The design of a small Bayesian hybrid-controlled trial utilizing time-to-event data from six published sources that used best alternative care to treat liver-predominant metastasis from primary ocular melanoma Authors: David Zahrieh, Xiaofei Wang, Svetomir N Markovic, Vera J Suman Published: 2026 ABSTRACT AbstractBackgroundTime-to-event endpoints, such as progression-free survival (PFS), are used to evaluate clinical activity in early-phase RCTs in cancer research. Because the power to detect a reduction in the hazard rate depends on the number of events observed, recruiting enough patients with a rare cancer can be difficult. Bayesian designs that leverage historical control data to augment the control arm are sought in rare cancers because there is a practical and ethical willingness to trade modest increases in bias and Type 1 error, particularly under prior-new data conflict, for meaningful reductions in variance and improved feasibility.PurposeThe purpose is to provide a real-world, from beginning to end application to design a time-to-event hybrid-controlled trial for a rare cancer with best alternative care (BAC) as the control that uses robust borrowing from six BAC PFS data sources and explicitly evaluates the operating characteristics across degrees of prior-new data conflict.MethodsA simulation-based approach for Bayesian sample size determination was used and calibrated to possess good frequentist Type 1 and 2 error properties.ResultsA Bayesian hybrid-controlled RCT resulted in a 28% reduction in the number of participants needed to detect a 50% reduction in the hazard rate compared with a traditional frequentist design with the same power.ConclusionsUtilizing historical BAC PFS data from 309 patients with liver-predominant metastasis from primary ocular melanoma can enhance trial efficiency by assigning fewer subjects to the control arm and potentially lead to the earlier availability of effective therapies to patients with this rare cancer.Keywords: Bayesian, Clinical trials, Meta-analytic predictive prior, Progression-free survival, Rare cancers, Time-to-event endpoint, Uveal melanoma Introduction To determine the clinical activity of experimental therapies in early-stage cancer research, time-to-event endpoints such as progression-free survival (PFS) are often used as the primary endpoint in phase II randomized clinical trials (RCTs). However, time-to-event endpoints require care at the design stage because the accrual and follow-up periods must be defined, and a censoring pattern must be assumed. Furthermore, because the power to detect a reduction in the hazard rate depends on the number of events observed, recruiting enough patients with a rare cancer in order to observe a sufficient number of events for a between-arm comparison with a time-to-event endpoint can be difficult. Bayesian designs that leverage historical control data to augment the control arm are sought in rare cancers because there is a practical and ethical willingness to trade modest increases in bias and Type 1 error, particularly under prior-new data conflict, for meaningful reductions in variance and improved feasibility. The literature has presented such Bayesian design applications and discussed these issues at length, albeit largely for binary or continuous endpoints [[1], [2], [3]]; applications to time-to-event endpoints are less common [4]. Accordingly, our focus is a time-to-event hybrid-controlled design that uses robust borrowing and explicitly evaluates operating characteristics (including Type 1 error) across degrees of prior-new data conflict. Because we have the tools to recreate patient-level time-to-event data from published RCTs and because we want to encourage innovative designs that augment the control arm in rare cancers, we provide herein a real-world, from start to finish application to design a phase II RCT using a time-to-event endpoint in ocular melanoma that metastasized to the liver. Primary ocular melanoma is a rare form of melanoma with six cases per million annually [5]. Although diagnosis typically occurs at an early localized stage, 50% of cases will experience metastatic recurrence; the liver is the most common metastatic site, involved in more than 90% of cases [6]. While an overall survival (OS) benefit from liver-directed therapy has not been demonstrated, it can improve objective response rates (ORRs) and PFS. Across six RCTs of liver-directed therapy for metastatic ocular melanoma published within ≤10 years, the median PFS ranged from 4.5 to 10.4 months [[7], [8], [9], [10], [11], [12]]. A recently published RCT with systemic Tebentafusp showed an OS benefit; however, that trial was limited to HLA-A2-positive patients, who comprise less than half of patients with ocular melanoma, and the comparator arm did not reflect the liver-directed BAC regimens considered relevant for the proposed RCT [13]. To date, a proven standard treatment has not been established for liver-predominant metastasis from primary ocular melanoma that is not limited to HLA-A2-positive disease. Selection of liver-directed therapy is based on clinician and institutional expertise, and may include hepatic arterial infusion of chemotherapy, embolization procedures, isolated hepatic therapy, or ablative therapy. To test experimental therapies, RCTs continue to be designed with best alternative care (BAC) as the control until a therapy is approved for this indication. Given that BAC therapies are associated with stubbornly poor and highly variable outcomes and coupled with the rarity of the disease, conventional phase II trial designs that utilize historical control data to estimate effect sizes for sample size and power calculations are suboptimal. Furthermore, some criteria given by Pocock (1976) [14] for the effective use of historical data, which places restrictions on historical data usage to reduce the risk of bias in the results, may be impractical [15,16]. In this rare cancer setting, methods that trade some bias in favor of a substantial reduction in variability are needed if we are to infer more definitive conclusions about the preliminary efficacy of the experimental therapies being evaluated. Herein, our objective was to go beyond conventional trial design approaches and optimize the use of historical control BAC PFS data in both the design and analysis of a new RCT evaluating whether an experimental therapy will prolong PFS compared with BAC for liver-predominant metastasis from primary ocular melanoma. To make use of all relevant evidence from the BAC therapies, a hierarchical meta-analytic predictive approach was used to judiciously combine the data from the new RCT with the historical control data, which is a Bayesian approach in the hybrid trial setting that is more promising compared with other borrowing methods, particularly when aggregated historical data from multiple data sources are available [1,17]. By way of example, we illustrate for a time-to-event endpoint how this Bayesian approach can improve clinical trial efficiency compared with conventional frequentist designs, thereby enabling earlier availability of effective therapies to patients with this devastating rare cancer. Methods Meta-analysis. Relevant historical control BAC PFS data from J=6 RCTs conducted within ≤10 years with a total of 309 patients were identified (Table 1) [[7], [8], [9], [10], [11], [12]]. Historical control data from a 7th RCT (IMCgp100-202) [13] in patients who received systemic chemotherapy with Tebentafusp were initially considered but later excluded (Supplemental Table 1). The reasons for exclusion were (1) Tebentafusp is limited to HLA-A2-positive patients; and (2) the comparator arm in the RCT did not use the liver-directed BAC regimens considered relevant for the new RCT proposed herein. Table 1Published randomized clinical trials (RCTs) of liver-directed therapies within ≤10 years.Table 1RCT (Phase)YearRelevant Historical Control BAC TherapiesbPFS (In Months)DescriptionnjdjEj=∑tiλˆjKMaMedian1SCANDIUM (Ph3)2022Isolated hepatic perfusion of melphalan4337388.660.09527.42FOCUS (Ph3)2022Percutaneous hepatic perfusion of melphalan9165855.250.07609.13SirTac (Ph2)2022Transarterial radioembolization2020137.830.14515.74Hughes (Ph3)2016Percutaneous hepatic perfusion of melphalan4432229.920.13926.25Valsecchi (Ph2)2015Immunoembolization2525314.400.079510.46EORTC 18021(Ph3)2014Hepatic arterial infusion of fotemustine8683624.040.13304.5Year = Year published. BAC = Best alternative care; KM = Kaplan-Meier; PFS = Progression-free survival.There are nj patients at risk of whichdj experienced a PFS event.Ej=∑ti (i=1,…,nj) represents the total observed survival time (complete + censored values).The term λˆj=dj/Ej is the estimated constant hazard rate.Assuming the dropout rate followed an exponential distribution, the term (nj−dj)/Ej was the constant dropout rate; the simple average of the dropout rates across the J=6 BAC therapies was 0.02.aReconstructed from the individual patient-level data from the published KM PFS curves.bBest alternative care therapies for the controls on Arm B in the new RCT. Published randomized clinical trials (RCTs) of liver-directed therapies within ≤10 years. Year = Year published. BAC = Best alternative care; KM = Kaplan-Meier; PFS = Progression-free survival. There are nj patients at risk of whichdj experienced a PFS event. Ej=∑ti (i=1,…,nj) represents the total observed survival time (complete + censored values). The term λˆj=dj/Ej is the estimated constant hazard rate. Assuming the dropout rate followed an exponential distribution, the term (nj−dj)/Ej was the constant dropout rate; the simple average of the dropout rates across the J=6 BAC therapies was 0.02. Reconstructed from the individual patient-level data from the published KM PFS curves. Best alternative care therapies for the controls on Arm B in the new RCT. To obtain the needed granular, patient-level information, the IPDfromKM Shiny app was used to reconstruct the individual patient-level data from the published Kaplan-Meier (KM) PFS curves for each of the J=6 RCTs [18]. Based on the graphical assessment of the negative log-transformed estimated survival probabilities plotted against time, the PFS times for each of the j control therapies closely followed an exponential distribution with a constant hazard rate λj (Supplemental Fig. 1A; Supplemental Fig. 1B). Table 1 presents the nj patients at risk for the jth historical control BAC therapy with the corresponding total exposure time Ej, of which dj experienced the PFS event. Here, Ej=∑ti (i=1,…,nj) represents the total observed survival time (complete + censored values). The estimated constant hazard rate, λˆj=dj/Ej, is also shown in Table 1 along with the median PFS obtained from the reconstructed KM curve. Let λc|λJ represent the control hazard rate given the J=6 historical control hazard rates λ1,…,λ6. To obtain the meta-analytic predictive (MAP) distribution, pMAP(λc|λJ), we applied a hierarchical model with data model dj|λj∼Poisson(λjEj) and parameter model log⁡(λj)=μ+ϵj, ϵj∼N(0,τ2). The term, τ2, is the between-trial heterogeneity parameter. A Normal(0,102) prior distribution was placed on μ, while for the parameter, τ, a half-normal prior with a standard deviation of σ/5=1/5=0.20 to allow for considerable heterogeneity across the J=6 historical control therapies [19]. Justification for the prior distribution placed on τ is provided in Supplemental Table 2, which provides pMAP(λc|λJ) results and posterior results for τ, for half-normal priors assuming moderate-to-considerable-to-large heterogeneity for the log-hazard as a function of the unit-information standard deviation σ=1 for a single exponentially distributed event [20]. Markov chain Monte Carlo (MCMC) sampling was used to estimate the MAP distribution based on three chains. For each chain, the first 750,000 of the 1M samples were discarded, such that each chain contributed 250,000 samples to the posterior distribution. The rstan package (v2.32.6, corresponding to Stan version 2.32.2) was used to fit the hierarchical model [21]. Using a two-component mixture of lognormal (LN) distributions, a parametric representation of pMAP(λc|λJ) was obtained. The R package mixR [22] was used to obtain the analytic expression for pMAP(λc|λJ), which was then used as the informative prior component for the hazard rate, λc, on the control arm in the study design of the new RCT. Study Design. The phase II RCT was designed to evaluate whether the experimental therapy (Arm A) would prolong PFS compared with BAC (Arm B) in patients with ocular melanoma with metastatic disease limited to the liver. We assumed a 2:1 randomization ratio favoring the experimental therapy, which is ethically and practically justifiable to increase trial participation in a rare cancer [3], a uniform accrual time of 24 months, and a follow-up time of 18 months. The total study time, therefore, is 42 months. The primary endpoint is PFS, defined as the time from randomization until disease progression or death. Disease evaluations are expected to take place every 2 months. Patients who do not experience disease progression or death are administratively censored at 42 months of total study time or at the time of the last known disease evaluation in the event of dropout. To determine the sample size, PFS times were assumed to follow an exponential distribution. Additionally, we assumed that patient dropouts would follow an exponential distribution with a common dropout rate of 0.02, which represented the simple average of the exponential dropout rates calculated from the reconstructed patient-level data from each of the J=6 control therapies (Table 1). Sample Size Determination. The simulation-based approach for Bayesian sample size determination was used to determine study size and was calibrated to possess good frequentist Type 1 and Type 2 error properties [23]. Let the null sampling prior for λc and for the hazard rate of the experimental therapy, λt, be denoted by πS0(λc) and πS0(λt), respectively, such that they are specified by a Dirac measure with the point-mass at the boundary value of the null parameter space, that is, πS0(λc)=πS0(λt)=δ(λ0), where λ0 is selected to be consistent with the mean and median of the MAP distribution. The null value, λ0, corresponds to a log hazards ratio of zero or Δ=log(λt/λc)=0. The alternative sampling value was chosen to reflect a 50% reduction in the hazard rate corresponding to Δ=log(λt/λc)=−0.6931. To compensate for the lack of a strong and consistent correlation between PFS and OS in metastatic ocular melanoma, a very large reduction in the hazard rate was targeted in this early phase II RCT [24]. We are interested in testing the one-sided hypothesis H0:Δ≥0 vs Ha:Δ<0. The experimental therapy will be worthy of further study in a confirmatory trial if there is a considerable chance of a positive treatment effect such that P(Δ<0|data)>0.90. To make use of all relevant evidence from the J=6 control therapies currently used in practice, a hierarchical MAP approach will be used to judiciously combine the new RCT data with the historical control data. A cautious incorporation of the historical control data will be made by using a robust mixture prior (RMP) that adds a weakly informative (robust) component, pR(λc), to the MAP prior via a pre-specified mixing weight (wR) to form a mixture distribution [25]. The analysis of the new RCT data will determine the appropriate weight to assign to the historical control data, based on the consistency between the historical control data and the new RCT control data on Arm B. The RMP can be expressed as: An LN (μ=ln[0.0900], σ=1.75) distribution with mean largely consistent with the mean of the pMAP(λc|λJ), albeit slightly lower to control the maximum one-sided Type 1 error inflation at 20%, and a much larger standard deviation serves as the robust component. A LN (ln[0.0900], 1.75) has 95% of the λc values in the range 0.0029 to 2.779, corresponding to a median PFS in the range 0.25 to 239 months, which includes all the plausible values for λc seen in pMAP(λc|λJ). From a philosophical standpoint, we are only interested in considering wR≥0.5. The mixing weight will be chosen to ensure that the maximum power is ≥ 80% and that the one-sided Type I error inflation in the presence of substantial prior-new data conflict in λc does not exceed 20%, a threshold consistent with the more liberal one-sided Type 1 error rates sometimes used in randomized phase II oncology screening designs [26], while recognizing that stricter control would be required for confirmatory trials. Here, the RMP serves as the fitting prior, πf(λc), for the hazard rate, λc. In the absence of relevant external information for the experimental therapy hazard rate, λt, and to allow the new RCT data to dominate the posterior, a sufficiently vague LN (0, 3) distribution serves as the fitting prior, πf(λt), for the hazard rate, λt. A LN (μ=ln[1], σ=3) has a very large standard deviation with 95% of the λt values in the range 0.0028 to 357.8, corresponding to a median PFS in the range 0.002 to 248 months. We started with a control-arm (Arm B) sample size of nc=12 and simulated 90,000 datasets given the aforementioned design assumptions. For each dataset generated, MCMC sampling based on three chains was used to estimate the posterior distribution of the log hazards ratio. For each chain, the first 75,000 of the 100,000 samples were discarded, such that each chain contributed 25,000 samples to the posterior distribution of the log hazards ratio or 75,000 samples total. Datasets generated under the null and alternative hypotheses were used to calculate the Type 1 error and power, respectively. Additionally, we calculated the Type 1 error and power for increasing degrees of prior-new data conflict in λc and for wR∈(0.5,0.6,0.7,0.8,0.9). Then, the control-arm sample size of nc was increased in increments of one until the study design achieved desirable frequentist properties. We report on the operating characteristics of the chosen Bayesian design and, for comparison, those of a conventional frequentist design based on the approximate permutation log-rank test given the small, unequally sized randomization arms [27,28]. Convergence diagnostics of the Markov chain to the posterior distribution for λc, λt, and log(λt/λc) were checked and are shown for a single dataset in the Supplemental Fig. 2. The simulations were conducted on a high-performance computing cluster. Results Fig. 1A shows the meta-analytic trial-specific estimates and 90% credible intervals (CIs) plus the posterior mean and prediction estimate for λc. The mean (SD) and median prediction estimate for λc was 0.1121 (0.0323) and 0.1078, respectively. The MAP distribution, pMAP(λc|λJ), is shown as a histogram in Fig. 1B. Using a two-component mixture of LN distributions, a parametric representation of pMAP(λc|λJ) was obtained and overlaid as a density curve. The parametric representation of pMAP(λc|λJ) is expressed analytically as: which provided a close approximation to pMAP(λc|λJ). Forest plot based on a meta-analysis allowing for considerable between-trial heterogeneity. Meta-analytic posterior predictive (MAP) distribution (histogram) and its parametric representation (density curve). Using a two-precisions approach [29], the effective sample size of the MAP prior pMAP(λc|λJ), expressed in terms of the number of patients, was 12 (Supplemental Table 3). The historical data will therefore contribute approximately 12 patients to the analysis of the new RCT. Based on the meta-analytic prediction of λc, we assume in the new RCT that the control hazard rate on Arm B will be λc=0.1100, corresponding to a median PFS of 6.3 months. We hypothesize an Arm A hazard rate of λt=0.0550 or a median PFS of 12.6 months, corresponding to a log hazards ratio, Δ=log(λt/λc)=−0.6931, or equivalently, a hazard ratio of 0.5. With fitting priors, and Table 2 displays the frequentist Type 1 error and power for nc∈(12,13,14) for the Bayesian design when there are no conflicts between the control data and historical sources, and the corresponding frequentist design. For example, a total of 27 events in 39 randomized patients (13 Arm B; 26 Arm A) led to 80.7% power to detect an improvement in median PFS from 6.3 months to 12.6 months; the estimated one-sided Type 1 error for the Bayesian design was 10.5%. The corresponding frequentist design yielded a power of 65.0%. Table 2Bayesian and frequentist design – Simulated operating characteristics based on 90,000 datasets when there are no conflicts between the control data and historical sources.Table 2Sample SizeaFitting PriorsUnder the Null HypothesisType 1 ErrorUnder the Alternative HypothesisPowerNo. of EventsBayesian DesignFrequentistDesignNo. of EventsBayesian DesignFrequentistDesignncntN122436πf(λc)=0.5×[0.635×LN(−2.2298,0.19002)+0.365×LN(−2.2184,0.37242)]+0.5×LN(ln⁡[0.09],1.75)πf(λt)=LN(ln⁡[1],3)300.10310.1004250.79230.6280132639320.10510.1017270.80680.6497142842340.10370.0991290.82720.6758aThe sample size accounts for patient dropout (exponential dropout rate of 0.02); therefore, there is no need to inflate N further. Bayesian and frequentist design – Simulated operating characteristics based on 90,000 datasets when there are no conflicts between the control data and historical sources. The sample size accounts for patient dropout (exponential dropout rate of 0.02); therefore, there is no need to inflate N further. For a single generated data set, it's instructive to visualize the estimated posterior distribution of the log hazards ratio when we reject the null hypothesis and when we fail to reject it, given that the data were generated under the alternative hypothesis. Consider the case in Table 2 where nc=13. Fig. 2A shows the posterior distribution for the log hazards ratio, log(λt/λc), for a single generated data set for which the null was rejected – the 90% quantile is to the left of the null value of zero. Fig. 2B shows the posterior distribution for the log hazards ratio, log(λt/λc), for a single generated data set for which we failed to reject the null – the 90% quantile is to the right of the null value of zero. Across the 90,000 data sets, the percentage of times we rejected the null was 80.7% compared with 65.0% power with the frequentist design (Table 2). Under the alternative hypothesis, we also recorded the mode of the posterior distribution for the log hazards ratio – the mean (SEM) over the 90,000 data sets was −0.6370(0.0015), or a hazards ratio of e−0.6370=0.5289, which represents a modest degree of bias relative to the targeted log hazards ratio, log(λt/λc)=−0.6931, or equivalently a hazards ratio of 0.5000. Posterior distribution for the log hazards ratio for a single generated data set under the alternative (nc=13) that led to a rejection of the null hypothesis. Posterior distribution for the log hazards ratio for a single generated data set under the alternative (nc=13) that led to a failure to reject the null hypothesis. Fig. 3A and B shows the operating characteristics (3A: Type 1 error; 3B: power) for increasing degrees of prior data conflict in λc and for wR∈(0.0,0.5,0.9) given a 2:1 randomization favoring the experimental therapy (26 Arm A; 13 Arm B). The chosen mixing weight of wR=0.5 resulted in a maximum one-sided Type 1 error inflation of 20% across increasing degrees of prior data conflict in the hazard rate for the control arm, λc; the maximum occurs when the observed control hazard rate in the new RCT is λc=0.0700. An observed hazard rate of 0.0700 on the control arm of the new trial corresponds to a median PFS of 9.9 months, which would be indicative of substantial prior-new trial data conflict. Given our confidence in the historical prior information, which was based on J=6 RCTs conducted within ≤10 years, the inflation of the one-sided Type 1 error in the presence of substantial, albeit unlikely prior-new trial data conflict is acceptable. Within the projected range of λc, a wR=0.5 also resulted in considerable gains in power relative to a frequentist design (Fig. 3B). The design operating characteristics based on the (A) Type 1 error and (B) power, are shown for increasing degrees of prior data conflict in the hazard rate for the control arm, λc, and for the mixing weight, wR∈(0.0,0.5,0.9), given a 2:1 randomization favoring the experimental therapy (26 Arm A; 13 Arm B). A mixing weight of wR=0.0 represents the historical prior only. A mixing weight of wR=0.5 applied to the robust component of the robust mixture prior (RMP) resulted in a maximum one-sided Type 1 error of 20% across increasing degrees of prior data conflict in the hazard rate for the control arm, λc, and considerable gains in power relative to the frequentist design. Sensitivity Analysis. The Supplemental Fig. 3A–C shows side-by-side the one-sided Type 1 error inflation for the selected prior used for the robust component, LN (μ=ln[0.0900], σ=1.75); a LN (μ=ln[0.1100], σ=1.75), which has a mean consistent with the mean of the MAP prior, pMAP(λc|λJ); and a LN (μ=ln[0.0900], σ=3), which has a substantially increased standard deviation. The latter two priors assumed for the robust component led to a maximum one-sided Type 1 error that modestly exceeded 20%. Apart from that, the one-sided Type 1 error inflation was similar across the three priors in the presence of increasing prior-new trial data conflict. Assuming the same design assumptions, we also conducted a sensitivity analysis to target a smaller effect, namely, a hazards ratio of 0.60, or a reduction in the hazard rate of 40%. Given a 2:1 randomization favoring the experimental therapy, 72 randomized patients (48 Arm A; 24 Arm B) are needed to achieve similar operating characteristics (Supplemental Fig. 4). Selected Study Design. If we randomize 39 patients (26 patients to receive the experimental therapy [Arm A] and 13 patients to receive BAC [Arm B]), and the RCT BAC data are consistent with the historical BAC data, a total of 27 events will provide the new RCT with 80.7% power to detect an improvement in median PFS from 6.3 months to 12.6 months. The estimated one-sided Type 1 error for this design is 10.5%. A traditional frequentist design with 1:1 randomization would require 54 patients (27 per arm) and 39 events to detect a hazard ratio of 0.5, with approximately the same power (Table 3), assuming a one-sided, approximate permutation log-rank test at an α = 0.10 level of significance and an exponential common dropout rate of 0.02. Put another way, the Bayesian design with 2:1 randomization favoring the experimental therapy (26 in Arm A; 13 in Arm B) reduced the number of trial participants by 28%. Table 3Simulated operating characteristics of the primary design, when there are no conflicts between the control data and historical sources, relative to alternative frequentist designs.Table 3DesignsSample SizebSimulated Operating CharacteristicsncntNUnder AlternativeUnder NullNo. of EventsPowerNo. of EventsOne-sidedαBayesian Design – 2:1 RandomizationConsiderable Heterogeneity in MAP Priora1326392780.7%320.1051Frequentist Design1:1 Randomization2727543980.0%440.10071:1 Randomization2626523878.7%430.10032:1 Randomization1326392765.1%320.1017aPrimary Bayesian design for the new RCT that assumed considerable between-trial heterogeneity in the meta-analytic predictive (MAP) distribution. The effective sample size of the MAP prior was 12 patients.bThe sample size accounts for patient dropout (exponential dropout rate of 0.02); therefore, there is no need to inflate N further. Simulated operating characteristics of the primary design, when there are no conflicts between the control data and historical sources, relative to alternative frequentist designs. Primary Bayesian design for the new RCT that assumed considerable between-trial heterogeneity in the meta-analytic predictive (MAP) distribution. The effective sample size of the MAP prior was 12 patients. The sample size accounts for patient dropout (exponential dropout rate of 0.02); therefore, there is no need to inflate N further. Discussion At the design stage, accrual and follow-up periods must be defined, and a censoring pattern must be assumed when using a time-to-event endpoint such as PFS as the primary endpoint in a phase II RCT. In an NCI Network Group clinical trial – the target setting for the example RCT – it is reasonable to accrue 39 patients with metastatic ocular melanoma limited to the liver over two years. Because it was straightforward to recreate the patient-level data from all six published RCTs, we had direct evidence to justify our assumption of an exponential dropout rate of 0.02. These assumptions are important to get right as they form the backbone of the design of any RCT with a time-to-event endpoint. Nonetheless, given the increase in design parameters for a time-to-event endpoint relative to binary and continuous endpoints, substantially more upfront time and computational resources are required to calibrate the Bayesian design to achieve good frequentist operating characteristics; notably, access to a high-performance computing cluster is crucial. There are additional points to consider regarding the acceptability of the trial design. First, for rare diseases, designs that assure participants that they will receive the experimental treatment favor patient participation and retention and reduce the risk of dropout [30,31]. Crossover to the experimental arm should be allowed, which will not affect the integrity of the PFS endpoint provided that clear documentation of disease progression is available beforehand. Second, for patients randomized to the BAC arm, the treating physician at each study center will need to decide the treatment, together with the patient. All available treatments, including surgery and other experimental treatments, should generally be tolerated. Selection of liver-directed therapy is based on clinician and institutional expertise, and may include hepatic arterial infusion of chemotherapy, embolization procedures, isolated hepatic therapy, or ablative therapy. Lastly, randomization should be stratified according to centrally assessed lactate dehydrogenase (LDH) status (normal; abnormal), which is included in the TNM staging system of melanoma and can have a powerful impact on PFS [32]. Given the small sample size, few patients enrolled on the BAC arm, and an accrual period of 24 months – coupled with research interests in correlative science endpoints in this early-phase setting – it is reasonable not to incorporate an interim analysis to stop the trial early for efficacy. If such an interim analysis is desired, however, the posterior probability approach used at the final analysis could be applied at the interim analysis within the context of a Bayesian group-sequential design, similar to a Z-score in a frequentist group-sequential design [23]. However, interim analysis for futility should be considered; for example, to evaluate the BAC hazard rate, λc, to see if the estimate falls below 0.0783, corresponding to the 10th percentile of the historical MAP distribution. If the BAC hazard rate is below the 10th percentile, this would provide evidence that the field has changed so drastically that we may need to retest everything. The study team should work with the Data and Safety Monitoring Board (DSMB) to determine next steps. If a redesign is desired – for example, based on monitoring events on the BAC arm – an independent statistician should be identified to collaborate with the DSMB to re-estimate the sample size. Lastly, a nonbinding futility analysis should also be considered at the design stage, say, if the HR > 1 [33]. Other Bayesian hybrid-controlled designs, which draw strength from historical data, could be considered as alternatives to the MAP prior method, including the use of a power prior, a commensurate prior, or an elastic prior [[34], [35], [36]]. In contrast to these other methods, the rationale for selecting the MAP prior approach was fourfold. First, the hierarchical model formulation accounts for between-trial heterogeneity across multiple historical data sources. Second, the approach offers flexibility to inform the posterior based on the relevance of the historical data to the new RCT data via a robust MAP prior; other than prespecifying the weight, w, we do not have to define upfront how different the historical data and new RCT data would need to be to be considered similar enough [37]. Additionally, the meta-analytic method can be extended straightforwardly to account for other sources of differential heterogeneity [38] and to account for trial-level covariates [17]. Lastly, the MAP approach has been recently shown to be more promising than the other currently available Bayesian hybrid-controlled designs [1,17]. As precedent, the design of an ongoing pediatric trial in multiple sclerosis (NCT04926818) with a binary response endpoint was accepted under the FDA's Complex Innovative Designs pilot program [39] and utilizes the Bayesian MAP approach [40]. Regardless of the chosen Bayesian hybrid-controlled design, regulatory and funding agencies recommend that sponsors provide the frequentist operating characteristics of the Bayesian analysis strategy, particularly focusing on the frequentist Type I error and power for all realistic scenarios [41,42]; thus, we carefully calibrated our design to achieve reasonable frequentist operating characteristics. There are limitations with the MAP prior method. Principally, the approach does not address systemic or hidden biases arising from shifts in patient-level baseline covariates or their outcomes between the historical data sources and the new RCT data. Methods have recently been developed to safeguard against these biases but were not considered in this example [43,44]. It is therefore important to thoughtfully select the historical data and justify its validity. The Bayesian approach used in the example is suitable for designs that emphasize the early identification of therapies showing clinical activity. That is not to say that the design approach cannot be considered in the late-stage setting; however, stricter control of the Type 1 error will be needed. Another limitation is that the design of the RCT assumed that the PFS times followed an exponential distribution. At the time of the final data analysis, if the constant hazard assumption is not satisfied within each randomized arm and/or the proportional hazards assumption is not satisfied, alternative data analysis approaches will be needed. While this may lead to reduced precision and an underpowered study, it ensures a suitable analysis method for the given data and that the analysis is based on the current, observed data. The observed results will then be used to inform sample size calculations for future studies. In such a setting, a Bayesian analysis of the data will need to adopt non-informative priors, and could include a restricted mean survival time analysis, which is a robust alternative when the time-to-event assumptions used in the design are violated [45]. When the assumptions are violated, the aim is to let the trial data determine the posterior distributions. The choice of alternative analysis strategies need to be prespecified in the statistical analysis plan. If an alternative primary analysis such as restricted mean survival time is used because the exponential or proportional-hazards assumptions are not tenable, the original posterior-probability decision threshold for the exponential hazards ratio model should not be carried over without recalibration; any alternative estimand, prior, and decision threshold would need to be prespecified and evaluated by simulation. Conclusion Using PFS as the primary endpoint, the Bayesian hybrid-controlled RCT resulted in a 28% reduction in the number of trial participants. Utilizing historical BAC PFS data, which comprised 309 patients with liver-predominant metastasis from primary ocular melanoma, can improve trial efficiency and potentially result in earlier availability of effective therapies to patients with this devastating rare disease. Funding support David Zahrieh is supported by the Innovation in Clinical Trial Design Award program through the Center for Clinical and Translational Science (CCaTS; UL1TR002377) at Mayo Clinic. Xiaofei Wang is partially supported by U01FD007934 from the FDA/HHS and R01AG06688 from the NIH/HHS. The contents of this article are those of the authors and do not necessarily represent the official views of, nor an endorsement by, the FDA/HHS, NIH, or the U.S. Government. CRediT authorship contribution statement David Zahrieh: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. Xiaofei Wang: Writing – review & editing. Svetomir N. Markovic: Writing – review & editing. Vera J. Suman: Supervision, Writing – review & editing. Declaration of competing interest The authors do not have any potential conflicts of interest. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.conctc.2026.101653. Supplementary data The following is the Supplementary data to this article: Multimedia component 1 Data availability No data are included in the research. Sample source code applied to synthetic data is available at https://github.com/dzahrieh/melanoma. References Associated Data Supplementary Materials Multimedia component 1 Data Availability Statement No data are included in the research. Sample source code applied to synthetic data is available at https://github.com/dzahrieh/melanoma.