- Journal List
- HHS Author Manuscripts
- PMC8478711
As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsem*nt of, or agreement with, the contents by NLM or the National Institutes of Health.
Learn more: PMC Disclaimer | PMC Copyright Notice
Clin Trials. Author manuscript; available in PMC 2021 Oct 1.
Published in final edited form as:
Clin Trials. 2021 Oct; 18(5): 562–569.
Published online 2021 Aug 21. doi:10.1177/17407745211028872
PMCID: PMC8478711
NIHMSID: NIHMS1714527
PMID: 34420417
Subodh Selukar,1 Susanne May,1 Dave Law,2 and Megan Othus3
Author information Copyright and License information PMC Disclaimer
The publisher's final edited version of this article is available at Clin Trials
Associated Data
- Supplementary Materials
Abstract
Background:
Platform trials facilitate efficient use of resources by comparing multiple experimental agents to a common standard of care arm. They can accommodate a changing scientific paradigm within a single trial protocol by adding or dropping experimental arms - critical features for trials in rapidly developing disease areas such as COVID-19 or cancer therapeutics. However, in these trials, efficacy and safety issues may render certain participant subgroups ineligible to some experimental arms, and methods for stratified randomization do not readily apply to this setting.
Methods:
We propose extensions for conventional methods of stratified randomization for platform trials whose experimental arms may differ in eligibility criteria. These methods balance on a prespecified set of stratification variables observable prior to arm assignment that remains the same across experimental arms. To do so, we suggest modifying block randomization by including experimental arm eligibility as a stratifying variable, and we suggest modifying the imbalance score calculation in dynamic balancing by performing pairwise comparisons between each eligible experimental arm and standard of care arm participants eligible to that experimental arm.
Results:
We provide worked examples to illustrate the proposed extensions. In addition, we also provide a formula to quantify the relative efficiency loss of platform trials with varying eligibility compared to trials with non-varying eligibility as measured by the size of the common standard of care arm.
Conclusions:
This article presents important extensions to conventional methods for stratified randomization in order to facilitate the implementation of platform trials with differing experimental arm eligibility.
Keywords: Platform Trials, Stratified Randomization, Varying Eligibility, COVID-19, SARS-CoV-2, Oncology
Background
The novel coronavirus COVID-19 spurred an unprecedented effort for rapid discovery of effective therapeutic and preventative agents. Some of these efforts involve the use of platform trials (I-SPY COVID-19, NCT04488081; DisCoVeRy, NCT04315948; RECOVERY, NCT04381936; SOLIDARITY, EudraCT Number 2020–000982-18), a type of master protocol design that compares multiple experimental arms to a common standard of care arm. Authors have previously described how to add and drop arms within a single protocol, making the design appealing for rapidly developing areas.1,2
In oncology, the LEAP trial (NCT03092674)3 was a platform trial for acute myeloid leukemia (AML), conducted by the NCI-funded SWOG Cancer Research Network. The trial evaluated several therapies with varying mechanisms of action, including immunotherapies and targeted agents. Because of these varying mechanisms, certain participant sub-populations were thought to potentially be harmed by specific therapies, so eligibility necessarily differed across arms. For example, patients with pre-existing autoimmune diseases were precluded from receiving checkpoint inhibitor immunotherapy. However, trial investigators did not want to modify trial-wide eligibility to address this, as a key desire for the trial was to be as inclusive as possible. Allowing for varying eligibility did not impact trial conduct beyond requiring appropriate treatment assignment and, thus, randomization. Other trial examples have acknowledged varying eligibility, but they did not detail their procedures for treatment assignment, and authors have called for more research on this issue.2,4,5
Investigators may wish to perform stratified randomization to ensure that important baseline prognostic factors remain balanced across study arms. Further, adjusting for stratification variables may also increase power during analysis.6 Existing methods for stratified randomization assume that all arms share the same eligibility criteria. As such, utilizing these methods requires applying the most stringent set of eligibility criteria: the intersection across every study arm’s eligibility. This may limit trial accrual and may hinder inference to the appropriate target populations.7,8
In this paper, we propose extensions to existing methods of both block randomization and dynamic balancing to appropriately perform stratified randomization in the setting of varying eligibility. We provide worked examples of both approaches, and we also briefly describe the efficiency (in terms of the size of the common standard of care arm) of platform trials when experimental arms differ in participant eligibility.
Methods
Extending Existing Methods for Stratified Randomization
In the analysis of a trial with differing experimental arm eligibility, each experimental arm is compared to the subset of participants in the standard of care arm who meet the experimental arm’s eligibility criteria. For example, consider a trial with three experimental arms, E1, E2 and E3. A participant randomized to the standard of care arm and eligible to experimental arms E1 and E2 but not E3 could be used in comparisons with E1 and E2 but is not used in a comparison with the experimental arm E3. This raises a difficulty of treatment assignment: how should we appropriately assign participants to maintain a desired allocation ratio, achieve balance across specified stratification factors and also accommodate this varying eligibility? This section outlines the modifications we propose to allow existing methods in block randomization and dynamic balancing to address this problem.
Throughout this paper, we consider a platform trial defined as a multi-arm trial with a single, common standard of care arm and at least two experimental arms that may start and/or end at different time points. We want to achieve balance on a prespecified set of stratification variables observable prior to arm assignment that remains the same across experimental arms. We assume all participants are eligible to the standard of care arm and at least one experimental arm for enrollment, and subjects are assigned to exactly one study arm. We describe 1:1 randomization in the examples, but each of the two proposed methods easily accommodates other allocation ratios in a manner corresponding to what would be done for the respective method without differing eligibility.
Block Stratified Randomization
In block stratified randomization, investigators create blocks (of fixed or varying size) for each stratum defined by the stratification variables and fill the blocks with random treatment assignments such that balance is achieved between arms within each block. A newly enrolled participant is matched with the corresponding stratum block and assigned to the next unassigned treatment.
We propose adding arm eligibility as an additional stratification factor to accommodate arms with differing eligibility. In other words, each possible combination of experimental arm eligibility contains its own nested set of the prespecified strata. With three experimental arms and differing eligibility criteria for each arm, a participant could be eligible to only one of the three, two of the three or all three arms, thus representing 7 (= 23 − 1) experimental arm eligibility strata. Additionally, these eligibility strata would themselves contain strata based on the prespecified stratification variables.
This simple extension addresses the problem of differing eligibility while achieving balance and targeting the desired allocation ratio. However, the number of possible eligibility combinations increases exponentially with the number of experimental arms. A trial with three experimental arms and just one binary stratification variable would have 14 strata, and many trials may desire more experimental arms and/or stratification factors.
Importantly, while the maximum number of eligibility combinations for K experimental arms is 2K − 1, this does not necessarily represent the number eligibility strata for a given trial, as these depend on exactly how the experimental arms differ in eligibility. To illustrate this, we describe two examples. First, consider a trial with two experimental arms. In addition to any trial-wide eligibility criteria, the first experimental arm (arm A) only recruits subjects over the age of 65 and the second (arm B) only recruits biomarker positive subjects. In this case, subjects can be eligible to both arms (older than 65 and biomarker positive) or exactly one (either older than 65 or biomarker positive but not both), so the number of eligibility combinations is fully 22 − 1 = 3. Suppose, instead, arm B had no additional restrictions beyond the trial-wide eligibility. In that case, while there are three possible combinations, only two eligibility strata are needed: (1) eligible to A and B or (2) only eligible to B, as no subjects would be only eligible to A. To expand on this, consider a different trial with three experimental arms (E1, E2 and E3). Beyond trial-wide eligibility criteria, E1 only recruits subjects younger than 65, E2 only recruits subjects younger than 75, but E3 has no additional restrictions. This results in just three eligibility strata: (1) eligible to E1, E2 and E3 (for subjects younger than 65), (2) eligible to E2 and E3 but not E1 (subjects aged 65–74), or (3) only eligible to E3 (subjects aged 75 and older).
Dynamic Balancing
A general scheme of dynamic balancing9 involves calculating the imbalance caused by provisionally assigning a participant to each eligible study arm and then using these imbalance scores to weight the participant’s randomization toward the arm that results in the least imbalance. As each new participant enters the study, this procedure is repeated. The method requires specification of how to compute an “imbalance score” and how to use these imbalance scores to weight study arm assignment.
To accommodate differing arm eligibility, we propose that investigators modify an existing dynamic balancing scheme’s imbalance score calculation. Consider pairwise calculations for each experimental arm and participants assigned to the standard of care arm who were eligible for that experimental arm. These pairwise calculations can be summarized into a single imbalance score (e.g., the maximum across the pairwise calculations) for adding the new participant to a given eligible study arm analogous to an imbalance score from the underlying dynamic balancing scheme.
For example, consider imbalance measured by differences in counts, defined as , of subjects of the same stratification factor level as a new participant. For each eligible experimental arm Ej (j = 1, 2, …) and corresponding standard of care subset we can tally the number who have the same stratification factor level and add to the tally the new participant. Then we conduct pairwise comparisons with the absolute differences . The imbalance score for adding the new participant to a given arm could then be the maximum of these differences. (We provide full details for this scheme in the Online Appendix.) Existing methods, which do not allow for varying eligibility, only calculate one count for the standard of care arm (T(C)) and that same count is used in all the difference calculations, e.g. .
The calculated imbalance scores may then be mapped to randomization weights as in dynamic balancing without varying eligibility criteria. As these weights can accommodate a desired allocation ratio, this method also resolves the issue of varying eligibility while achieving balance and targeting the desired allocation ratio.
Stratified Randomization when Adding or Dropping Experimental Arms
Dropping an experimental arm does not require modifications to the above procedures. The dynamic balancing algorithm would no longer compute an imbalance score for adding to the dropped arm and would not compute pairwise differences with that arm. The block randomization method would discontinue the strata that include the dropped arm. For example, if an arm E3 were dropped with E1 and E2 continuing, then strata for eligibility to E1 only, E2 only or both E1 and E2 would continue, but other strata that include eligibility to E3 would not be used for future participants.
However, the above methods do require further specification after the addition of an experimental arm. The methods involve the eligibility and arm assignment of previously-randomized participants, so we must specify how each algorithm accommodates these existing participants but also ensures only concurrently randomized participants are compared to account for potential changes in participant characteristics over time (e.g., more recent participants may have systematic differences in prior care compared to those enrolled earlier).
In order to only analyze concurrently randomized participants (i.e., not include prior standard of care participants in the analysis of a new arm), we recommend not using previously-randomized standard of care participants in randomization calculations for the new arm - even if they would have been eligible for the new arm. But to ensure maximal use of the existing data, calculations involving any continuing arms should continue to include previously-randomized standard of care participants eligible to the continuing arms.
In block randomization, this means, upon the addition of the new arm, the creation of a new stratum for participants only eligible to the new arm and also new strata for participants eligible to the new arm and combinations of the continuing arms. For example, if a new arm E3 is added with continuing experimental arms E1 and E2, then four new strata would be added: one for participants eligible to only E3, one each for E1 and E3 or E2 and E3 and one for E1, E2 and E3 eligibility. The previously-existing strata would continue for new participants eligible to any combination of the continuing arms E1 and E2 but not E3.
This recommendation does not require any significant changes to the dynamic balancing procedure after the addition of the new arm. When a new participant is eligible to continuing experimental arms, the imbalance score calculations will continue to include all participants previously assigned to the continuing arms and those assigned to standard of care but eligible to the continuing arms. Calculations involving the new arm will only include participants assigned since the addition of the new arm and eligible to the new arm.
In the STAMPEDE trial, investigators implemented a different method for randomization after adding a new experimental arm: they restarted the stratified randomization process after the addition of each arm.4 While this seems logistically simpler, from a simulation study summarized in the Online Appendix, we conclude that the algorithm we suggest above causes less deviation from the desired allocation ratio in the balancing process. The choice may have practical impacts to statistical operating characteristics in sequential monitoring of the study.
Results
Worked Examples
Block Randomization
For block randomization, we illustrate the proposed extension with an example with one binary stratification factor of biomarker status (positive or negative). Consider a trial with a standard of care arm, C, and two experimental arms, E1 and E2. Subjects can be eligible for any combination of the experimental arms but all are eligible for C and at least one experimental arm.
In Table 1, we provide sample treatment assignments using blocks of size 6 and 1:1 randomization (other allocations are straightforward to implement). Suppose we have already randomized several participants (assignments struck through), and the next participant to be randomized is eligible to E1 and E2 and is biomarker positive. As shown in the table (in bold and italics), the new participant would be assigned E1.
Table 1.
One block of treatment assignments for participants in each of the strata defined by arm eligibility (each combination of experimental arms E1 and E2) and biomarker status (positive or negative)
| Eligible to El Only | Eligible to E2 Only | Eligible to El and E2 | |||
|---|---|---|---|---|---|
| Positive | Negative | Positive | Negative | Positive | Negative |
| El | |||||
| El | C | E2 | E2 | ||
| C | C | C | E2 | C | C |
| El | C | E2 | C | E2 | El |
| C | El | E2 | C | C | El |
| ⋮ | ⋮ | ⋮ | ⋮ | ⋮ | ⋮ |
Open in a separate window
Dynamic Balancing
To illustrate dynamic balancing, we consider a trial with three experimental arms, E1, E2 and E3, a standard of care arm, C, and one binary stratification factor for biomarker status (positive or negative). Again, subjects can be eligible for any combination of the experimental arms but all are eligible for C and at least one experimental arm.
We depict the eligibility and treatment assignments of already-randomized participants with a Venn Diagram (Figure 1a). Each point labeled E1, E2, E3 or C represents the treatment assignment of an already-randomized participant, and the circle(s) containing that point represent the eligibility of that participant. Points lying in the intersection of multiple eligibility circles indicate the participant is eligible to more than one experimental arm. To use existing randomization methods, it would require that these three circles overlap completely: in other words, all subjects must be eligible for all experimental arms.
Eligibility and Study Arm Assignment Partway through Accrual. Each currently-randomized participant is represented by their arm assignment, and their location within the circle indicates their experimental arm eligibility prior to randomization.
Figure 1a describes the eligibility and assignments partway through accrual. In this case, for example, the blue (top) circle represents eligibility for E2. One participant randomized to E2 was only eligible for E2, another was eligible for E2 and E3 and one was eligible for all experimental arms. Three participants randomized to C were eligible for E2 (two were eligible for all arms and one was eligible for E2 and E1) and would be used in imbalance calculations made for new participants eligible for E2.
Suppose the trial enrolls a new biomarker positive participant eligible to enroll on arms E1, E2 and C but not E3. In Figure 1b, we illustrate how the dynamic balancing algorithm will not incorporate the participants randomized to E3 or standard of care participants who were only eligible to E3 (struck-through). The calculation includes standard of care participants who were eligible to all arms or any combination of E1 and E2.
For this example, we implement an extended Poco*ck-Simon9 procedure. We focus on the counts of already-randomized participants of the same biomarker status and compute the imbalance from adding the new participant to each eligible study arm. To calculate the imbalance due to adding to a given arm, we do the following: For each eligible experimental arm, we compute the absolute difference between the count in the experimental arm and those in C who could have been randomized to that experimental arm. The imbalance score for adding to the given study arm is then the largest among these absolute differences. (See Online Appendix for full details.)
In Table 2, we cross-tabulate the number randomized to each eligible experimental arm and the number randomized to C who were eligible for that experimental arm by stratification level. The tallies in bold represent the participants with the same level of the stratification factor as the new participant.
Table 2.
Tallies by Arm and by Stratification Factor Level†
| E1 | CE1 | |
|---|---|---|
| Negative | 1 | 2 |
| Positive | 3 | 2 |
Open in a separate window
| E2 | CE2 | |
|---|---|---|
| Negative | 1 | 2 |
| Positive | 2 | 1 |
Open in a separate window
†The two tables provide the counts by biomarker status of those participants randomized to E1, E2 or CE1 and CE2, the standard of care arm participants who were eligible for E1 and E2, respectively. The bold numbers represent the counts for participants of the same biomarker status as the new participant.
Table 3 details the full calculations of the modified dynamic balancing procedure by examining the imbalance due to adding the new participant to each eligible study arm. The right column of each row gives the imbalance score due to the addition to a given arm. We see that adding the new participant to E1 or E2 gives an imbalance of 2, while adding to C gives an imbalance score of 0.
Table 3.
Algorithm Procedure for New Study
Open in a separate window
Open in a separate window
†Each row indicates the addition of the new participant to a given study arm. The tables show the relevant tallies modified for this addition in bold and the right column gives the imbalance score computed by adding the new participant to that study arm.
With C having the lowest imbalance score after addition of the new participant, we give it the largest weight in a weighted randomization procedure (in which the user could also specify the allocation ratio).
Efficiency of Platform Trials with Differing Arm Eligibility
Ventz et al.2 discuss the potentially dramatic reductions in sample size when conducting a platform trial compared to multiple independent two-arm studies. However, they also observe that the additional flexibility to add arms mid-study incurs an efficiency loss in sample size compared to randomizing all experimental arms initially.
We can describe the efficiency of platform trials with differing arm eligibility by evaluating the proportion of subjects randomized to the standard of care arm. In a two-arm study with 1:1 allocation, the proportion randomized to standard of care is . In a multi-arm trial with K experimental arms and equal allocation, it is . Platform trials with differing arm eligibility are maximally efficient when they resemble a multi-arm trial with equal allocation (corresponding to all shared eligibility) and minimally efficient when resembling a two-arm study (corresponding to each participant being eligible to exactly one experimental arm).
Using the law of total probability, we can fully characterize the proportion randomized to the standard of care arm when we assume balance is reached (see Online Appendix). We perform this calculation under the assumption that that all experimental arms are started and ending at the same time, but a more general framework allowing for adding and dropping arms is possible, if needed. The proportion depends on the probability of each combination of experimental arm eligibility. As expected, if the probability of subjects being eligible to all experimental arms is high, the platform trial has higher efficiency (i.e., more closely resembles a multi-arm trial with equal allocation); as this probability decreases, the efficiency decreases.
As an example, suppose a platform trial has two experimental arms, E1 and E2, and a subject’s probability of being eligible to both experimental arms is α, and the probability of being eligible for only E1 is and the probability of being eligible for only E2 is also . Each experimental arm plans to enroll 100 subjects at a 1:1 allocation with the standard of care arm.
Consider three scenarios: α1 = 1, α2 = 0 and α3 = . The scenario with α1 is a situation with no varying eligibility, while α2 has completely distinct eligibility between experimental arms, and α3 represents an intermediate scenario. The probability of being randomized to standard of care (under assumptions given in the Online Appendix), is , and , respectively.
A different way to assess the efficiency of platform trials with differing arm eligibility is by comparing the relative sample sizes of the standard of care arm (differences in the total size of the trial due to differing eligibility would only occur via the size of this arm). In the above example, the sizes of the standard of care arm would be 100, 200 and based on the planned allocation and sample sizes. With these calculations, we reach the same conclusion as above: if the probability of subjects being eligible to all experimental arms is high, the platform trial has higher efficiency.
LEAP Trial Example
In this section, we describe the decision-making and treatment assignment process for the LEAP trial (clinicaltrials.gov ID: NCT03092674), which used the method for dynamic balancing described above. As stated in the background section, the LEAP trial was a platform trial for AML: in particular, it evaluated AML therapies in medically less-fit older adults, a patient population that suffers from therapeutic resistance and reduced chemotherapy tolerance. A key goal of the trial was to be “as unrestrictive as possible and to provide treatment options for the real-world patient.”3
The trial was designed to begin with a standard of care arm, azacytidine monotherapy, and three experimental arms: (1) nivolumab in combination with azacytidine; (2) midostaurin in combination with azacytidine; and (3) decitabine in combination with cytarabine.3 (Arm 3 was only going to open after Phase 2 accrual to arms 1 and 2 was complete.) Notably, the nivolumab combination arm used a checkpoint inhibitor to activate an immune response, and the midostaurin combination was known to have possible cardiac toxicities. As such, while some patients could be ineligible to these two experimental arms, they would still be recruited if they were eligible to at least one experimental arm and the standard of care. Excluding all patients with preexisting autoimmune disease or heart risks would have limited the generalizability of study outcomes.
The study team anticipated that most patients would be eligible to all study arms, so allowing varying eligibility was not expected to substantially increase trial sample size but would allow for better generalizability and provide wider access to trial therapies.
In addition, the investigators also wanted to balance randomization on age, performance status and FLT3 mutation status, which are important prognostic variables for patients with AML.3 A chance imbalance on these variables would complicate the interpretations of the analysis, so the investigators chose to employ stratified randomization.
They opted to use dynamic balancing, as opposed to block randomization, based on the number of possible eligibility combinations, number of stratifying variables and the planned sample size set. The imbalance score calculations for LEAP were identical to the calculations used in the example for dynamic balancing above.
For randomization weights, the study employed telescoping weights of 0.75 and 0.25, if eligible to 2 arms, or 0.75, 0.1875 and 0.0625, if eligible to 3 arms, with the highest weight given to the arm with the smallest imbalance score. This choice of weights was based, in part, on a paper by Brown et al.,11 who assessed different weighting schemes of dynamic balancing. However, their paper only considered trials with two arms and a limited number of weights, and research on optimal weighting in more general settings has not been done.
The trial was closed to accrual after observing an unexpected safety signal in the nivolumab combination arm.12 The study randomized 78 subjects: 26 patients were randomized to the standard of care, 26 to the midostaurin combination arm and 26 to the nivolumab combination arm. Seven patients were not eligible for the nivolumab combination arm and one was not eligible for the midostaurin combination arm.
We observe that approximately 10% of randomized subjects (8 of 76) were not eligible to all experimental arms. In this example, we see that that differing eligibility did not substantively impact the efficiency of the trial.
Conclusions
Platform trials offer a framework to efficiently conduct randomized studies within a developing research area. In some platform trials, if experimental arms differ in eligibility, existing methods for stratified randomization cannot be naively employed. This paper proposes extensions to existing methods that properly account for differing arm eligibility.
While straightforward, the extensions we outline in this manuscript pave the way for more flexible platform trials using stratified randomization. With current methods for stratified randomization, trials must enroll only participants eligible to all experimental arms, limiting accrual and generalizability. The extensions we describe here address this issue directly and obviate such a requirement.
Our proposed methods are not intended to replace important conversations regarding the increased trial complexity by allowing varying experimental arm eligibility. As shown in this paper, such trials require adjustment to the treatment assignment process, and the complexity also affects other aspects such as trial logistics. While some scenarios may justify these complications, others may benefit from simpler, common eligibility criteria across arms. It is an important role of biostatisticians to scrutinize proposed eligibility criteria and evaluate whether the advantages of differing eligibility outweigh the increased complexity.
We note that it is valid to implement simple randomization to assign subjects to eligible treatment arms, and simple randomization is straightforward to implement. Researchers can employ stratified randomization to increase power6 or to prevent imbalance by chance on important prognostic variables, but the increased costs in implementation may not always be justified in this complex setting. Again, it is a role of biostatisticians for a given study to gauge whether the increase in computational complexity justifies the benefits of stratified randomization.
This manuscript outlines one modification for varying eligibility each for block randomization and dynamic balancing, but other possible modifications are possible. For block randomization, our recommendation dramatically increases the number of strata as the number of treatment arms increases. As such, this recommendation may only be practical with small numbers of stratification factors and/or few eligibility strata, and, while the dynamic balancing algorithm may be more difficult to implement, it does not suffer from this problem. Separately, our recommendation for dynamic balancing is flexible to many choices, but the efficiency may depend strongly on the underlying chosen dynamic balancing scheme.
We also describe the efficiency of platform trials with differing experimental arm eligibility. As the number of participants eligible to all experimental arms increases, the platform trial becomes more efficient: the size of the common standard of care arm decreases to resemble the size in a multi-arm trial with non-varying eligibility across study arms. As such, a trial may be less efficient than expected if the proportion of participants ineligible to arm(s) is higher than anticipated, and this directly impacts the planned sample size and costs for the trial. Researchers can estimate a crude maximum size of the standard of care arm by assuming participants would each be eligible to exactly one experimental arm and refine this estimate with the formula in the appendix based on anticipated participant eligibilities.
Throughout this article, we assume all experimental arms require balance on the same set of prespecified stratification factors. This serves to simplify the presentation, but careful implementation can also allow investigators to employ the proposed extensions with differing subsets of the trial’s set of stratification factors across experimental arms. This may be desirable if, for example, one arm has a small sample size that limits the number of allowable stratification factors for that arm. The implementation would require a user to first map each combination of participant arm eligibilities to the joint set of stratification factors needed for the arms of that combination. Then, the user would modify the proposed extensions by using these combinations in place of the full set of stratification factors: in block stratified randomization, use these combinations to form the nested strata within corresponding eligibility strata and, in dynamic balancing, use the combination corresponding to the new participant’s eligibility for the balancing calculations.
The implementation of these algorithms to properly address varying arm eligibility can require non-trivial effort when developing a study. But we believe that the methods we present here will be important in the design of more flexible platform trials that can accommodate differing eligibility criteria.
Supplementary Material
1
Click here to view.(156K, pdf)
Acknowledgements
The authors gratefully acknowledge the helpful feedback from the editors and referees.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Cancer Institute (NCI) [grant number T32CA09168]. In addition, this investigation was supported in part by the following PHS/DHHS grant numbers awarded by the National Cancer Institute (NCI), National Clinical Trials Network (NCTN) to SWOG: CA180888 and CA180819.
Footnotes
Declaration of conflicting interests
The authors declare that there is no conflict of interest.
References
1. Hirakawa Akihiro, Asano Junichi, Sato Hiroyuki, and Teramukai Satoshi. Master protocol trials in oncology: Review and new trial designs. Contemporary Clinical Trials Communications, 12:1:8, 2018. [PMC free article] [PubMed] [Google Scholar]
2. Ventz Steffen, Alexander Brian M., Parmigiani Giovanni, Gelber Richard D., and Trippa Lorenzo. Designing clinical trials that accept new arms: An example in metastatic breast cancer. Journal of Clinical Oncology, 35(27):3160:3168, 2017. [PubMed] [Google Scholar]
3. Walter Roland B., Michaelis Laura C., Othus Megan, Uy Geoffrey L., Radich Jerald P., Little Richard F., Hita Sandi, Saini Lalit, Foran James M., Gerds Aaron T., and et al. Intergroup leap trial (s1612): A randomized phase 2/3 platform trial to test novel therapeutics in medically less fit older adults with acute myeloid leukemia. American Journal of Hematology, 93(2), 2017. [PMC free article] [PubMed] [Google Scholar]
4. Masters L Rauchenberger M Van Looy N Diaz-Montana C Gannon M James N Maughan T Parmar MKB Brown L Sydes MR; STAMPEDE Hague D, Townsend S and FOCUS4 investigators.Changing platforms without stopping the train: experiences of data management and data management systems when adapting platform protocols by adding and closing comparisons. Trials, 20(294), 2019. [PMC free article] [PubMed] [Google Scholar]
5. Pistone T Onaisi R Sitta R Journot V-Nguyen D Peiffer-Smadja N Crémer A Bouchet S Darnaud T Poitrenaud D Piroth L Binquet C Michel JF Lefèvre B Lebeaux D Lebel J Dupouy J Roussillon C Gimbert A Wittkop L Thiébaut R Orne-Gliemann J Joseph JP Richert L Anglaret X Malvy D; COVERAGE study group Duvignaud A, Lhomme E. Home treatment of older people with symptomatic sarscov-2 infection (covid-19): A structured summary of a study protocol for a multi-arm multi-stage (mams) randomized trial to evaluate the efficacy and tolerability of several experimental treatments to reduce the risk of hospitalisation or death in outpatients aged 65 years or older (coverage trial). Trials, 21(846), 2020. [PMC free article] [PubMed] [Google Scholar]
6. Anderson G, Crowley J. Use of Covariates in Randomization and Analysis of Clinical Trials. Handbook of Statistics in Clinical Oncology. CRC Press, third edition, 2011. [Google Scholar]
7. Vermorken JB, Parmar MKB, Brady MF, Eisenhauer EA, Hogberg T, Ozols RF, Rochon J, Rustin GJS, Sagae S, and Verheijen RHM Clinical trials in ovarian carcinoma: study methodology. Annals of Oncology, 16:viii20 – viii29, 2005. 3rd International Ovarian Cancer Consensus of the GCIG, September, 2005: Germany. [PubMed] [Google Scholar]
8. George SL Reducing patient eligibility criteria in cancer clinical trials. Journal of Clinical Oncology, 14(4):1364–1370, 1996. [PubMed] [Google Scholar]
9. Poco*ck Stuart J. and Simon Richard. Sequential treatment assignment with balancing for prognostic factors in the controlled clinical trial. Biometrics, 31(1):103, 1975. [PubMed] [Google Scholar]
10. Walter Roland B., Appelbaum Frederick R., Tallman Martin S., Weiss Noel S., Larson Richard A., and Estey Elihu H. Shortcomings in the clinical evaluation of new drugs: Acute myeloid leukemia as paradigm. Blood, 116(14):2420–2428, 2010. [PMC free article] [PubMed] [Google Scholar]
11. Brown Sarah, Thorpe Helen, Hawkins Kim, and Brown Julia. Minimization—reducing predictability for multi-centre trials whilst retaining balance within centre. Statistics in Medicine, 24(24):3715–3727, 2005. [PubMed] [Google Scholar]
12. Hay Annette E., Assouline Sarit, Walter Roland B., Little Richard F., Moseley Anna, Sperling M Gail Annie Im, Foran James M., Radich Jerald P., Fang Min, and et al. Accrual barriers and detection of early toxicity signal in older less-fit patients treated with azacitidine and nivolumab for newly diagnosed acute myeloid leukemia (aml) or high-risk myelodysplastic syndrome (mds) in the swog 1612 platform randomized phase ii/iii clinical trial. Blood, 134(Supplement 1):3905–3905, 2019. [Google Scholar]
