Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Aug 20.
Published in final edited form as: Pediatr Blood Cancer. 2026 Jul 25;73(9):e70593. doi: 10.1002/1545-5017.70593

On Hierarchical Composite Endpoints in Pediatric Cancer Supportive Care: Illustrative Examples From Two Multi-Center Phase-III Randomized Clinical Trials

Willem H Collier 1, Mark Zobeck 2,3, Adam J Esbenshade 4, Christopher C Dvorak 5, Lillian Sung 6, David Freyer 7,8, Sarah Alexander 6, Etan Orgel 7,8, Nicole J Ullrich 9, Zach Prudowsky 2,3, Brian Fisher 10,11, Caitlin W Elgarten 12
PMCID: PMC13487803  NIHMSID: NIHMS2199679  PMID: 42499270

Abstract

Pediatric supportive care clinical trials often involve multiple clinically important outcomes, complicating trial interpretation. Hierarchical composite endpoints (HCEs) provide a framework to integrate key outcomes according to clinical importance. We performed post hoc analyses using HCE in two randomized trials conducted by the Children’s Oncology Group. We found that HCE can be more sensitive overall endpoints for detecting treatment benefit as well as found that HCE can support harmonized trial conclusions in the presence of intervention benefits and harms. These analyses illustrate the potential of HCE to improve the interpretability of complex pediatric supportive care trials and support consideration of their prospective use.

Keywords: hierarchical composite endpoints, randomized clinical trials, supportive care

1 I. Introduction

Optimizing outcomes for pediatric cancer patients relies on the successful advancement of evidence-based supportive care to ease treatment burden and mitigate long-term side effects of cancer therapy. The Children’s Oncology Group (COG) Cancer Control (CCL) Committee oversees the design and completion of pediatric supportive care clinical trials [1]. Interpreting supportive care trials remains challenging, at least in part, due to a multiplicity of key clinical outcomes identified a priori as integral to interpretation of intervention effects [2–4]. For example, a supportive care trial of bacterial prophylaxis would have multidimensional hypothesized benefits, including reductions in neutropenic fever episodes, empiric antibiotic exposure, bacteremia, and/or death.On the other hand, some trials necessitate assessment of whether an intervention interferes with cancer-directed treatment (e.g., through disease outcomes). In this article, we discuss how outcome plurality uniquely complicates interpretation of pediatric supportive care trials, and we describe an analytical approach, which can be used to address these challenges.

Traditionally, separate analyses are performed for each distinct outcome identified as relevant to evaluate comparative effectiveness in a supportive care trial [2–4]. However, trials are typically powered for analysis of an a priori defined primary outcome, often leading to underpowered analyses of other pertinent outcomes [3, 5]. This can be problematic for two reasons. First, secondary outcomes can still reflect clinically important benefits of an intervention, but their standalone analysis, especially if underpowered, cannot be used towards an efficacy conclusion. Second, separate analyses can lead to conflicting findings of benefit and harm. An approach to handling outcome multiplicity is the use of a composite endpoint. Traditional composites impart clinical equivalence to each component, which can mask the treatment effect on outcomes that occur less frequently. The hierarchical composite endpoint (HCE) is an endpoint construction strategy developed to overcome such challenges [6]. HCEs are a composite wherein each outcome is a priori ranked according to clinical importance. In the analysis, patients are compared using this hierarchy, which effectively gives higher weight to the more clinically important outcomes. Initially devised for trials of heart failure [6, 7], HCEs have gained traction and have subsequently been used in pivotal Phase III trials [7–13].

We postulate HCE will be a useful tool for supportive care trials in cancer. We leveraged data from two previously completed trials conducted through COG to perform post hoc HCE analyses: ACCL0934, a trial of levofloxacin prophylaxis, and ACCL0431, a trial assessing an otoprotectant [2, 3]. These analyses provide an empirical assessment of HCE to motivate consideration of this approach in supportive care.

2 I. Methods

2.1 I. Overview of HCE Construction and Analysis Using Win-Statistics

Constructing an HCE is a two-step process. First, the key outcomes of the trial are defined as HCE components. Second, HCE component outcomes are ranked in order of clinical importance. It is imperative that each individual HCE component be clinically important should any one component strongly drive overall effect size. Non-cancer examples are available [7, 13–15].

A class of metrics, termed “win statistics,” was developed to quantify randomized treatment effects on HCE [6, 7]. Win-statistics are flexible in that they can handle HCE components of mixed data types (e.g., time-to-event and continuous) and can be estimated using nonparametric methods (limiting mathematical assumptions) [6, 16–19]. Among several frequently used win-statistic metrics, we use the win-odds ratio due to its interpretability in the presence of “ties” (defined below) [20, 21]. Win-odds ratios reflect the odds that a randomly selected patient from the active arm has a better overall outcome on the basis of the HCE than a randomly selected control arm patient, and a win-odds greater than 1 implies benefit of randomization to the active option [21].

The nonparametric approach to estimation of the win-odds ratio is as follows: Firstly, define P1 as the probability of a randomly selected treatment arm patient has a better overall outcome, defined by the HCE, than a randomly selected control arm patient; P2 : The probability of a control arm patient has a better overall outcome; Ptie: The probability of a tie. The win-odds ratio is (P1 + Ptie)/(P2 + Ptie). For a fixed follow-up period for which every patient has non-missing data on HCE components, these quantities can be estimated by performing head-to-head comparisons of each pair of active and control arm patients. For P1, calculate the proportion, across all possible pairwise comparisons for which the treatment arm patient “wins” over the control arm patient based on the algorithm: Each two patients are first compared using the most clinically important HCE component, and if a win (e.g., one patient experiences the event and the other does not) cannot be determined, the patients are compared with respect to the second component, and so on. If, across HCE components, a winner could not be determined, a tie has resulted.

2.2 I. Summary of Included Randomized Trials

Both ACCL0934 and ACCL0431 have been previously published (see Supporting Information) [2, 3]. ACCL0934 was a multicenter, open-label, randomized Phase 3 trial comparing levofloxacin to no prophylaxis (1:1 randomization) for the prevention of blood-stream infections (BSIs) in patients undergoing chemotherapy for treatment of acute leukemia (Cohort 1), and separately, for those undergoing hematopoietic cell transplantation (HCT, Cohort 2). The primary outcome of ACCL0934 was centrally adjudicated BSI, with additional outcomes collected, including neutropenic fevers (FN), severe infections, severe adverse events (e.g., CTCAE grade 4), Clostridioides difficile infections, and deaths. The infection observation period on ACCL0934 was tied to the period of severe neutropenia (ANC < 200 cells/μL).

ACCL0431 was a multicenter, open-label randomized Phase 3 trial of sodium thiosulfate (STS) for the prevention of cisplatin-induced hearing loss. Patients with different malignancies were enrolled (e.g., medulloblastoma and osteosarcoma). Participants were randomized 1:1 to STS or no STS. Trial endpoints included hearing loss following therapy (primary endpoint), renal or hematologic toxicities, cancer relapse, progression, secondary malignancy, and death. Patients were followed for the ototoxicity primary outcome up to approximately 4 weeks following completion of cisplatin therapy, and for disease outcomes until the trial was administratively completed [3].

Written informed consent/assent of participants or their legal guardians was obtained prior to enrollment on ACCL0934 or ACCL0431, whichever was relevant. Both studies were registered with ClinicalTrials.gov (ACCL0934: NCT01371656; ACCL0431: NCT00716976). Each study was initially approved by the US National Cancer Institute (NCI) Central Institutional Review Board (IRB), and subsequently by each COG site IRB, if necessary.

2.3 I. Data Analyses Performed

2.3.1 I. ACCL0934 Hypothetical HCE

For ACCL0934, the first HCE (HCE1) included the following components, ordered from most to least clinically important: (1) death (all cause), (2) severe infection, (3) BSI, and (4) days of FN. Infectious outcomes of varying clinical severity were chosen, which facilitated an endpoint hypothesized to be more sensitive to the effect of bacterial prophylaxis than BSI alone. The second HCE (HCE2) was expanded to include potential harms of levofloxacin per the ranking as follows: (1) death, (2) severe infection, (3) other serious adverse event, (4) BSI, (5) C difficile diarrhea, and (6) the duration of FN. Death attributable to BSI was substituted for all-cause death for analyses presented in the Supporting Information. Days of FN were continuous and other outcomes binary.

2.3.2 I. ACCL0431 Hypothetical HCE

For ACCL0431, HCE1 pertained to disease and hearing loss outcomes per the following ordering from most to least clinically important: (1) death, (2) oncologic event (cancer relapse, progression, or a secondary malignancy), and (3) hearing loss. This HCE was chosen such that the analysis could weigh observed benefits of STS (hearing loss) with observed harms (death and oncologic events) previously reported [3]. All outcomes were analyzed as time to event. For HCE2 a fourth component was included: (4) targeted renal or hematologic toxicity (binary outcome).

Hearing loss was categorized using the International Society of Paediatric Oncology (SIOP) severity grading system as used in post hoc analyses; this outcome was dichotomized with SIOP Grade ≥ 2 (communication-impacting hearing loss) defined as hearing loss present [22]. Analysis of the original American Speech-Language-Hearing Association (ASHA)-defined outcome is presented in the Supporting Information.

The HCE definitions were developed by the clinician co-authors on this study, with consensus reached across all such contributors.

2.3.3 I. Cohorts and Follow-Up Time

For ACCL0934, analyses were performed separately for acute leukemia and HCT cohorts, and all patients eligible in the primary publication were included. Patients were followed for the protocol-defined infection observation period.

For reanalysis of ACCL0431, all eligible participants except those who withdrew within 1 day of enrollment were included. Four participants without SIOP gradable evaluations were included but were assumed not to have had a hearing loss event. This ensured participation for analysis of disease outcomes. For win-statistical analysis, a fixed follow-up period must be defined. We performed analysis with follow-up at truncated at 1-year post-enrollment and the another at 3 years. We chose 1 year to approximate the observation time intended for the ototoxicity outcome. We sought to evaluate the impact of extended disease outcome monitoring and chose 3 years for illustration. Additional analysis was restricted to the cohort with localized disease only, the population for which STS has regulatory approval [23].

2.3.4 I. Statistical Analysis

We estimated win-odds ratios using the nonparametric pairwise comparison approach described. For ACCL0934, all patients were considered to have completed the infection observation period. For ACCL0431, to follow the nonparametric approach, each pairwise comparison utilized the minimum follow-up time between the two patients. A closed form, asymptotic variance estimator was used to compute standard errors (and, hence, confidence intervals [CIs] and p values) [17, 19]. Additional analyses, including an inverse probability of censoring weighted (IPCW) approach to address early censoring in ACCL0431, bootstrap-based inference of ACCL0431, and alternative win-statistic summary measures, are presented in the Supporting Information [24].

Demonstrative visual aids were generated. For ACCL0934, we utilized a combined bar and forest plot, showing the percentage of pairwise comparisons resulting in a win for each arm within each HCE component [7]. For the ACCL0431, we utilized the Maraca plot [25].

All analyses were implemented using R version 4.5.0, employing the WINS and MARACA packages [26, 27].

3 I. Results

3.1 I. Example 1: ACCL0934

The trial enrolled 624 eligible participants. There were 549 (90%), who completed the infection observation period, with the remainder completing between 10 and 92 days (median 22).

Table 1 summarizes HCE outcomes by arm. In the acute leukemia cohort, BSI cumulative incidences were 43% on control and 22% on levofloxacin. In the HCT cohort, BSI cumulative incidences were 17% on control and 11% on levofloxacin. Across HCE components, there tended to be a reduction in event incidence or duration on levofloxacin relative to control.

TABLE 1 |.

ACCL0934 outcome summaries and win-statistic analysis results.

 Event summaries for HCE components
Acute leukemia group
 HCT group
Levofloxacin Control  Levofloxacin Control
Overall patients (N) 96 99 210 208
Death (all cause)—n (%) (In HCE 1 and HCE 2) 6 (6) 7 (7) 25 (12) 25 (12)
Severe infection—n (%) (In HCE 1 and HCE 2) 5 (5) 10 (10) 6 (3) 8 (4)
Other serious adverse event—n (%) (In HCE 2 Only) 1 (1) 3 (3) 4 (2) 0 (0)
Bloodstream infection—n (%) (In HCE 1 and HCE 2) 21 (22) 43 (43) 23 (11) 36 (17)
C-diff associated diarrhea—n (%) (In HCE 2 only) 2 (2) 7 (7) 5 (2) 9 (4)
Days of febrile neutropenia—median (quartiles)(In HCE 1 and HCE 2) 4 (2.9) 6 (3.12) 2 (0.5) 3 (1.6)
Patients with no above-listed events and no FN episode—n (%) 12 (13) 9 (9) 47 (22) 20 (10)
Win-statistic analysis results
Acute leukemia group
 HCT group
Win-odds (ratio)—estimate Win-odds (ratio)—estimate
Endpoint (95% CI) p value (95% CI) p value
BSI only 1.55 (1.17–2.05) 0.003 1.14 (0.99–1.30) 0.065
HCE 1 1.74 (1.23–2.46) 0.002 1.28 (1.02–1.60) 0.031
HCE 2 1.77 (1.25–2.50) 0.001 1.27 (1.01–1.58) 0.037
Open in a new tab

Abbreviations: BSI, blood stream infection; HCE, hierarchical composite endpoint; HCT, hematopoietic cell transplantation.

Table 1 also summarizes our win-statistical analyses. For analysis of BSI alone, randomization to levofloxacin was beneficial in both cohorts, but only the estimate for the acute leukemia (AL) cohort achieved statistical significance (AL: win-odds [WO] = 1.55, p value = 0.003; HCT: WO = 1.15, p value = 0.065), as per the primary manuscript. For the analysis of HCE1, the estimated win-odds ratios were above 1 and reached statistical significance in both cohorts (AL: WO = 1.74, p value = 0.002; HCT: WO = 1.28, p value = 0.031). The effect sizes and p values under analysis of HCE2 were like those of analysis of HCE1, suggesting minimal impact of the added toxicity outcomes. In an additional analysis where all-cause deaths were replaced with death due to BSI in the HCE analyses, the win-odds effect sizes increased, and corresponding p values decreased (Table S1, HCE1: AL: WO = 1.83, p value = 0.001; HCT: WO = 1.39, p value = 0.004). The use of alternative win-statistics and methods to calculate standard errors yielded similar conclusions (Tables S2 and S3).

Figures 1 and 2 illustrate how each outcome influenced the estimated win-odds effect sizes. For the acute leukemia cohort, the largest absolute difference in wins across arms was observed in the BSI component. In contrast, in the HCT cohort, the largest absolute difference in wins occurred in the FN component.

FIGURE 1 |.

FIGURE 1 |

Open in a new tab

Win percentages and corresponding win-odds ratios in the acute leukemia cohort of ACCL0934. Panel (A) displays the percentage of patient level pairwise comparisons that resulted in a win within the HCE component displayed, broken up by randomized arm. For example, across pairwise comparisons performed on the BSI outcome, 26% resulted in a win for the levofloxacin arm patient. Panel (B) lists the estimated win-odds and plots the point estimate and corresponding 95% confidence interval width in forest plot format. Each win-odds is based on the HCE that includes the outcome listed in the corresponding row of the bar plot (Panel A) and all outcomes displayed above, if relevant. AE: other severe (non-severe infection) adverse event; BSI: true blood stream infection; Cdiff: Clostridioides difficile-associated diarrhea; Dth: death (all cause), FN: febrile neutropenia; HCE: hierarchical composite endpoints; SevInf: severe infection.

FIGURE 2 |.

FIGURE 2 |

Open in a new tab

Win percentages and corresponding win-odds in the HCT cohort of ACCL0934. Panel (A) displays the percentage of patient level pairwise comparisons that resulted in a win within the HCE component displayed, broken up by randomized arm. For example, across pairwise comparisons performed on the BSI outcome, 10% resulted in a win for the levofloxacin arm patient. Panel (B) lists the estimated win-odds and plots the point estimate and corresponding 95% confidence interval width in forest plot format. Each win-odds is based on the HCE that includes the outcome listed in the corresponding row of the bar plot (Panel A) and all outcomes displayed above, if relevantAE: other severe (non-severe infection) adverse event; BSI: true blood stream infection; Cdiff: Clostridioides difficile-associated diarrhea; Dth: death (all cause); FN: febrile neutropenia; HCE: hierarchical composite endpoints; HCT: hematopoietic cell transplantation; SevInf: severe infection

3.2 I. Example 2: ACCL0431

The trial enrolled 125 eligible participants; one was excluded due to study withdrawal on Day 1. Among the overall cohort, four were censored prior to 1 year (median follow-up 9 months; range 7–11), and 16 prior to 3 years (median 30 months; range 7–35). Among the localized cohort, three patients were censored prior to 1 year (median 9 months; range 7–11), and 11 prior to 3 years (median 34 months; range 7–35).

Table 2 summarizes HCE outcomes. Among the overall cohort, nearly 30% of patients on the control arm experienced clinically significant hearing loss, compared to only 5% of patients on the STS arm. The absolute reduction in hearing loss was similar in the localized disease cohort. Among the overall cohort, there was a larger proportion of patients who died or experienced an oncologic event on the STS arm, which was more pronounced with 3-year follow-up. Regardless of follow-up, there was a similar proportional incidence of death or oncologic events across arms among the localized cohort.

TABLE 2 |.

ACCL0431 outcome summaries and win-statistic analysis results.

Event summaries for HCE components
Overall cohort: 1-year follow-up
 Overall cohort: 3-years follow-up
 Localized cohort: 1-year follow-up
 Localized cohort: 3-years follow-up
STS Control STS Control STS Control STS Control
Overall patients (N) 60 64 60 64 38 38 38 38
Death—n (%) (In HCE 1 and HCE 2) 7 (12) 6 (9) 17 (28) 8 (12) 3 (8) 3 (8) 6 (16) 4 (11)
Oncologic event—n (%) (In HCE 1 and HCE 2) 14 (23) 12 (19) 26 (43) 22 (34) 8 (21) 8 (21) 14 (37) 13 (34)
Hearing Loss—n (%) (In HCE 1 and HCE 2) 3 (5) 17 (27) 3 (5) 18 (28) 2 (5) 8 (21) 2 (5) 9 (24)
Targeted renal or hematologic toxicity—n (%)(In HCE 2 only) 31 (52) 41 (64) 31 (52) 41 (64) 23 (61) 29 (76) 23 (61) 29 (76)
Patients with no above-listed event—n (%) 14 (23) 9 (14) 7 (12) 6 (9) 10 (26) 5 (13) 4 (11) 6 (16)
Win-statistic analysis results
Overall cohort: 1-year follow-up
 Overall cohort: 3-years follow-up
 Localized cohort: 1-year follow-up
 Localized cohort: 3-years3 follow-up
Win-odds (ratio)—estimate Win-odds (ratio)—estimate Win-odds (ratio)—estimate Win-odds (ratio)—estimate
Endpoint (95% CI) p value (95% CI) p value (95% CI) p value (95% CI) p value
Hearing loss only 1.55 (1.17–2.06) 0.003 1.60 (1.19–2.15) 0.002 1.38 (0.99–1.90) 0.056 1.45 (1.03–2.05) 0.036
HCE 1 1.27 (0.89–1.81) 0.182 1.00 (0.68–1.45) 0.985 1.20 (0.78–1.85) 0.402 1.09 (0.69–1.73) 0.701
HCE 2 1.24 (0.84–1.85) 0.28 1.00 (0.67–1.49) 0.994 1.06 (0.66–1.71) 0.816 1.05 (0.65–1.70) 0.844
Open in a new tab

Note: Oncologic event is defined as cancer relapse, progression, or secondary malignancy. Hearing loss is defined as SIOP grade ≥2.

Abbreviation: HCE, hierarchical composite endpoint.

The results of win-odds analyses are also displayed in Table 2. Analysis of hearing loss alone resulted in estimated win-odds reflective of benefit of randomization to STS versus control (1-year follow-up in the overall cohort: WO = 1.55; 95% CI: 1.17–2.06). In the overall cohort, win-odds effect sizes were closer to the null under analysis of HCE inclusive of death and disease outcomes relative to analysis of hearing loss alone (HCE1 with 3-year follow-up in the overall cohort: WO = 1.00; 95% CI: 0.69–1.45). The estimated win-odds differed meaningfully depending on the follow-up duration used (HCE1 overall cohort 1-year follow-up: WO = 1.27; 3-years follow-up: WO = 1.00). For the localized-only cohort, the estimated win-odds remained above 1 in all analyses. Additional analyses of the overall cohort are presented in Tables S4–S6. The use of IPCW did not meaningfully change the primary result (Table S5: HCE1 WO = 1.01; 95% CI: 0.69–1.48). Use of the ASHA-defined hearing endpoint, analyzed under the smaller cohort with ASHA-evaluable hearing assessments, resulted in a win-odds estimate of 1.10 under HCE1 with 3-years follow-up (Table S6; 95% CI: 0.72–1.69).

Figure 3 displays time-to-event curves for components of HCE1 under analysis of the overall cohort. The final panel is used to display censoring times. This figure illustrates the mechanism through which the estimated win-odds ratio converged to the null when follow-up was up to 3 years: The across-arm difference in oncologic events increased over time, balancing out the beneficial effect observed for hearing loss.

FIGURE 3 |.

FIGURE 3 |

Open in a new tab

Maraca Plot (time-to-event curves) for HCE components among the ACCL0431 overall cohort with follow-up truncated at 3 years. The x-axis reflects time, where for each HCE component (death, oncologic event, and hearing loss), Time 0 is represented by the bold grey vertical bar on the left of the component, and the maximum follow-up time (3 years) is represented by the bar to the right. For the censoring time column of the figure (right most), the first bold grey vertical bar represents the minimum censoring time observed. Oncologic events include relapse, secondary malignancy, or disease progression. SIOP Gr2+ HL refers to the binary SIOP grade 2 or higher hearing loss outcome. Censoring Time refers to the time a patient withdrew consent or was last seen if an HCE event hadn’t previously occurred. Except for the time-to-death curve, event curves displayed are among patients who did not experience the events listed to the left. This figure illustrates the mechanism through which the estimated win-odds (null value of WO = 1) arose based on the ACCL0431 overall cohort analysis with 3-year follow-up. There was a larger cumulative incidence of death and of oncologic events on the STS relative to control arm. Among patients who did not die or experience an oncologic event, there were more hearing loss events on control than on STS.

4 I. Discussion

Supportive care trials in pediatric cancer are challenging to design and interpret due to the plurality of key outcomes relevant to interpreting the overall effects of an intervention [2, 3]. HCEs and their analyses using win-statistics are powerful but nuanced tools developed for this challenge. These are employed to estimate an overall summary measure of the comparative effectiveness of treatments evaluated through a principled approach to handling heterogeneity in clinical importance across outcomes [7, 13]. We reanalyzed two previously conducted COG supportive care trials to evaluate the use of HCE in practice.

As in the primary analysis for ACCL0934, when analyzing the BSI primary endpoint alone, the estimated win-odds ratios favored levofloxacin prophylaxis in both cohorts, but the corresponding p value was only statistically significant for the acute leukemia cohort [2]. Under analysis of HCE, the win-odds ratios were greater than 1 and achieved statistical significance in both cohorts. The HCEs captured additional benefits of levofloxacin, including reductions in severe infections and neutropenic fevers. Relative to BSI alone, the HCEs were more sensitive to the broader benefits of levofloxacin, and win-statistic effect sizes were larger, yet with sufficiently narrow CIs to allow for efficacy conclusions. In general, an HCE should only be used if it will facilitate a scientifically meaningful conclusion. If this criterion is met, there may be scenarios in which the HCE analysis will facilitate a more statistically efficient trial than powering for separate analysis of individual outcomes that make up the HCE.

In the primary manuscript for ACCL0431, a statistically significant reduction in the proportional incidence of hearing loss was reported for STS compared to control [3]. However, considering all participants in aggregate, lower survival was observed on STS versus control. The conflicting directionality in the treatment effect on these outcomes posed challenges to trial interpretation. Our analyses that leverage an HCE inclusive of both ototoxicity and oncologic outcomes facilitate the conclusion that patients from the broader cohort do not have higher odds of overall benefit when randomized to STS relative to control. This harmonized interpretation could be viewed as a strength of the HCE approach, potentially generalizable to other supportive care trials that go on to exhibit both benefits and harms.

We also conducted analyses in the localized tumor cohort of ACCL0431, which highlighted different considerations. There, death and oncologic outcomes were similarly distributed across randomized arms, but a larger total number of such events were observed when follow-up was longer. We may expect the win-odds ratios to indicate the benefit of randomization to STS due to the ototoxicity reduction and similar distributions of death and oncologic events. Using longer follow-up led to weaker win-odds effect sizes under HCE analyses. This reflects a known property of win-statistical methods. When the highest ranked HCE outcome occurs more frequently, even if equivalently distributed between arms, the pairwise comparison methodology will find the two groups as more similar, attenuating the estimated effect size.

We emphasize several additional takeaways. Although win-odds ratios provide a single measure of effect size with respect to multiple outcomes, the overall effect size should not be interpreted in the absence of understanding which of the HCE-embedded outcomes differ most markedly across arms. Simple summary statistics and visual aids were critical for such transparency in our analyses. For example, the figure used for ACCL0934 showed FN duration, the lowest ranking HCE component chosen, was needed for an efficacy conclusion using the win-odds ratio in the HCT cohort. In alignment with the HCE literature, if constructing a trial protocol around analysis of an HCE, we advocate for declaring planned analyses of the individual HCE-embedded outcomes, even if not powering for those analyses, and corresponding visual aids [7, 25].

Next, regarding ACCL0431, when analyzing the overall cohort with follow-up limited to 1 year, the estimated win-odds were greater than 1, whereas when follow-up was extended to 3 years the estimated win odds ratios were reduced to the null (1.00) due to additional disease events after 1 year of follow-up. This illustrates that win-statistic metrics are follow-up time dependent [16]. If prospectively designing a trial around an HCE, the follow-up period may naturally extend from a well-defined at-risk period. For example, infectious outcome trials may anchor the analysis to a period of severe neutropenia. In other settings where HCE-relevant events do not pertain to an easily defined at-risk period, investigators may choose an extended follow-up duration likely to capture all or most such events on the basis of prevailing data and reasonable subject matter expertise. In either case, the HCE analysis interpretation must strictly pertain to the follow-up period chosen, and this decision should be transparent and strongly justified.

We also consider our analyses in the context of challenges emphasized in the HCE literature. There is an abundance of work describing the role of censoring in win-statistical analysis [16, 24, 28]. To motivate this issue, consider that the pairwise comparison procedure cannot be employed in a straightforward way when one participant is censored earlier than another. Although there are statistical methods proposed for these scenarios (e.g., IPCW methods), these can be more challenging to implement or may introduce barriers in trial design (censoring weights will be unknown in the design phase) [16, 24]. In our ACCL0431 analysis, the IPCW sensitivity analysis did not meaningfully change the results, which was likely due to several factors including large effect sizes and limited attrition. However, this finding may not generalize to other settings, such as those with more censoring, fewer events, or smaller effect sizes. Conveniently, many supportive care applications give rise to short follow-up periods wherein minimal early censoring may be expected. ACCL0934, among other infectious outcome trials conducted with COG, serves as important cases in demonstrating the feasibility in following patients for such durations [2, 4, 29, 30].

We acknowledge several limitations. First, the implications of our analyses are dependent upon trial designs not intended for the use of HCE. Second, the study team had knowledge of study results prior to constructing HCE. Third, prevailing limitations in the data could not be overcome. Additional outcomes may have been considered. For ACCL0431, HCE analysis could not overcome the possibility of unbalanced randomization in prognostic factors as a potential explanation for observed survival differences across STS and control groups with disseminated disease, as previously discussed [31]. Finally, this report pertains to just two previously conducted trials. These trials were chosen because of key illustrative differences and because of known quandaries in their interpretation. A comprehensive evaluation of COG-conducted supportive care trials under HCE was beyond the scope of this methodologic evaluation.

Overall, we highlight two opportunities to improve supportive care trials in pediatric cancer under use of HCE. First, HCE can encompass multiple clinically important outcomes, which represent an array of potential benefits of an intervention. There, an HCE may better represent the complete outcome experience from a patient-centered perspective and can lead to a more sensitive overall endpoint, potentially facilitating lower sample sizes. Second, HCE can also encompass outcomes that represent tradeoffs of benefits and harms, harmonizing trial conclusions relative to separate analyses of relevant outcomes. HCE construction is complex and must involve multiple key stakeholders, including not only researchers and physicians but also patient advocates. Nonetheless, this approach should be considered for future supportive care trials.

Supplementary Material

Supplementary Material

Additional supporting information can be found online in the Supporting Information section.

Acknowledgments

This work was supported by the following mechanisms. The COG NCTN Operations Center Grant (U10CA180886) supported conduct of ACCL0934 and ACCL0431. The corresponding author’s effort was supported by the COG National Cancer Institute Community Oncology Research Program (NCORP) Research Base Grant (UG1CA189955). CWE is supported by an NHLBI Career Development Award (K23HL161309). These funders did not have a role in the design or reporting of the analyses presented in this manuscript, nor did the funders influence the decision to pursue publication.

Conflicts of Interest

C.D. reports consulting for Alexion, Inc., unrelated to this work. E.O. reports consulting (outside the scope of this work) for Jazz Pharmaceuticals and Syndax Pharmaceuticals. B.T.F. has received research funding from Pfizer and Merck and served on a data safety monitoring board for a study performed by Astellas, all for unrelated research. C.W.E. has received research funding from Jazz Pharmaceuticals for unrelated research. All other authors have no COI to report.

Abbreviations:

BSI

blood stream infection

COG

Children’s Oncology Group

FN

neutropenic fever

HCE

hierarchical composite endpoint

HCT

hematopoietic cell transplantation

SIOP

International Society of Paediatric Oncology

STS

sodium thiosulfate

Footnotes

Disclosure

The funder did not play a role in the design of the study; the collection, analysis, and interpretation of the data; the writing of the manuscript; and the decision to submit the manuscript for publication. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Ethics Statement

For both trials analyzed (ACCL0934 and ACCL0431), the protocol, all protocol amendments, and informed consent documents were reviewed and approved by the National Cancer Institute’s Central Institutional Review Board, and the institutional review board’s of participating centers. Written informed consent and, if applicable, assent were obtained from participants or their guardians prior to participation on trial, whichever trial relevant.

Data Availability Statement

The COG data sharing policy describes the release and use of COG individual subject data for use in research projects in accordance with National Clinical Trials Network (NCTN) Program and NCI Community Oncology Research Program (NCORP) guidelines. Only data expressly released from the oversight of the relevant COG data and Safety Monitoring Committee (DSMC) are available to be shared. Data sharing will ordinarily be considered only after the primary study manuscript is accepted for publication. For Phase 3 studies, individual-level de-identified datasets that would be sufficient to reproduce results provided in a publication containing the primary study analysis can be requested from the NCTN/NCORP Data Archive at https://nctn-data-archive.nci.nih.gov/. Data are available to researchers who wish to analyze the data in secondary studies to enhance the public health benefit of the original work and agree to the terms and conditions of use. Requests for access to COG protocol research data should be sent to datarequest@childrensoncologygroup.org. Data are available to researchers whose proposed analysis is found by COG to be feasible and of scientific merit and who agree to the terms and conditions of use. For all requests, no other study documents, including the protocol, will be made available, and no end date exists for requests. In addition to above, release of data collected in a clinical trial conducted under a binding collaborative agreement between COG or the NCI Cancer Therapy Evaluation Program (CTEP) and a pharmaceutical/biotechnology company must comply with the data sharing terms of the binding collaborative/contractual agreement and must receive the proper approvals.

References

  • 1.Esbenshade AJ, Sung L, Brackett J, et al. , “Children’s Oncology Group’s 2023 Blueprint for Research: Cancer Control and Supportive Care,” Pediatric Blood & Cancer 70, no. S6 (2023): e30568, 10.1002/pbc.30568. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Alexander S, Fisher BT, Gaur AH, et al. , “Effect of Levofloxacin Prophylaxis on Bacteremia in Children With Acute Leukemia or Undergoing Hematopoietic Stem Cell Transplantation: A Randomized Clinical Trial,” JAMA 320, no. 10 (2018): 995–1004, 10.1001/jama.2018.12512. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Freyer DR, Chen L, Krailo MD, et al. , “Effects of Sodium Thiosulfate Versus Observation on Development of Cisplatin-Induced Hearing Loss in Children With Cancer (ACCL0431): A Multicentre, Randomised, Controlled, Open-Label, Phase 3 Trial,” Lancet Oncology 18, no. 1 (2017): 63–74, 10.1016/S1470-2045(16)30625-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Fisher BT, Zaoutis T, Dvorak CC, et al. , “Effect of Caspofungin vs. Fluconazole Prophylaxis on Invasive Fungal Disease Among Children and Young Adults With Acute Myeloid Leukemia,” JAMA 322, no. 17 (2019): 1673–1681, 10.1001/jama.2019.15702. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Freyer DR, Orgel E, Knight K, and Krailo M, “Special Considerations in the Design and Implementation of Pediatric Otoprotection Trials,” Journal of Cancer Survivorship: Research and Practice 17, no. 1 (2023): 4–16, 10.1007/s11764-022-01312-x. [DOI] [PubMed] [Google Scholar]
  • 6.Pocock SJ, Ariti CA, Collier TJ, and Wang D, “The Win Ratio: A New Approach to the Analysis of Composite Endpoints in Clinical Trials Based on Clinical Priorities,” European Heart Journal 33, no. 2 (2012): 176–182, 10.1093/eurheartj/ehr352. [DOI] [PubMed] [Google Scholar]
  • 7.Pocock SJ, Gregson J, Collier TJ, Ferreira JP, and Stone GW, “The Win Ratio in Cardiology Trials: Lessons Learnt, New Developments, and Wise Future Use,” European Heart Journal 45, no. 44 (2024): 4684–4699, 10.1093/eurheartj/ehae647. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Walker H, McLeman L, Meyran D, et al. , “Co-Designing a Novel Ordinal Endpoint for an Adaptive Platform Trial, BANDICOOT, in Pediatric Hematopoietic Stem Cell Transplant,” Transplantation and Cellular Therapy 31, no. 5 (2025): 321.e1–321.e12, 10.1016/j.jtct.2025.01.894. [DOI] [Google Scholar]
  • 9.Maurer MS, Schwartz JH, Gundapaneni B, et al. , “Tafamidis Treatment for Patients With Transthyretin Amyloid Cardiomyopathy,” New England Journal of Medicine 379, no. 11 (2018): 1007–1016, 10.1056/NEJMoa1805689. [DOI] [PubMed] [Google Scholar]
  • 10.Kosiborod MN, Esterline R, Furtado RHM, et al. , “Dapagliflozin in Patients With Cardiometabolic Risk Factors Hospitalised With COVID-19 (DARE-19): A Randomised, Double-Blind, Placebo-Controlled, Phase 3 Trial,” Lancet Diabetes & Endocrinology 9, no. 9 (2021): 586–594, 10.1016/S2213-8587(21)00180-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Gillmore JD, Judge DP, Cappelli F, et al. , “Efficacy and Safety of Acoramidis in Transthyretin Amyloid Cardiomyopathy,” New England Journal of Medicine 390, no. 2 (2024): 132–142, 10.1056/NEJMoa2305434. [DOI] [PubMed] [Google Scholar]
  • 12.Sorajja P, Whisenant B, Hamid N, et al. , “Transcatheter Repair for Patients With Tricuspid Regurgitation,” New England Journal of Medicine 388, no. 20 (2023): 1833–1842, 10.1056/NEJMoa2300525. [DOI] [PubMed] [Google Scholar]
  • 13.Gasparyan SB, Buenconsejo J, Kowalewski EK, et al. , “Design and Analysis of Studies Based on Hierarchical Composite Endpoints: Insights From the DARE-19 Trial,” Therapeutic Innovation & Regulatory Science 56, no. 5 (2022): 785–794, 10.1007/s43441-022-00420-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Evans SR, Rubin D, Follmann D, et al. , “Desirability of Outcome Ranking (DOOR) and Response Adjusted for Duration of Antibiotic Risk (RADAR),” Clinical Infectious Diseases: An Official Publication of the Infectious Diseases Society of America 61, no. 5 (2015): 800–806, 10.1093/cid/civ495. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Little DJ, Gasparyan S, Schloemer P, et al. , “Validity and Utility of a Hierarchical Composite End Point for Clinical Trials of Kidney Disease Progression: A Review,” Journal of the American Society of Nephrology 34, no. 12 (2023): 1928–1935, 10.1681/ASN.0000000000000244. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Mao L, “Defining Estimand for the Win Ratio: Separate the True Effect From Censoring,” Clinical Trials: London, England) 21, no. 5 (2024):584–594, 10.1177/17407745241259356. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Dong G, Li D, Ballerstedt S, and Vandemeulebroecke M, “A Generalized Analytic Solution to the Win Ratio to Analyze a Composite Endpoint Considering the Clinical Importance Order Among Components,” Pharmaceutical Statistics 15, no. 5 (2016): 430–437, 10.1002/pst.1763. [DOI] [PubMed] [Google Scholar]
  • 18.Gasparyan SB, Folkvaljon F, Bengtsson O, Buenconsejo J, and Koch GG, “Adjusted Win Ratio With Stratification: Calculation Methods and Interpretation,” Statistical Methods in Medical Research 30, no. 2 (2021):580–611, 10.1177/0962280220942558. [DOI] [PubMed] [Google Scholar]
  • 19.Bebu I and Lachin JM, “Large Sample Inference for a Win Ratio Analysis of a Composite Outcome Based on Prioritized Components,” Biostatistics 17, no. 1 (2016): 178–187, 10.1093/biostatistics/ kxv032. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Dong G, Huang B, Verbeeck J, et al. , “Win Statistics (Win Ratio, Win Odds, and Net Benefit) Can Complement One Another to Show the Strength of the Treatment Effect on Time-to-Event Outcomes,” Pharmaceutical Statistics 22, no. 1 (2023): 20–33, 10.1002/pst.2251. [DOI] [PubMed] [Google Scholar]
  • 21.Brunner E, Vandemeulebroecke M, and Mütze T, “Win Odds: An Adaptation of the Win Ratio to Include Ties,” Statistics in Medicine 40, no. 14 (2021): 3367–3384, 10.1002/sim.8967. [DOI] [PubMed] [Google Scholar]
  • 22.Orgel E, Knight KR, Villaluna D, et al. , “Reevaluation of Sodium Thiosulfate Otoprotection Using the Consensus International Society of Paediatric Oncology Ototoxicity Scale: A Report From the Children’s Oncology Group Study ACCL0431,” Pediatric Blood & Cancer 70 (2023):e30550, 10.1002/pbc.30550. [DOI] [Google Scholar]
  • 23.U.S. Food and Drug Administration, FDA Approves Sodium Thiosulfate to Reduce the Risk of Ototoxicity Associated With Cisplatin in Pediatric Patients With Localized, Non-Metastatic Solid Tumors (U.S. Food and Drug Administration, 2024), https://www.fda.gov/drugs/resources-information-approved-drugs/fda-approves-sodium-thiosulfate-reduce-risk-ototoxicity-associated-cisplatin-pediatric-patients?utm_source=chatgpt.com. [Google Scholar]
  • 24.Dong G, Mao L, Huang B, et al. , “The Inverse-Probability-of-Censoring Weighting (IPCW) Adjusted Win Ratio Statistic: An Unbiased Estimator in the Presence of Independent Censoring,” Journal of Biopharmaceutical Statistics 30, no. 5 (2020): 882–899, 10.1080/10543406.2020.1757692. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Karpefors M, Lindholm D, and Gasparyan SB, “The maraca Plot: A Novel Visualization of Hierarchical Composite Endpoints,” Clinical Trials (London England) 20, no. 1 (2023): 84–88, 10.1177/17407745221134949. [DOI] [PubMed] [Google Scholar]
  • 26.Cui Y and Huang B, WINS: The R WINS Package. R package version 1.5.1. 2025, https://cran.r-project.org/web/packages/WINS/index.html.
  • 27.Karpefors M, Gasparyan SB, Anderson K, and Huhn M, Maraca: The Maraca Plot: Visualizing Hierarchical Composite Endpoints. R package version 1.0.0 (2025), https://cran.r-project.org/web/packages/maraca/index.html. [Google Scholar]
  • 28.Li H, Chen WC, Lu N, Tang R, and Zhao Y, “The Elusiveness of the Win Ratio Parameter in the Presence of Missing Data,” Therapeutic Innovation & Regulatory Science 58, no. 3 (2024): 431–432, 10.1007/s43441-024-00645-2. [DOI] [PubMed] [Google Scholar]
  • 29.Zerr DM, Milstone AM, Dvorak CC, et al. , “Chlorhexidine Gluconate Bathing in Children With Cancer or Those Undergoing Hematopoietic Stem Cell Transplantation: A Double-Blinded Randomized Controlled Trial From the Children’s Oncology Group,” Cancer 127, no. 1 (2021): 56–66, 10.1002/cncr.33271. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Dvorak CC, Fisher BT, Esbenshade AJ, et al. , “A Randomized Trial of Caspofungin vs. Triazoles Prophylaxis for Invasive Fungal Disease in Pediatric Allogeneic Hematopoietic Cell Transplant,” Journal of the Pediatric Infectious Diseases Society 10, no. 4 (2020): 417–425, 10.1093/jpids/piaa119. [DOI] [Google Scholar]
  • 31.Orgel E, Villaluna D, Krailo MD, Esbenshade A, Sung L, and Freyer DR, “Sodium Thiosulfate for Prevention of Cisplatin-Induced Hearing Loss: Updated Survival From ACCL0431,” Lancet Oncology 23, no. 5 (2022): 570–572, 10.1016/S1470-2045(22)00155-3. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Material
NIHMS2199679-supplement-Supplementary_Material.docx (45.7KB, docx)

Data Availability Statement

The COG data sharing policy describes the release and use of COG individual subject data for use in research projects in accordance with National Clinical Trials Network (NCTN) Program and NCI Community Oncology Research Program (NCORP) guidelines. Only data expressly released from the oversight of the relevant COG data and Safety Monitoring Committee (DSMC) are available to be shared. Data sharing will ordinarily be considered only after the primary study manuscript is accepted for publication. For Phase 3 studies, individual-level de-identified datasets that would be sufficient to reproduce results provided in a publication containing the primary study analysis can be requested from the NCTN/NCORP Data Archive at https://nctn-data-archive.nci.nih.gov/. Data are available to researchers who wish to analyze the data in secondary studies to enhance the public health benefit of the original work and agree to the terms and conditions of use. Requests for access to COG protocol research data should be sent to datarequest@childrensoncologygroup.org. Data are available to researchers whose proposed analysis is found by COG to be feasible and of scientific merit and who agree to the terms and conditions of use. For all requests, no other study documents, including the protocol, will be made available, and no end date exists for requests. In addition to above, release of data collected in a clinical trial conducted under a binding collaborative agreement between COG or the NCI Cancer Therapy Evaluation Program (CTEP) and a pharmaceutical/biotechnology company must comply with the data sharing terms of the binding collaborative/contractual agreement and must receive the proper approvals.

RESOURCES