(Neo)adjuvant approvals for solid tumors by the US Food and Drug Administration 2010-2023

Abstract

Background

(Neo)adjuvant systemic treatments aim to improve overall survival by reducing recurrence risk; however, there can be challenges, such as reliance on surrogate endpoints, censoring, loss to follow-up, and impact on quality of life. We examine the design, reporting, and outcomes of US Food and Drug Administration–approved drugs for solid tumors.

Methods

We searched Food and Drug Administration (neo)adjuvant approvals for solid tumors 2010-2023 with time-to-event endpoints. Trial characteristics, endpoints, and censoring were evaluated descriptively. Statistical designs and sample size calculations were assessed for reproducibility and transparency.

Results

A total of 24 studies supporting 25 approvals met inclusion criteria. Common tumor types included melanoma (7 of 24, 29%), breast (6 of 24, 25%), and non-small cell lung cancer (6 of 24, 25%). Among 21 trials reporting overall survival, 52% (11 of 21) showed benefit at any time, but only 36% (4 of 11) reported overall survival benefits at the time of approval and/or primary publication. In 21 (88%) trials the expected number of events could be reproduced, in 11 (46%) trials sample size calculations could be attempted (though calculations often differed), and only 14 (58%) stated expected absolute numerical outcomes. Censoring rules were applied inconsistently; only 36% defined rules adequately. Quality-of-life outcomes were reported in 79%, with 31% reporting clinically meaningful declines in one or more domains.

Conclusions

(Neo)adjuvant trials often lack transparency in sample size justifications and censoring rules. Approximately half demonstrated an overall survival improvement at any time, and one-third showed a decline in quality of life. Greater standardization in trial designs and reporting is needed to ensure trials align with patient and health-care priorities.

Introduction

(Neo)adjuvant systemic treatments are given to reduce recurrence and improve overall survival for patients diagnosed with cancer. When delivered prior to definitive treatment (typically surgery), neoadjuvant treatment can also improve surgical outcomes or facilitate less extensive surgery. However, some patients treated with (neo)adjuvant drug-based therapies may have been cured with surgery alone. Therefore, (neo)adjuvant trials must carefully balance the potential benefits in terms of improved overall survival and/or reduced recurrence with the negative impacts on quality of life (QOL) and risks of short- and long-term toxicities.

To increase trial feasibility and decrease the time for drugs to reach market, intermediate endpoints are frequent approval endpoints in the (neo)adjuvant setting.¹ However, for most tumors they have no proven surrogacy with overall survival.^2,3 Demonstrating overall survival benefits from (neo)adjuvant treatments is complicated by the long natural history for some cancers and the availability of effective therapies in the postrecurrence (metastatic) setting. The extended follow-up required to reach surrogate time-to-event endpoints and overall survival endpoints in the (neo)adjuvant setting increases the likelihood of incomplete follow-up, missing data, and censoring, which may cause bias and complicate the interpretation of efficacy.⁴ (Neo)adjuvant studies often run over many years, and changes in baseline prognosis over time can also complicate study design and results interpretation. Despite the existence of regulatory recommendations⁵ and guidelines to report censoring and handling of missing data,^6,7 the use standardized approaches are inconsistent.⁸

The analysis and interpretation of QOL data are also unique in the (neo)adjuvant setting. Unlike the metastatic setting, where interventions may improve QOL by palliating symptoms, in the (neo)adjuvant setting, treatments may be more likely to cause a decline in QOL. With the emergence of immunotherapy in the (neo)adjuvant setting, understanding the chronic QOL impacts of (neo)adjuvant treatments and the duration of detrimental QOL impacts is increasingly important. Furthermore, some patients who experience detrimental QOL from treatment might have been cured through surgery alone or derive similar benefits from treatment if and/or when they recur.

Acknowledging the differences in biology and outcomes between cancer types and variations in patients’ preferences,^9,10 cross-tumor comparisons of the design and reporting of (neo)adjuvant clinical trials may provide insights into what constitutes success and value for patients across oncology. In this review, we provide an in-depth analysis of the design and reporting of (neo)adjuvant clinical trials in oncology supporting US Food and Drug Administration (FDA) registration between 2010 and 2023, examining statistical design and reproducibility, QOL, and censoring.

Methods

Study inclusion criteria

We searched Drugs@FDA to identify trials supporting cancer drug approvals for solid organ malignancies (excluding lymphoma) between January 2010 and December 2023. We included adjuvant and neoadjuvant trials with a time-to-event endpoint. We excluded trials in the metastatic setting and single arm trials.

Data extraction and synthesis

Data extraction was performed by 3 authors (C.M.V., A.M., and B.E.W.). For each approval, the FDA labels were reviewed, and data were extracted from the corresponding peer-reviewed publications, supplemental appendices, and protocols. The primary and subsequent publications associated with each approval were identified through MEDLINE (host: PubMed). The European Society for Medical Oncology (ESMO) Magnitude of Clinical Benefit (MCBS) score associated with each drug approval was extracted from the ESMO website.¹¹

Food and Drug Administration approval and trial characteristics

For each trial, we extracted, drug name, drug class (grouped into immunotherapy, monoclonal antibodies [excluding immunotherapy], small molecular therapies, and other [including chemotherapy and drug combinations]), approval date, FDA designations (Priority Review, Breakthrough Therapy, Orphan Drug, Fast Track), indication, tumor site, treatment intent (neoadjuvant, adjuvant, or perioperative), study phase, blinding, total sample size, randomization ratio, and type of control arm. For the primary endpoint and overall survival endpoints, we extracted the hazard ratio (HR), 95% confidence interval, and P value for time-to-event endpoints and the absolute difference in outcome between experimental and control groups.

Clinical trial design: sample size, power, and anticipated magnitude of effect

From the protocol, supplemental appendices, and/or manuscript, we extracted the sample size calculation including planned power, expected hazard ratio, alpha, and estimated sample size based on the expected number of events and expected absolute numerical outcomes in the experimental and control groups. The statistical plans were evaluated as follows: (1) was the expected number of events provided; (2) could the number of events calculation be reproduced; (3) was the hypothesized primary outcome in the experimental and control groups provided; (4) if not provided, could it be calculated using the available information; and (5) could the expected sample size calculation be reproduced. Our goal was not to verify exact numerical agreement but rather to evaluate design assumptions, transparency, and the clinical relevance of the study at the time of design.

Censoring, sensitivity analysis, and missing data plans

From the protocol and supplemental appendices, we extracted information on censoring rules and planned sensitivity analysis. Adapting from similar work in the metastatic setting,⁸ we considered censoring rules sufficient if they reported (1) censoring rules for participants starting a new anticancer therapy and (2) censoring rules for participants with at least 2 sequential missed outcome assessments.

From the CONSORT diagrams, we quantified the proportion of patients with early drug discontinuation and withdrawal of consent or loss to follow-up. Early drug discontinuation may be a metric of drug tolerability and acceptance for patients treated in the (neo)adjuvant setting. We defined early drug discontinuation as drug cessation before completion of the planned protocol duration for any reason other than disease recurrence or death. Using the CONSORT diagrams, supplemental appendices, and Kaplan–Meier curves, we quantified the number of censored patients for each study.

Quality of life data

Protocols were reviewed to determine whether there was a prespecified QOL endpoint, which QOL scales and instruments were used, the assessment timepoints, and the planned QOL analysis. We determined whether QOL results were published in the primary article or separate publication or presented in conference format. Any changes in the primary QOL endpoint or any key QOL domains identified by the study were extracted. In instances of deterioration, we documented whether it pertained to global QOL or a subdomain.

Statistical analysis

Data were reported descriptively as proportions, medians, and ranges. We compared the mean paired differences in the proportion of withdrawal of consent or loss to follow-up and early drug discontinuation between the experimental and control arms. All analyses were performed using STATA version 12.0 (StataCorps LP, College Station, TX, USA). No inferential statistical testing was planned in this study because of the small sample size of studies limiting statistical power.

Results

Cohort selection and trial characteristics

Among the 316 FDA approvals for solid tumor malignancies during the study period, 31 were in the curative-setting. Six were excluded (3 approvals supported by a single-arm trial,^12-14 1 for nonmetastatic prostate cancer with biochemical recurrence,¹⁵ and 2 for non–time-to-event primary endpoints [pathologic complete response¹⁶ and response rate¹⁷]). One study supported 2 approvals,¹⁸ resulting in a final cohort of 25 FDA approvals supported by 24 studies (Figure S1).

Characteristics of the 24 included studies are provided in Table 1. Most (n = 21, 88%) studies were conducted in the adjuvant setting, 2 (8%) in the neoadjuvant setting, and 1 (4%) in the perioperative setting. The most common tumors were melanoma (29%), breast (25%), and non-small cell lung cancer (25%). The median sample size of breast trials was 2877 compared with 795 for all other cancer types. No trials used overall survival as a primary or coprimary endpoint. Overall, 21 (88%) trials reported overall survival data at any timepoint; however, only 10 trials reported overall survival data at the time of FDA approval and/or primary publication, and in some cases these data were immature. Eleven (52%) trials demonstrated an overall survival benefit at any time, but only 4 reported an overall survival improvement at the time of primary publication and/or FDA approval. Quality of life was a secondary endpoint in 21 (88%) trials. Most (19 of 24) studies scored A on the ESMO MCBS v1.1 online tool, indicating high value (Table 1).

Table 1.

Design characteristics and endpoint definitions of the included (neo)adjuvant clinical trials.

	No. (%)
Primary endpoint
Event-free survival, recurrence-free survival, disease-free survival	23 (96)
Progression-free survival	1 (4)
Overall survival reported endpoint, any time	21/24 (88)
At the time of FDA approval and/or primary publication	10/21 (48)
Later publication	11/21 (52)
Statistically significant overall survival benefit demonstrated, any time	11/21 (52)
At the time of FDA approval and/or primary publication	4/11 (36)
Later publication	7/11 (64)
Sponsor
Industry	18 (75)
Cooperative	2 (8)
Both	4 (17)
Blinding
Blind, single or double	18 (75)
Open-label	6 (25)
Year of approval
2010-2012	2 (8)
2013-2015	1 (4)
2016-2018	6 (25)
2019-2021	9 (38)
2022-2023	6 (25)
Approval designation
Orphan drug	8/24 (33)
Breakthrough therapy	6/24 (25)
Priority review	15/24 (63)
Approval based on subgroup	3 (13)
Drug class
Immunotherapy	14 (59)
Monoclonal antibodies	2 (8)
Targeted therapy	7 (29)
Other, peginterferon	1 (4)
Malignancy site
Breast	6 (25)
Lung	6 (25)
Melanoma	7 (29)
Gastrointestinal	2 (8)
Genitourinary	3 (13)
Approval setting
Adjuvant	21 (88)
Neoadjuvant	2 (8)
Perioperative	1 (4)
Control group
Active control	9 (37)
Placebo or best supportive care alone	15 (63)
ESMO MCBS
Score A	19 (79)
Score B	2 (8.3)
Score C	1 (4)
No evaluable benefit	1 (4)
Score 4a	1 (4)

Abbreviations: ESMO = European Society for Medical Oncology; FDA = Food and Drug Administration; MCBS = Magnitude of Clinical Benefit.

^aThis study on perioperative treatment for stage III lung cancer was scored according to the palliative ESMO MCBS.

Statistical design, reproducibility, and transparency

Protocols were available for 92% (n = 22) of studies. The expected number of events needed to perform the statistical analysis for the primary endpoint could be calculated (exactly or to close approximation) in 88% (n = 21) of studies and could not be reproduced in 3 (Tables S1 and S2). In 46% (n = 11), the expected sample size could be calculated with the available information, but in most (n = 10) the calculated number did not match the sample size reported in the protocol. In 54% (n = 13), there was insufficient information on the required assumptions in the protocol and associated publications to attempt a sample size calculation.

Only 58% (n = 14) of studies reported expected absolute numerical outcomes in the experimental and control groups, a key element for understanding the potential clinical relevance of any trial. We were able to estimate the expected outcomes for an additional 4 studies. The expected magnitudes of benefit the studies were designed to detect varied widely across tumor types. Lung cancer studies were designed to detect an absolute difference in median event-free survival or disease-free survival (DFS) ranging from 9 to 24 months. For melanoma, recurrence-free survival timepoint analyses at 3-5 years anticipated benefits ranging from 10% to 13% or an absolute recurrence-free survival gain of 6 months. For renal cell carcinoma RCC, urothelial, esophageal, and gastrointestinal stromal cell tumors, study protocols reported expected median DFS gains of 10-22 months or 2- to 5-year DFS landmark gains ranging from 9% to 16% (Table S3). In comparison, breast cancer studies were designed to detect relatively small improvements in DFS/event-free survival at 3-5 years, ranging from 2.9% to 6.5%.

Censoring rules, sensitivity analysis, and missing data

Censoring rules were defined adequately in the trial protocols in 36% (n = 8 of 22 studies; see Table 2). While 68% (n = 15) planned to censor patients upon initiation of a new anticancer therapy, 32% (n = 7) provided insufficient information. Approximately 27% (n = 6) of studies censored patients once they had missed at least 2 scheduled assessments, meaning that any events occurring after the censoring time would not be included in the analysis, and 50% of studies provided insufficient information. It was unclear in all protocols whether patients who missed at least 2 scheduled assessments were also censored for overall survival assessment. Recognizing differences in definitions of surrogate endpoints (ie, DFS vs relapse-free survival), 8 studies censored patients with new primary malignancies, while 5 considered secondary cancer as an event. Most studies (91%, n = 20) described a planned sensitivity analysis regarding censoring rules. However, few (33%, 8 of 24) studies presented any of the described sensitivity analysis in publications. Only 25% (6 of 24) provided clear information on censoring in lifetables associated with the Kaplan–Meier curves. No studies discussed the impact of competing events in (neo)adjuvant trials in oncology.

Table 2.

Reporting of withdrawal of consent or lost to follow-up, early drug discontinuation, and informative censoring in adjuvant US Food and Drug Administration approvals between 2010 and 2023.

	No.
Was the protocol available for review?	(n = 24)
Yes	22
No	2
Were censoring rules adequately defined in the protocol?	(n = 22)
Yes	4
No	18
Were patients censored at initiation of new anticancer treatment for surrogate endpoints?	(n = 22)
Yes	15
No	0
Unclear, not specified	7
Were patients censored in the primary efficacy analysis if they missed more than 2 visits and progressed or died thereafter?	(n = 22)
Yes	6
No	3
Unclear, no information	13
Were patients with new primary malignancies censored?	(n = 22)
Yes	8
No	7
Unclear, not specified	7
Was a sensitivity analysis for the surrogate endpoint regarding alternate censoring rules described in the protocol or paper?	(n = 22)
Yes	20
No	1
Unclear, no information on censoring available	1
Was censoring explicitly reported in the publication (eg, a table showing the number censored by arm and/or timepoint)?	(n = 24)
Yes	6
No	18
Were treatment discontinuation and patient outcomes (event vs censored) separately reported at final analysis?	(n = 24)
Yes	7
No	17

Among the 24 included studies, 23 provided sufficient information on causes of early drug discontinuation and withdrawal of consent or loss to follow-up (Table 3). For one perioperative study, the design complicated the summation of early drug discontinuation and withdrawal of consent or loss to follow-up, and the study was excluded.¹⁹ The proportion of patients withdrawal of consent or loss to follow-up in the experimental and control groups was similar in most studies (median paired difference = 0%, mean paired difference [SD] = 0.03% [1.3%]), and only 5 showed a more than 2% difference in the proportion withdrawal of consent or loss to follow-up. The proportion of patients with early drug discontinuation for all causes other than recurrence or death was numerically higher in the experimental group in all but 2 studies (mean paired difference = 14%, range = <38% to ≥47%) (Figure S2). Most early drug discontinuation was for adverse events, which were higher in the experimental arm (median paired difference = 11.9%, range = <37% to ≥47%) (Figure S3). A higher proportion of patients discontinued drug in the control arm for progression (mean paired difference = −9.9%) (Table 3).

Table 3.

Comparison of WCLFU and causes of EDD between experimental and control arms.

	Experimental arm	Control arm	Paired difference (experimental—control), %
Proportion WCLFU
Mean (SD), %	1.9 (2.4)	2.0 (2.2)	0.03 (1.3)
Median (range), %	0.4 (0-8)	1.2 (0-6.9)	0 (−3.2 to 2.5)
Proportion not treated
Mean (SD), %	0.9 (0.8)	0.9 (1.0)	0.004 (1.1)
Median (range), %	0.8 (0-3)	0.6 (0-4.1)	0.2 (−3.5 to 1.8)
Proportion EDD
Mean (SD), %	26.0 (10.9)	11.9 (9.3)	14.1 (15.9)
Median (range), %	25.5 (5.6-56)	11.1 (2-51)	15.5 (−38 to 47)
Discontinued for death
Mean (SD), %	0.13 (0.27)	0.09 (0.2)	0.04 (1.8)
Median (range), %	0 (0-1.1)	0 (0-0.76)	0 (−0.5 to 0.42)
Discontinued for objective progression
Mean (SD), %	13.9 (11.8)	23.8 (21.0)	−9.9 (15.8)
Median (range), %	10.3 (0-34.8)	19.3 (0-75)	−7.2 (−49.5 to 22.0)
Discontinued for adverse event
Mean (SD), %	18.3 (9.9)	6.5 (9.3)	12.0 (15.0)
Median (range), %	17.9 (5.6-51.6)	4.7 (0-47)	11.9 (−36.9 to 47.4)
Discontinued for other
Mean (SD), %	1.5 (2.3)	1.5 (2.0)	0.01 (1.6)
Median (range), %	0.7 (0-9.4)	0.9 (0-7.6)	0 (−5.7 to 3.8)

Abbreviations: EDD = early drug discontinuation; WCLFU = withdrawal of consent or lost to follow-up.

Quality of life

Quality of life was a defined secondary endpoint in 88% (21 of 24) of studies (Table 4). The most frequently used QOL scales were the EORTC QLQ-C30 (European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire - Core 30), EQ-5D, and FACT (functional assessment of cancer therapy). The reporting and planned statistical analysis of QOL was not well described in protocols. For studies with QOL as a secondary endpoint, no protocols clearly described what aspect of QOL was of primary interest (ie, global QOL, specific subdomain). Most protocols defined the minimal clinically important difference (typically a 10-point change on a 100-point scale). However, no protocol adequately described how the analysis would be conducted (ie, mean change in the group or paired changes by patient). The duration of QOL assessments also varied. Twelve trials ceased QOL evaluations at a specified timepoint (12-60 months after random assignment or treatment completion depending on the study), 3 trials stopped assessments upon treatment completion or discontinuation, 1 extended evaluation until the end of the study follow-up, 2 continued until confirmed progression and/or the initiation of a new cancer therapy,^20,21 and 2 trials did not specify when QOL assessments were discontinued.

Table 4.

Reporting, measurement, and impact of QOL outcomes in (neo)adjuvant trials.

	No. (%)a
QOL was a secondary endpoint	21/24 (88)
Commonly used QOL scales
EORTC QLQ-C30	14/21 (67)
EQ-5D	16/21 (76)
FACT	4/21 (19)
QOL data published	19/24 (79)
Main paper	6/19 (31)
Separate paper	10/19 (53)
Abstract presented at an oncology meeting	3/19 (16)
Did the investigational agent provide clinically meaningful improvements in QOL?	0/19 (0)
The investigational agent led to a significant decline in overall QOL	2/19 (11)
The investigational agent led to a significant decline in a specific QOL element	4/19 (21)
The investigational agent led to significant decline in global or specific symptom QOL without improving overall survival	5/6 (83)

Abbreviation: QOL = quality of life.

^aDenominators reflect the number of trials that published QOL data (19 of 24), not all trials with a planned QOL endpoint (21 of 24).

Among the 79% (19 of 24) of studies that published QOL data, 31% (n = 6) reported QOL results in the main trial publication, 53% (n = 10) published them in a separate article, and 16% (n = 3) presented results as an abstract at a professional meeting. None of the trials demonstrated a clinically meaningful improvement in QOL. In 31% (6 of 19), the experimental agent led to a clinically meaningful decline in at least 1 aspect of QOL, and only 1 of these studies demonstrated an improvement in overall survival. Two demonstrated a clinically significant detriment in overall QOL, and 4 reported declines in specific QOL-related symptoms (eg, diarrhea, appetite loss, insomnia) (Table S4). The duration of QOL declines was not reported clearly in any trials.

Discussion

In this comprehensive analysis of (neo)adjuvant clinical trials in oncology with time-to-event endpoints supporting regulatory approval, we identified several challenges in trial design, endpoint selection, and the reporting and handling of censoring. Most studies used an intermediate endpoint, and approximately half had reported any overall survival outcomes at the time of the FDA approval and/or primary publication. Quality of life was a secondary endpoint in most trials, and only one-third reported a decline in QOL. Although most studies included in this cohort had high ESMO MCBS scores, our analysis provides complementary information to this metric of clinical value by highlighting underreported aspects of trial design and transparency such as censoring and reproducibility that may impact the interpretation of clinical benefit. Finally, there was incomplete reporting of the expected absolute improvements in the primary endpoint, making it difficult for patients and doctors to understand the absolute benefits that the trial was designed to detect.

Cross-tumor comparisons within this cohort highlight how standards for (neo)adjuvant trial design vary across tumor types. The average size of clinical trials in breast cancer was substantially larger than any other tumor types. Outcomes in breast cancer are generally more favorable, and events are observed typically over a longer period of follow-up than in other tumor types. Therefore, a target hazard ratio translates to a smaller absolute improvement in outcome compared with the same hazard ratio applied to a disease with poor prognosis, such as lung cancer, may translate to a much larger absolute improvement. Trial designs must balance feasibility with patient preferences, ensuring the benefits they aim to detect are clinically meaningful for patients and health systems.²² Patient preferences vary depending on how questions are asked, demographics, socioeconomic status, and level of education.²³ Research has shown that patients are often willing to accept chemotherapy for modest survival gains, though preferences can vary.^24,25 Patients often have limited knowledge of the definitions of intermediate endpoints, and surveys have shown that patients tend to expect that all endpoints are a variation of living longer.²⁶ In the metastatic setting, when patients are fully informed of the definition of progression-free survival, more than one-half would decline treatment if not also associated with improved overall survival.²⁷ Similar work evaluating patient’s perspectives and understanding of surrogate endpoints in the (neo)adjuvant setting is needed.

Quality of life assessments lacked clarity on primary objectives and planned statistical analysis. Unsurprisingly, among this cohort of (neo)adjuvant trials where patients are being treated with curative intent, none demonstrated a clinically meaningful improvement in QOL. Somewhat reassuringly, only 31% reported clinically meaningful declines in QOL associated with the experimental agent. However, no trials adequately described the duration of these negative QOL effects, or whether they were reversible. This is increasingly important in trials evaluating (neo)adjuvant immunotherapy, where chronic toxicities such as endocrinopathies may have long-term detrimental impacts on patient QOL. Among the trials showing a decline in QOL, only 1 demonstrated a statistically significant improvement in overall survival. Further exploration is needed to understand how patients in the (neo)adjuvant setting value QOL declines in exchange for improvements in surrogate endpoints without proven survival advantage, particularly when there are risks of chronic and lasting toxicities.

Although not specific to the adjuvant setting, in recent work using discrete choice experiments, patients showed a strong preference for high certainty of survival benefit and were willing to wait almost 2 years for stronger evidence to emerge.²⁸ Although overall survival data were ultimately reported in most trials, more than half of the studies had not provided overall survival data at the time of FDA approval and/or primary publication. Where overall survival is a secondary endpoint, studies may be underpowered to detect statistically significant differences, even if the effect size is clinically important. Overall survival may also be difficult to demonstrate for cancers with long natural history and multiple metastatic treatment options. In such cases, research should focus on understanding which intermediate endpoints are most meaningful for patients and prioritizing those with the strongest surrogacy with overall survival.29-31 For example, DFS is the most frequently used endpoint for adjuvant trials in breast cancer. However, whether patients value metastasis-free survival or distant DFS over DFS requires further evaluation. For other tumor types, such as lung cancer, where prognosis remains relatively poor and 5-year survival rates are modest in patients with stage II and III disease, collecting and reporting overall survival data should be strongly encouraged. However, even in such trials, overall survival may take 7-10 years to report, substantially increasing study costs and the risk of loss to follow-up.

Statistical plans often lacked the transparency to reproduce the 3 main elements of interest: sample size, number of expected events, and expected magnitudes of benefit. In keeping with prior recomendations,³² statistical protocol for any (neo)adjuvant trial should include the planned power and alpha, median follow-up, the number of events required to trigger interim and final analyses, and the expected primary outcomes in the control and intervention arms, expressed in absolute measures wherever possible, anticipated hazard ratio, total sample size, accrual assumptions, and the method used for the sample size calculation. These elements are essential to understanding whether the trial is designed to detect a clinically meaningful difference. Although the relative magnitude of benefit (ie, the expected hazard ratio) was provided in all studies and forms the basis of power calculations and trial designs, it contributes little information regarding the underlying risk assumptions and anticipated outcomes. These assumptions are fundamental for clinicians, patients, and health systems to determine whether a clinical trial is designed to answer a clinically meaningful question. Relative benefits can be misleading–a study powered to detect a relative risk reduction of 50% (ie, a target HR of 0.5) could equate to a reduction in disease recurrence from 50% to 25%, which is clinically meaningful, or a reduction from 2% to 1%, which most would perceive as low value. Whereas our focus was on statistical protocols, aligning these elements with transparent and comprehensive reporting in published articles reporting trial results is equally important. These recommendations complement broader recommendations for transparent clinical trial reporting within trials and protocols found in the recently published CSO RCT Common Sense Oncology Randomized Controlled Trial checklist.³³

Finally, censoring rules were poorly defined in many protocols and poorly reported in most trial publications. Recent work by Vitale et al.³⁴ similarly highlighted poor reporting of censoring patterns in randomized trials of immunotherapy in advanced cancers and demonstrated imbalances in censoring between treatment arms. In contrast, our study examines (neo)adjuvant trials supporting regulatory approval across multiple types of systemic therapy and extends beyond reporting to evaluate prespecified censoring rules, endpoint definitions, competing events, and the transparency of statistical protocols. Including censoring data in a table, rather than in a lifetable format, hinders the ability to evaluate the timing of censoring events (ie, unbalanced early censoring will create more bias in the results than late censoring). Patients who are censored may be more unwell than those who remain in the trial, which can bias the results.^35,36 An extreme example of the potential consequences of censoring are illustrated in the MonarchE trial.¹⁸ When a sensitivity analysis was performed, assuming all censored patients experienced an event (analogous to time to treatment failure), the benefit of the treatment in the intervention arm was attenuated.³⁷ In the current analysis, one-third of studies censored patients if they missed at least 2 scheduled appointments, and half provided insufficient information. Missed appointments may affect the ability to accurately collect the timing of recurrent or progressive events, which may introduce bias, and examples in the metastatic setting illustrate the impact on absolute effects at set timepoints (eg, median progression-free survival) where estimates can vary based on whether censoring occurs prior to or after the set timepoint.³⁸ However, we believe that missed appointments should not lead to censoring for overall survival endpoints in the neoadjuvant setting, and additional clarity regarding censoring rules for overall survival data are needed. In (neo)adjuvant trials, competing events may also substantially influence the interpretation of treatment effects. Their presence and impact depend on how the outcome and estimand are defined. Data on competing events were not well reported in our cohort of trials. When cancer-specific survival is used as an endpoint, noncancer mortality constitutes a competing event. Similarly, for longitudinal QOL outcomes, death from any cause precludes further assessment and should be treated as a competing outcome rather than as noninformative censoring. Although harmonization in the handling and reporting of censoring and intercurrent events could facilitate comparisons across trials, differences may be appropriate depending on the primary endpoint and estimand of interest. Instead, we emphasize the need for transparent reporting of censoring rules and the handling of intercurrent events and the use of sensitivity analyses when appropriate.

This study has limitations. By focusing on (neo)adjuvant studies with FDA approval, the sample size of the cohort is small, and the findings cannot be generalized to the metastatic setting or (neo)adjuvant trials not supporting FDA approval. The small number of neoadjuvant studies in our cohort limited comparisons between adjuvant and neoadjuvant study designs. Our analysis focused on censoring rules, and further research is needed to evaluate the impact of unbalanced censoring on trial outcomes. Finally, we focused our data extraction on the primary publications and supplemental appendices from these publications. However, we recognize that on occasion, analyses presented to regulators for approval may differ because of updated data or additional sensitivity analyses. To mitigate this, we also extracted data from any subsequent publications after FDA approval. However, our analysis was limited to data available in the scientific literature, which may be incomplete.

This study of (neo)adjuvant oncology trials with time-to-event endpoints highlights challenges in trial design, reporting, and interpretation. Improved transparency across multiple domains are needed (including statistical design, censoring, and QOL). The oncology community should continue to advocate for well-designed studies to detect meaningful absolute improvements in outcome. Future research should also focus on improved reporting of QOL data in the (neo)adjuvant setting and understanding how patients value alternative endpoints in the (neo)adjuvant setting to inform future trial designs.

Author contributions

Consolacion Molto Valiente (Conceptualization, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing), Abhenil Mittal (Conceptualization, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing), Dongsheng Tu (Formal analysis, Methodology, Writing—review & editing), Eitan Amir (Investigation, Writing—review & editing), Joseph Pater (Investigation, Writing—review & editing), Janet Dancey (Investigation, Writing—review & editing), Bishal Gyawali (Investigation, Writing—review & editing), Christopher M. Booth (Investigation, Writing—review & editing), and Brooke Wilson (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing)

Supplementary material

Supplementary material is available at JNCI: Journal of the National Cancer Institute online.

Funding

No funding was provided for this work.

Conflicts of interest

Brooke E. Wilson reports consulting fees or honoraria from AstraZeneca, Gilead, and Novartis and an institutional grant from BMS, outside of the submitted work. Consolacion Molto Valiente reports consulting fees or honoraria from AstraZeneca and Merck, outside of the submitted work. Abhenil Mittal reports consulting fees or honoraria from Gilead, Janssen, Roche, BMS, AstraZeneca, and Knight Therapeutics, outside of the submitted work. Eitan Amir reports consulting fees from Novartis (to institution) and from Gilead (to self). Bishal Gyawali, who is a JNCI associate editor and co-author on this article, was not involved in the editorial review or decision to publish the manuscript.

Data availability

The data that support this work were extracted from publications that are publicly available. The database generated through this work is available from the corresponding author upon reasonable request.