

A total of 5188 records were screened (4200 from the initial literature search, plus 988 from the forward citation search). The 175 eligible identified records using an HCE as primary outcome were allocated to a total of 92 unique randomized clinical trials (RCTs could have multiple records such as results publication, published protocol, and registry record) (Table 1; flowchart available in the eFigure and overview of individual studies in eTable 1 in Supplement 1). These RCTs had an estimated total of 79 567 planned participants. Most included trials were drug trials (40 of 92 [43.5%]), conducted in the fields of cardiology (40 of 92 [43.5%]), were non–industry sponsored (62 of 92 [67.4%]), and conducted in multiple centers (84 of 92 [91.3%]) (stratified by sponsor type in eTable 2 in Supplement 1). The Figure shows the number of trials by medical field initiating patient recruitment each year, from the earliest included trial starting recruitment in 1975 to the most recent in 2025. While cardiology remained the dominant field, there was an increase in trials related to other medical fields (eg, infectious diseases), alongside a strong overall increase in the use of HCEs in the past decade (Figure). A total of 47 out of 92 included trials were completed (51.1%), with 45 having results published; 34 trials (37.0%) were recruiting or had ongoing follow-up, 6 trials (6.5%) were prematurely discontinued, and 4 trials (4.3%) had not started recruitment yet.

Table 1. General Characteristics of Included Randomized Clinical Trials (RCTs) Using a Hierarchical Composite End Point as Primary Outcome.

Characteristic	Total RCTs, No (%) (N = 92)
Planned sample size, median (IQR)	400 (240-750)
Sponsor
Nonindustry	62 (67.4)
Industry	30 (32.6)
Medical field
Cardiology	40 (43.5)
Infectious diseases	10 (10.9)
Intensive care	9 (9.8)
Neurology	7 (7.6)
Othera	26 (28.3)
Trial status
Completed	47 (51.1)
Recruiting or ongoing follow-up	34 (37.0)
Premature discontinuation	6 (6.5)
Not yet recruiting	4 (4.3)
Unclear	1 (1.1)
Type of intervention
Drug	40 (43.5)
Health service delivery including diagnostics	15 (16.3)
Medical device	13 (14.1)
Surgical or invasive procedure	10 (10.9)
Otherb	14 (15.2)
No. of centers
Multicenter	84 (91.3)
Single center	8 (8.7)
International vs national
International	47 (51.1)
National	45 (48.9)
Design
Parallel group RCT	86 (93.5)
Crossover RCT	2 (2.2)
Factorial design	2 (2.2)
Otherc	2 (2.2)
Randomization unit
Individual	89 (96.7)
Cluster	3 (3.3)
Treatment groups
2	81 (88)
3	6 (6.5)
4	5 (5.4)
Trial objective
Superiority	90 (97.8)
Noninferiority	2 (2.2)
Patient masking
No	55 (59.8)
Yes	37 (40.2)
Physician or health care practitioner masking
No	59 (64.1)
Yes	33 (35.9)
Outcome assessor masking
No	30 (32.6)
Yes	52 (56.5)
Not reported	10 (10.9)

^a Total of 26 trials, including: 4 orthopedics; 3 gastroenterology; 3 cardiothoracic surgery; 2 pneumology; 2 nephrology; 2 oncology; 2 emergency medicine; 2 angiology; and 1 each of rheumatology, hematology, endocrinology, pediatrics, urology, and geriatrics.

^b Total of 14 trials, including: 4 biological interventions; 3 airway management; 2 behavioral, lifestyle, or psychological intervention; 2 mixed interventions (drug and procedure, behavioral and drug); and 1 each radiation, complementary therapy, or pharyngeal electrical stimulation.

^c Two trials: ENABLE was 2 separate parallel group RCTs analyzed combined, and LIVE-SMART sequential randomization.

Figure. Frequency of Recruitment Start of Randomized Clinical Trials Using a Hierarchical Composite Primary End Point.

^aA total of 26 trials, including: 4 orthopedics; 3 gastroenterology; 3 cardiothoracic surgery; 2 pneumology; 2 nephrology; 2 oncology; 2 emergency medicine; 2 angiology; and 1 each of rheumatology, hematology, endocrinology, pediatrics, urology, and geriatrics. For 4 trials the anticipated recruitment data was used; 1 did not report start date (2 years prior to publication was estimated).

Domains of the HCEs

The highest ranked component of the 92 HCEs was mortality for 74 trials (80.4%) (Table 2). The remaining 19.6% (18 of 92) could be grouped into the following 4 categories: disease-specific clinical outcome events (6 events; eg, asthma exacerbations), functional or recovery (6 events; eg, return to work, fracture union or clinical recovery), treatment failure or necessity of therapy escalation (4 events), and safety or adverse events (2 events).

Table 2. Characteristics of Components Included in Hierarchical Composite End Points.

Characteristic	RCTs, No (%) (N = 92)
Highest ranked HCE component
Mortalitya	74 (80.4)
Disease specific clinical outcomeb	6 (6.5)
Functional or recoveryc	6 (6.5)
Treatment failure or necessity of therapy escalationd	4 (4.3)
Safety or adverse eventse	2 (2.2)
Handling of ties when both had an event on the highest ranked componentf
Taking into account lower component	21 (22.8)
Time to event taken into account	22 (23.9)
Counted as tie	19 (20.7)
Nonparametric rank methods (Wilcoxon or Mann-Whitney)	3 (3.3)
Unclear or not mentioned	35 (38.0)g
Lowest ranked HCE component
Continuous	59 (64.1)
Binary	21 (22.8)
Time-to-event	4 (4.3)
Otherh	8 (8.7)
Handling continuous componentsf
Direct comparison of the values	73 (71.6)
Predefined minimum threshold difference	11 (10.8)
Unclear	17 (16.7)
Other method usedi	1 (1.1)
No. of components, median (IQR)	4 (3-5)
No. of components per HCE
2	22 (23.9)
3	22 (23.9)
4	23 (25.0)
5	12 (13.0)
6	4 (4.3)
7	9 (9.8)

Abbreviations: HCE, hierarchical composite end point; RCT, randomized clinical trial.

^a Including 53 with all-cause mortality, 8 time to death, 7 cardiovascular death, 2 hospital mortality, 2 fatal perioperative or periprocedural bleeding, 1 mortality due to liver failure, and 1 death due to thrombosis.

^b Including asthma exacerbation, change in seizure frequency, major amputation, total rate of heart failure admissions, and swallowing safety.

^c Including clinical outcome, radiographic fracture union, efficacy plus safety, return to work, clinical, recovery, and adequate clinical response.

^d Including necessity of intensive care unit care during hospitalization and failure of noninvasive ventilation.

^e Including safety assessment and serious adverse events.

^f Multiple answers possible from a single hierarchical composite end point.

^g For 17 of these studies (48.6%), information was only available from registries.

^h Including 3 categorical, 3 ordinal, and 2 unclear.

ⁱ Model-based approach.

The lowest ranked component of the HCE was continuous in 59 trials (64.1%) and binary in 21 trials (22.8%). The 92 HCEs had a total of 345 components, with a median (IQR) of 4 (3-5) components per HCE (range, 2-7) (Table 2). Of all 345 components, 184 components (53.3%) were binary, 95 (27.5%) continuous, 41 (11.9%) time-to-event, 11 (3.2%) were ordinal or ranking, and 11 (3.2%) were categorical (eTable 3 in Supplement 1). Handling of tie breaks when both pairs had an event on the highest ranked component of the HCE (eg, both patients died) was either not reported (35 of 92 [38.0%]) or handled by taking into account the lower hierarchy component (21 of 92 [22.8%]), time-to-event (22 of 92 [23.9%]), or counted as ties (19 of 92 [20.7%]); multiple approaches could be used within a single trial.

Terminology, Development, and Rationale

The terminology “hierarchical composite end point” or “hierarchical composite outcome” was used in 62 of 92 identified trials (67.4%). Other terms used were “rank composite” (12 of 92 [13.0%]), and DOOR (10 of 92 [10.9%]) (Table 3). DOOR end points were primarily used by infectious disease trials and RCTs in neurology (key characteristics by medical field available in eTable 4 of Supplement 1). In more than 80% of the RCTs, authors did not provide any explanation of how they developed or selected the hierarchical order (60 of 73 [82.2%]; excluding 19 RCTs for which only information from a clinical trials registry was available). Six trials (8.2%) cited a reference to an established hierarchy, and for 2 trials each (2.7%) consulting patient representatives or using an expert committee to define the hierarchy were reported (Table 3). A justification for the selection of an HCE was reported in 39 of 73 RCTs (53.4%). The top 3 reported reasons were reporting general advantages compared with traditional end points (26 of 73 [35.6%]), that an HCE better reflects overall health (24 of 73 [32.9%]), and that the HCE increases statistical power (16 of 73 [21.9%]). Furthermore, 3 (4.1%) switched from a traditional end point to an HCE when realizing that they could not meet the originally targeted sample size due to recruitment issues. Only 10 RCTs (13.7%) discussed potential limitations of HCEs, most commonly noting difficulties in interpreting the effect size (5 of 73 [6.8%]) (Table 3).

Table 3. Development, Rationale, and Limitations of Hierarchical Composite End Points.

Characteristics	RCTs, No./total No. (%)
Terminology used for end point
Hierarchical composite end point/outcome	62/92 (67.4)
Rank composite	12/92 (13.0)
DOOR	10/92 (10.9)
Othera	8/92 (8.7)
Development of hierarchy of HCEb
Nothing reported	60/73c (82.2)
Provided information on HCE development	13/73 (17.8)
Hierarchy already established (reference provided)	6/73 (8.2)
Patient and public involvement	2/73 (2.7)
Expert committee	2/73 (2.7)
Otherd	4/73 (5.5)
Justification for use of HCEb
No justification reported	34/73 (46.6)
Justification reported	39/73 (53.4)
Advantage vs traditional composite end points	26/73 (35.6)
Better reflects overall health	24/73 (32.9)
Increase powere	16/73 (21.9)
Otherf	8/73 (11.0)
Limitations of HCEb
No limitations reported	63/73 (86.3)
Limitations reported	10/73 (13.7)
Effect size difficult to interpret or unclear	5/73 (6.8)
Otherg	7/73 (9.6)

Abbreviations: DOOR, desirability of outcome ranking; HCE, hierarchical composite end point; RCT, randomized clinical trial.

^a Including 4 with win ratio approach, 2 self-declared as “composite” but analyzed with hierarchy, and 2 net treatment benefit.

^b Multiple answers could be reported and counted toward multiple categories.

^c Including 73 studies; excluding 19 studies with only registry entry available.

^d Including 2 with rationale provided by authors, 1 suggested by data monitoring committee, and 1 simulation.

^e Including 3 in which started trials faced recruitment issues and therefore changed the originally planned end point to a hierarchical composite end point during the trial conduct.

^f Including 5 to better reflect and handle missing data, 1 for safety aspects at higher hierarchy where no difference is expected, 1 technical advantage (no need for the proportionality of hazards), and 1 combining continuous and binary end points.

^g Including 2 with dominance of less important components, 2 due to compensatory nature, may mask worse results for individual end points, 1 method may overestimate effect, 1 adjustment not possible or difficult, and 1 limited generalizability.

Sample Size Estimation, Analysis, and Reporting of Results

The sample size assumptions for all components in the control group and the expected improvement for all components were reported in 43.8% (32 of 73) and 47.9% (35 of 73), respectively (Table 4). Approximately half of the trials (40 of 73 [54.8%]) adjusted the sample size for loss to follow-up or provided a rationale why it was not needed. Only 1 RCT (1.4%) reported how much the individual components contribute to the overall power of the trial.

Table 4. Sample Size Estimation, Analysis, and Reporting of Results for Hierarchical Composite End Points.

Characteristic	RCTs, No. (%)a
Protocol, results, or statistical analysis plan available (n = 73)
Assumptions for control group reported for each component
For all components	32 (43.8)
Only for single components (ie, not for all)	12 (16.4)
Not at all	29 (39.7)
Expected improvement for each component reported
For all components	35 (47.9)
Only for single components (ie, not for all)	10 (13.7)
Not at all	28 (38.4)
α Reported (type I error)	64 (87.7)
Power reported	69 (94.5)
Each component’s contribution (eg, % or proportion) to overall trial power indicated	1 (1.4)
Correction for loss to follow-up (or providing rationale why not done)	40 (54.8)
Reporting results (45 RCTs)
Results presentationb
Table with results for each component	20 (44.4)
Bar charts	10 (22.2)
Flowcharts	8 (17.8)
Results without illustration	6 (13.3)
Otherc	14 (31.1)
Primary analysis of hierarchical composite end point
GPC	29 (64.4)
GPC using win ratiod	21 (46.7)
Rank order method	11 (24.4)
Probabilistic index model	2 (4.4)
Othere	3 (6.6)
Additional partial credit analysis	1 (2.2)
Handling of missing data
No imputation or negligible missing	14 (31.1)f
Multiple imputation	8 (17.8)
Censoring on last observation	3 (6.6)
Single imputation	2 (4.4)
Otherg	5 (11.1)
Unclear	13 (28.8)
Missing data for each component reported	4 (8.9)
Adjustment for covariates in analysis	8 (17.8)
Each component included as a secondary end point	28 (62.2)
Pairwise comparison conductedh	33 (73.3)
No. of patient pairs reported	18 (40.0)
Win ratio/difference for each component reported	16 (35.6)
No. of ties for each component reported	11 (24.4)
Weight of each component on overall effect reported	6 (13.3)

Abbreviations: GPC, generalized pairwise comparison; RCT, randomized clinical trial.

^a For the sample size estimation and planned analysis only, studies were included for which at least a protocol, results publication, or statistical analysis plan was available (73 studies); for the analysis, only studies with results publication were considered (45 studies).

^b Multiple answers could be reported and counted toward multiple categories.

^c Including 9 with tables, 3 with forest plots, 2 with Kaplan-Meier curves (1 of those with additional table).

^d Including 19 (90.5%) conducted unmatched win ratio and 1 (4.8%) matched win ratio. One study reported using stratified win ratio (4.8%).

^e Including 1 analysis of covariance, 1 nonlinear mixed-effects model, and 1 model-based approach with adjustment and pairwise comparison as sensitivity analysis.

^f Including 8 reporting a sensitivity analysis using imputed data (4 studies) or tipping point analysis (2 studies), and 1 using inverse probability weighted analysis or different scenarios.

^g Including 2 studies with next hierarchy, 1 last observation carried forward, 1 combination of imputing and censoring, and 1 decision by independent outcome committee in case of missing data.

^h Including 4 studies that conducted a pairwise comparison not as the primary analysis.

Published results were available for 45 trials. The presentation of published results consisted of tables with results for each component (20 of 45 [44.4%]), flowcharts (8 of 45 [17.8%]), bar charts (10 of 45 [22.2%]), and results without illustration (6 of 45 [13.3%]). Two-thirds did not include each component of the primary end point as a secondary end point (28 of 45 [62.2%]). Only 6 trials (13.3%) reported the weight of each component on the overall effect size, and 4 (8.9%) reported the proportion of missing data for each component of the HCE. Adjusting the analysis for covariates (eg, stratification factors) was rarely conducted (8 of 45 [17.8%]). The most frequently used analysis methods were generalized pairwise comparisons (29 of 45 [64.4%]) and rank order method (11 of 45 [24.4%]) (Table 4). The most commonly reported summary measure was the win ratio (21 of 45 [46.7%]; 19 unmatched win ratios, 1 matched win ratio, and 1 stratified win ratio). A partial-credit analysis assigning varying weights to different outcome components was reported in 1 RCT (Table 4).

Among the 29 trials conducting a pairwise comparison, we identified a total of 11 RCTs that reported the overall wins, losses, and unresolved ties in the results publication. The proportion of remaining unresolved ties among these 11 trials ranged from 0.3% up to 77.3% (median [IQR] 23.8% [3.1%-31.8%]) (eTable 5 in Supplement 1). One reported win odds and win ratio, the remaining 10 reported win ratio.

Discussion

Our scoping review revealed that RCTs with HCEs as their primary end point have become more frequent, primarily in cardiovascular trials, but with growing adoption in other medical fields. This is in agreement with a recently published systematic review focusing on win ratio which showed an increased use beyond cardiovascular trials.²⁸ The main arguments reported by trialists for using an HCE were the general advantages over traditional composite end points, that HCEs can better reflect overall health of patients, and the increase of power when using an HCE. Our scoping review reveals substantial heterogeneity and shortcomings in how investigators plan, analyze, and report the use of HCEs in RCTs. Given the potential methodological pitfalls associated with HCEs, as highlighted in the current research,^17,29,30 the development of clear, standardized guidelines is essential to mitigate these risks. For example, when developing an HCE, Ajufo and colleagues¹⁷ recommend that several stakeholders, including patients, trialists and regulators, should be included to ensure that the HCE is patient-centered and clinically relevant. However, our scoping review has shown that for most RCTs (over 80%), it was not reported at all how the hierarchy was developed. Furthermore, reporting of the sample size calculation, so readers can understand the underlying assumptions for each component,³¹ was often incomplete. Only about half of the trials provided the expected event rates or values in the control group for each component, the anticipated improvement in the intervention group, and details on how the sample size was adjusted for missing data.

Our findings have several implications for clinicians, trialists, and policymakers. These stakeholders should be aware when assessing effect estimates of an HCE that the overall effect may be strongly driven by its lowest component.¹⁷ Therefore, all included components should be clinically meaningful to patients, and the contribution of each component to the overall effect size (ie, proportion of overall decisions made on this component) should be reported. Among the RCTs included in our study, only 13% reported the weight of each component on the overall effect size. Another frequently discussed limitation of HCEs (in particular for the win ratio) is that the win ratio does not include unresolved ties in the analysis. This is especially concerning when the proportion of unresolved ties is high.²⁹ In the 11 RCTs that performed pairwise comparisons and reported unresolved ties, this proportion was substantial, with a median of 23.8% and a maximum of 77.3% in one of the trials. To account for this problem, the use of win odds, which allocates half the ties to the intervention and the other half to the control group, is often proposed.^32,33 In our review, we identified only 1 RCT that reported win odds, and this trial had a negligible proportion of unresolved ties (0.6%). It is noteworthy that Pocock and colleagues argue that “it is a myth that ties are a problem using the win ratio,” suggesting that the win difference can be calculated for each component to quantify the absolute benefit.³⁰ Yet, none of the trials we identified reported the win difference. We suggest that these different target parameters (win ratio, win odds, win difference or net treatment benefit, and probabilistic index) be reported in parallel to provide complementary measures of both the absolute and relative treatment effect.^34,35

One additional remaining question is how to best present results of HCEs. Our review revealed a wide range of approaches, many of which lacked transparency. We suggest that a flowchart clearly illustrating the hierarchy and decision-making process may offer the most intuitive and comprehensive solution. The flowchart should not only include the overall wins and losses but be complemented by detailed numerical information for each component (such as the counts of wins, ties, and losses; eg, as presented by Mentz and colleagues³⁶), along with its weight on the overall effect and corresponding win difference, enabling readers to clearly understand the contribution of each element to the overall effect. The optimal approach to integrating covariate adjustment into both the analysis and presentation of results remains uncertain and warrants further investigation.³⁰

Limitations

While our scoping review provides a broad overview of RCTs employing HCEs as a primary end point, we acknowledge several limitations. First, because HCEs combine features from ordinal and composite end points,¹⁹ clearly defining an end point as an HCE is not always straightforward. For operational purposes we used a combination of either self-declaration or the use of analytical approaches typical for HCEs. Consequently, we may have included studies whose primary end point is not universally recognized as an HCE. In addition, we may have missed trials using HCEs which are not clearly labeled as such. Second, our scoping review provides only a snapshot of the current trial landscape. A large proportion of trials was still ongoing, with only limited information (eg, only from trial registries) and no published results available. It is therefore possible that certain aspects (eg, reporting quality of study results) may change. Third, our study did not allow assessment of other important considerations, such as resource savings compared with traditional end points, differences in effect sizes between composite end points and HCEs, or the level of acceptance of HCEs among decision-makers. Alternative designs such as economic evaluations, statistical simulations, and qualitative studies are needed to answer these remaining questions.

Conclusions

In this scoping review, we showed that the use of HCEs is expanding across medical fields in both academic and industry-sponsored trials, offering multiple potential advantages for evaluating clinically meaningful, patient-centered outcomes in RCTs. At the same time, our review showed considerable heterogeneity and shortcomings in how HCEs are developed, analyzed, and reported, with key steps such as hierarchy justification or handling of ties often being insufficiently described. This limits the interpretability of trial findings and complicates evidence synthesis. Establishing clear methodological standards and reporting recommendations for RCTs using HCEs would support more consistent implementation and enhance the validity and interpretability of HCE-based evidence for clinicians, patients, researchers, and decision-makers.

Supplement 1.

eAppendix 1. Full search strategy conducted in different databases

eAppendix 2. Data extraction form

eTable 1. Table of all included studies

eTable 2. Characteristics of trials stratified by sponsor

eTable 3. Categories of components used in the 92 identified hierarchical composite endpoints

eTable 4. Sub-group analysis of key characteristics by medical field

eTable 5. Studies that conducted a pairwise comparison and reported winners, losers, and ties for the hierarchical endpoint

eFigure. Flow chart

Supplement 2.

Data Sharing Statement