 Research
 Open access
 Published:
Practical basket design for binary outcomes with control of familywise error rate
BMC Medical Research Methodology volumeÂ 23, ArticleÂ number:Â 52 (2023)
Abstract
Background
A basket trial is a type of clinical trial in which eligibility is based on the presence of specific molecular characteristics across subpopulations with different cancer types. The existing basket designs with Bayesian hierarchical models often improve the efficiency of evaluating therapeutic effects; however, these models calibrate the type I error rate based on the results of simulation studies under various selected scenarios. The theoretical control of familywise error rate (FWER) is important for decisionmaking regarding drug approval.
Methods
In this study, we propose a new Bayesian twostage design with one interim analysis for controlling FWER at the target level, along with the formulations of type I and II error rates. Since the difficulty lies in the complexity of the theoretical formulation of the type I error rate, we devised the simulationbased method to approximate the type I error rate.
Results
The proposed design enabled adjustment of the cutoff value to control the FWER at the target value in the final analysis. The simulation studies demonstrated that the proposed design can be used to control the wellapproximated FWER below the target value even in situations where the number of enrolled patients differed among subpopulations.
Conclusions
The accrual number of patients is sometimes unable to reach the predefined value; therefore, existing basket designs may not ensure defined operating characteristics before beginning the trial. The proposed design that enables adjustment of the cutoff value to control FWER at the target value based on the results in the final analysis would be a better alternative.
Background
Recent developments in molecular biology and genomics have enabled the classification of patients with common organspecific cancers into several subpopulations depending on their molecular profiles derived using nextgeneration genomic sequencing. Clinical development of molecular targeted therapies has recently become pivotal and has been accelerated by the emergence of master protocol trials that assess the combination of several molecular markers and their targeted therapies by employing multiple substudies for single or multiple tumor types [1,2,3]. A basket trial is a type of clinical trial in which eligibility is based on the presence of a specific molecular characteristic across subpopulations with different types of cancers. This feature assumes that a fairly accurate prediction can be made regarding the response of a subpopulation with molecular characteristics to a targeted treatment. Based on this hypothesis, traditional Bayesian hierarchical models (BHMs) based on information borrowing among subpopulations to improve the efficacy of therapeutic effect evaluation are used [4]. However, in some cases, the therapeutic effects among subpopulations with common molecular characteristics are heterogenous, indicating that exchangeability between these subpopulations cannot be assumed. Development of flexible Bayesian methods that account for exchangeable and nonexchangeable subpopulations by extending the BHM or Bayesian model averaging has been garnering interest in addressing the aforementioned issue [5,6,7,8,9,10,11,12,13]. However, the existing basket designs calibrate the type I and II error rates as well as the familywise error rate (FWER) based on the results of simulation studies under several selected scenarios because of the methodological nature of Bayesian methods.
Explicit control of the FWER at the target value is also important in basket trials, particularly in registration trials for drug approval [14,15,16]. In this study, we propose new formulations of subpopulationspecific type I and II error rates and FWER by extending the type I error rate in Bayesian sequential design for singlearm phase II trials with binary outcomes proposed by Shi and Yin (2019) [17]. Based on these formulations, we developed a twostage Bayesian design with a binary endpoint controlling the FWER. In the first stage, we performed futility analysis to exclude subpopulations without efficacy based on the Bayesian posterior probability. In the second stage, we determined whether targeted therapy was effective for each subpopulation. The proposed design enables adjustment of the cutoff value to control the FWER in the final analysis according to the number of accumulated patients. Using the proposed method, the exact type I and II error rates and FWER can be calculated; however, it involves timeconsuming numerical experiments to determine the sample size. We, therefore, devise a timesaving simulationbased method to approximate the type I error rate. Simulation studies were conducted under various scenarios and settings to evaluate the operating characteristics of the proposed method.
The proposed design is presented in the Methods section. The operating characteristics of the proposed design are described in the Results sections. Finally, the utility of the proposed design is summarized in the Discussion and conclusion section.
Methods
This section is divided into four parts: the trial framework, the estimation method for posterior probability, the definitions of type I/II error rates and FWER, and the numerical calculation method for type I/II error rates.
Trial framework
We propose a twostage design using binary endpoints with J (\(j = 1, \cdots , J\)) subpopulations. The notation of the trial framework is presented in TableÂ 1. For each subpopulation, the trial performs one interim analysis for futility stopping when the number of patients included in that subpopulation reaches a prespecified value. Consequently, J interim analyses were performed. Subpopulation j was excluded from the study when the futility stopping boundary was obtained. The final analysis was performed when all remaining subpopulations achieved a predefined number of subjects. Supposing that \(\{j\}\) refers to the index of the subpopulations excluding subpopulation j, \(\{j\}\) can be calculated as \(\{j\} \ (=\{1,\cdots ,j1,j+1,\cdots ,J\})\).
In the interim analysis for subpopulation j, we estimated the posterior probability \(\Pr (p_j >p_0 \mid \textbf{n}^{*},\textbf{r}^{*})\), where \(\textbf{n}^{*}\) and \(\textbf{r}^{*}\) represent the vectors of data with respect to the number of patients and responses in interim analysis for subpopulation j, respectively. Subpopulation j was excluded from the study when the posterior probability of \(\Pr (p_j>p_0 \mid \textbf{n}^{*},\textbf{r}^{*})\) was less than the prespecified threshold of \(c_{1,j}\).
In final analysis after completing patient enrollment, we estimated \(\Pr (p_j>p_0 \mid \textbf{n},\textbf{r})\) where \(\textbf{n}\) and \(\textbf{r}\) refer to the vectors of data with respect to the number of patients and responses in the final analysis, respectively. Notably, \(n_j=n_j^{*}\) at the final analysis when the subpopulation j satisfied the futility criterion during the interim analysis for subpopulation j. The investigational treatment was declared as effective in subpopulation j when the posterior probability of \(\Pr (p_j>p_0 \mid \textbf{n},\textbf{r})\) exceeded the prespecified threshold of \(c_{2,j}\).
The two thresholds, \(c_{1,j}\) and \(c_{2,j}\), were determined for each subpopulation to maintain FWER at the target level using the method described in the Numerical calculation section.
Posterior probability estimation
We estimate the posterior probabilities of \(\Pr (p_j >p_0 \mid \textbf{n}^{*},\textbf{r}^{*} )\) of Eq.Â (1) and \(\Pr (p_j>p_0 \mid \textbf{n},\textbf{r})\) of Eq.Â (2) based on the BHM [18]. \(\theta _j\) itself be a logodds \(\theta _j=\textrm{logit} (p_j)\); the prior probability of \(\theta _j\) sets the normal distribution with mean \(\mu\) and variance \(\tau ^{2}\) as follows:
The hyperprior parameter of \(\mu\) is assumed to follow a normal distribution with a mean \(\tilde{\mu }\) and variance \(\tilde{\sigma }^2\),
The hyperprior parameter of \(\tau\) is assumed to follow the gamma distribution with mean \(\nu /\xi\) and variance \(\nu /\xi ^2\).
The posterior joint distribution for the parameters for subpopulation j in the final analysis (replace \(\textbf{n}\) with \(\textbf{n}_j^{*}\) for interim analysis) is as follows:
where \(L(\theta _j \mid n_j, r_j) = \left( {\begin{array}{c}n_j\\ r_j\end{array}}\right) p_j^{r_j}(1p_j)^{n_jr_j}\),and \(\varvec{\theta }=(\theta _1,\theta _2,\cdots , \theta _J)\) is the vector of the log odds for the true response rate for subpopulation j, and f is the probability density function for each parameter.
In the proposed method, \(\Pr (p_j>p_0 \mid \textbf{n},\textbf{r})\) of Eq.Â (2) is calculated considering the prior distribution of \(p_0\) based on Thall and Simon (1994) [18] as follows:
where \(p_0\) is assumed to follow the beta distribution with hyperparameters \(a_0\) and \(b_0\). The values of \(a_0\) and \(b_0\) are set as the mean of the beta distribution on \(p_0\) (i.e., \(p_0=a_0/(a_0+b_0)\)) based on historical information. \(F(p_j \mid \textbf{n},\textbf{r})\) are the cumulative distribution functions of \(p_j\) calculated from the posterior samples of the posterior joint distribution of Eq.Â (6) using the Hamilton Monte Carlo method.
Definitions
Subpopulationspecific type I error rate and FWER
The type I error rate in the Bayesian sequential design with the binary endpoint proposed by Shi and Yin (2019) [17] was defined as the sum of the probabilities for all possible cases in which the truly ineffective treatment was incorrectly declared as effective. As the response rates in the basket trials can be correlated among subpopulations, and the BHM borrows information among subpopulations, formulating the definition of subpopulationspecific type I error rate should account for the observed response rates of the remaining subpopulations. To this end, we first introduce the probability of declaring the therapeutic effect in the final analysis under a null response rate of \(p_0\) for the subpopulation j.
where I indicates the indicator function and Bin() is the probability density function of the binomial distribution. We extend \(\tilde{\alpha _j}(\textbf{n}^{*},\textbf{r}^{*},\textbf{n},\textbf{r})\) to the subpopulationspecific type I error rate for subpopulation j, which accounts for all possible combinations of \(n_{\{j\}}^{*}\),\(r_{\{j\}}^{*}\),\(n_{\{j\}}\), \(r_{\{j\}}\) as follows:
where \(w=\prod _{j^{\prime }\in \{j\}}Bin(r_{j^{\prime }}^{*};n_{j^{\prime }}^{*},p_{j^{\prime }})Bin(r_{j^{\prime }}r_{j^{\prime }}^{*};n_{j^{\prime }}n_{j^{\prime }}^{*},p_{j^{\prime }})\); i.e., \(\alpha _j\) is the weighted average of \(\tilde{\alpha _j}(\textbf{n}^{*},\textbf{r}^{*},\textbf{n},\textbf{r})\) for all possible values with respect to \(n_{\{j\}}^{*}\),\(r_{\{j\}}^{*}\),\(n_{\{j\}}\), and \(r_{\{j\}}\) for the subpopulations \(\{j\} \ (=\left\{ 1,\cdots ,j1,j+1,\cdots ,J \right\} )\), which are weighted by the corresponding probability of w. As possible values, we also consider the values of pattern with which the subpopulation j met the futility criterion at the interim analysis for subpopulation j, and then we set \(n_j=n_j^{*}\) with that pattern. The FWER is defined as \(\alpha _{FWER}=1\prod _{j=1}^J \left\{ 1\alpha _j \right\}\) under the assumption that the true response rates in all subpopulations represent null response rates, \(p_j=p_0\) (\(j=1,\cdots ,J\)).
Subpopulationspecific type II error rate
To define the subpopulationspecific type II error rate for subpopulation j, we also introduced the probability of declaring the therapeutic effect during the final analysis with the alternative response rate of \(p_1\) for subpopulation j as follows:
to obtain the subpopulationspecific type II error rate for subpopulation j.
where \(v=\prod _{j^{\prime }\in \{j\}}Bin(r_{j^{\prime }}^{*};n_{j^{\prime }}^{*},p_{j^{\prime }})Bin(r_{j^{\prime }}r_{j^{\prime }}^{*};n_{j^{\prime }}n_{j^{\prime }}^{*},p_{j^{\prime }})\). We define the subpopulationspecific power for the subpopulation j as \(1\beta _j\).
Numerical calculation
Extensive calculations are required to determine the values of \(\alpha _j\), \(\beta _j\), and \(\alpha _{FWER}\), even when the number of subpopulations is small. For example, when calculating \(\alpha _j\) with \(J=4\), \(n_j^{*}=10\) (\(j=1,\cdots ,J\)) and \(n_j=20\) (\(j=1,\cdots ,J\)), the calculations are repeated for 700 million combinations of the values of \(\textbf{r}^{*}\) and \(\textbf{r}\); therefore, the burdens and time required to perform calculations for obtaining the exact values of of \(\alpha _j\), \(\beta _j\), and \(\alpha _{FWER}\) are enormous and unrealistic. Instead of the exact values of \(\alpha _j\), \(\beta _j\), and \(\alpha _{FWER}\), we propose to use values of \(\alpha _j\), \(\beta _j\), and \(\alpha _{FWER}\) calculated using simulated data and call those the wellapproximated \(\alpha _j\), \(\beta _j\), and \(\alpha _{FWER}\) in this study. To this end, we devised a numerical approach for determining the thresholds that can approximately control the FWER at the target level. Specifically, given the values of \(n_j^{*}\), \(n_j\), \(p_0\), \(p_1\), \(a_0\), \(b_0\), \(c_{1,j}\) and \(c_{2,j}\) (\(j=1,\cdots ,J\)), we calculated the wellapproximated values of \(\alpha _j\) and \(\beta _j\) using simulated data in five steps as described in the following paragragh. Subsequently, wellapproximated FWER was calculated from \(\alpha _{FWER}=1\prod _{j=1}^J \left\{ 1\alpha _j \right\}\) using the wellapproximated values of \(\alpha _j\).

Step 1
We set the parameters of \(n_j^{*}\), \(n_j\), \(p_0\), \(p_1\), \(a_0\), \(b_0\) (\(j=1,\cdots ,J\)), under which the arbitrary ranges of \(c_{1,j}\) and \(c_{2,j}\) were specified, and all combinations of these values were enumerated.

Step 2
We randomly selected subpopulation j from a multinomial distribution with equal success probability among subpopulations and generated a binary response data for a patient from the Bernoulli distribution with the true response rate (i.e., \(p_0\) or \(p_1\)) in the subpopulation j. This step was repeated until the number of patients in any subpopulation reached \(n_j^{*}\) (\(j=1,\cdots ,J\)).

Step 3
After reaching \(n_j^{*}\) patient enrollment in subpopulation j, futility evaluation based on Eq.Â (1) is performed. If subpopulation j satisfied the futility criterion in the interim analysis, subpopulation j was excluded from the trial. Otherwise, patient enrollment was continued for subpopulation j.

Step 4
After completing patient enrollment for all subpopulations except for the excluded subpopulations in the previous step, the posterior probability for subpopulation j was evaluated based on Eq.Â (2).

Step 5
Steps 2â€“4 were repeated up to 10,000 times, and the empirical values of \(\alpha _j\) (i.e., proportion of declaration that the investigational treatment was effective under the assumption of \(p_0\)) and \(\beta _j\) (i.e., proportion of declaration that the investigational treatment was ineffective under the assumption of \(p_1\)) were calculated.
Throughout these steps, determination of the two thresholds, \(c_{1,j}\) and \(c_{2,j}\), was explored for each subpopulation to maintain a wellapproximated FWER of \(\alpha _{FWER}\) at the target level. If two and more combinations of \(c_{1,j}\) and \(c_{2,j}\) were obtained to control the wellapproximated FWER \(< 5\%\), then we selected the values of \(c_{1,j}\) and \(c_{2,j}\) with the lowest average of wellapproximated values of \(\beta _j\) among subpopulations with the true response rates in all subpopulations are under \(p_1\), i.e., \(p_j = p_1 (j=1,\cdots ,J)\).
Results
We evaluated the operating characteristics through simulation studies. This section is divided into four parts: simulation settings, investigation of the accuracy of the numerical approach, simulation results under various scenarios, and operating characteristics with different sample sizes among subpopulations.
Simulation settings
We evaluated the operating characteristics of the proposed method in a simulation study. The values of \(p_0\) and \(p_1\) were set to 0.05 and 0.25, respectively. The number of subpopulations J was set to four. The true response rates of each subpopulation assumed in each scenario with \(J=4\) are listed in TableÂ 2. The maximum number of accrual patients in each subpopulation was 20 (i.e., \(n_j=20\)). For each subpopulation, one interim analysis for futility was conducted when the number of accrued patients reached 10 (i.e., \(n_j^{*}=10\)). We assumed that \(p_0\) followed a beta distribution with parameters Beta(10,Â 190) and a mean of 0.05 based on the analysis reported by Hirakawa et al. (2018) [19].
In the proposed method, we used \(\tilde{\mu } = 1.734\) (i.e., \(p_j=(0.05+0.25)/2=0.15\)) and \(\tilde{\sigma }^2=10\) with a weak (or strong) degree of borrowing information, i.e., (\(\nu ,\xi\))=(2,Â 20) (or (2,Â 2)). We estimated the posterior distribution using the rstan package in R software.
In simulation studies, we compared the proposed method with an independent method. For the independent method, the posterior distributions of \(\Pr ( p_j > p_0 \mid \textbf{n}^{*},\textbf{r}^{*})\) and \(\Pr ( p_j > p_0 \mid \textbf{n},\textbf{r} )\) were calculated independently as follows:
where \(F_{Beta}\) refered the distribution function of the beta distribution, and \(a_{ind}\) and \(b_{ind}\) were set to 0.6 and 1.4, respectively, based on the analysis reported by Hirakawa et al. (2018) [19].
We conducted 10,000 simulations for each scenario. We have presented wellapproximated type I error rates, power (i.e., one minus type II error rate), and FWER here.
Accuracy of numerical approach for determining thresholds for declaration of the therapeutic effect
Before conducting the simulation studies, we examined the accuracy of the numerical values for \(c_1\) and \(c_2\) determined using the numerical approach described in the Numerical calculation section. We used common values for \(c_{1,j}\) and \(c_{2,j}\) between subpopulations (i.e., \(c_1 = c_{1,1} = \cdots = c_{1,J}\) and \(c_2 = c_{2,1}=\cdots =c_{2,J}\)) because the true null and alternative response rates were the same among subpopulations. To reduce the computational burden of the calculations, we compared the exact FWER with a wellapproximated value using the proposed numerical approach in the independent method. We set the ranges for \(c_{1}\) and \(c_{2}\) from 0.2 to 0.4 in increments of 0.1 and from 0.95 to 0.99 in increments of 0.01. Other simulation settings were the same as those described in the Simulation settings section. The exact FWER was calculated from \(1\prod _{j=1}^{J} \left\{ 1\tilde{\alpha _j} (n_j^{*},r_j^{*},n_j,r_j) \right\}\) because the values of the proposed numerical method in the independent method were not affected by the results of other subpopulations. TableÂ 3 shows the exact and wellapproximated FWER under Scenario 1. We found that the observed differences between the exact and wellapproximated values of FWER were negligible.
Simulation results
For the hyperprior parameter of \(\tau\), which is assumed to be followed by the gamma distribution with mean \(\nu /\xi\) and variance \(\nu /\xi ^2\), \((\nu ,\xi )=(2,20)\) and (2,Â 2) were used in the proposed method with weak and strong information borrowing, respectively. In the following simulation studies, the values of \(c_1\) and \(c_2\) were selected to achieve a wellapproximated FWER of lower than and nearest to 5% using the proposed numerical approach. The ranges for \(c_{1}\) and \(c_{2}\) were 0.2â€“0.4 in increments of 0.1 and of 0.95â€“0.99 in increments of 0.01. We used \(c_1=0.2\) and \(c_2=0.98\) for the proposed method with weak information borrowing, \(c_1=0.2\) and \(c_2=0.95\) for the proposed method with strong information borrowing, and \(c_1=0.2\) and \(c_2=0.99\) for the independent method. Using these values for \(c_1\) and \(c_2\), the wellapproximated FWER for independent, BHMweak, and BHMstrong methods were 1%, 4%, and 4%, respectively. Note that the empirical FWER (i.e., proportion of declaration that investigational treatment was effective in at least one subpopulation among all subpopulations under the assumption that true response rates in all subpopulations are \(p_0\) (i.e., scenario 1)) for independent, BHMweak, and BHMstrong methods were 1%, 4%, and 4%, respectively . We found that the observed differences between the wellapproximated and empirical values of FWER were negligible.
TableÂ 4 shows the wellapproximated type I error rates and power when using \(c_1\) and \(c_2\) which were selected to maintain a wellapproximated FWER \(<5\%\) in 10,000 simulations. In the four scenarios in which at least one subpopulation had a true response rate of 0.05 (in scenarios 1, 2, 3, and 4), the average wellapproximated type I error rates of independent, BHMweak, and BHMstrong methods across the four subpopulations were 0%, 1%, and 3%, respectively.
Under the four scenarios in which at least one subpopulation had a true response rate of 0.25 (i.e., scenarios 2, 3, 4, and 5), the average wellapproximated power of independent, BHMweak, and BHMstrong methods across subpopulations were 57%, 71%, and 85%, respectively. The wellapproximated power of the BHMstrong methods increased with the increase in the number of subpopulations, with a true response rate of 0.25. In the scenario with all subpopulations showing a true response rate of 0.25 (i.e., scenario 5), the averaged wellapproximated power of independent, BHMweak, and BHMstrong methods across subpopulations were 57%, 71%, and 91%, respectively.
Operating characteristics with different samples sizes among subpopulations
We often encounter a situation in which the number of enrolled patients differs from that determined at the planning stage of the trials. In this section, the operating characteristics of the proposed method were calculated using the sample sizes at the final analysis \(({\mathrm n}_1,{\mathrm n}_2,{\mathrm n}_3,{\mathrm n}_4)=(20,20,20,5),(20,20,20,1),(20,20,20,30),(5,5,5,40),\;\mathrm{and}\;(15,\;10,\;5,\;1)\)Â in scenario 1 with the values of \(c_1\) and \(c_2\). These values were selected to achieve a wellapproximated FWER of lower than and closest to 5% with the sample size of \((n_1, n_2, n_3, n_4) = (20, 20, 20, 20)\) at the planning stage of the trials. Other simulation settings were the same as those used in the previous section.
TableÂ 5 shows wellapproximated FWER under the scenario 1 using different sample sizes among subpopulations. According to TableÂ 5, the BHMstrong method controlled the wellapproximated FWER to below 5%. The wellapproximated FWER exceeded the target level 5% when using \((n_1, n_2, n_3, n_4) = (5, 5, 5, 40)\) for the independent method and \((n_1, n_2, n_3, n_4) = (20, 20, 20, 1)\) and (15,Â 10,Â 5,Â 1) for the BHMweak method. To suppress this inflation of FWER, the value of \(c_2\) was redetermined using the actual sample size at final analysis. For the independent method, \(c_2\) was redetermined as 0.995, and the wellapproximated FWER was \(0.4\%\) under \((n_1, n_2, n_3, n_4) = (5, 5, 5, 40)\). For the BHMweak method, \(c_2\) was redetermined as 0.99, and the wellapproximated FWER were 4.3% and 4.8% with \((n_1, n_2, n_3, n_4) = (20, 20, 20, 1)\) and \((n_1, n_2, n_3, n_4) = (15, 10, 5, 1)\), respectively.
Operating characteristics with weak informative prior for the variance \(\tilde{\sigma }^2\) of BHM
We investigated the operating characteristics of the proposed method with BHM under the assumption of weak informative prior for the variance \(\tilde{\sigma }^2\) of the hyperprior parameter of \(\mu\) (TableÂ 6). We set \(\tilde{\sigma }^2\) to 25 or 50; other simulation settings were the same as those described in the Simulation settings section. In the following simulation studies, the values of \(c_1\) and \(c_2\) were selected to achieve a wellapproximated FWER of lower than and nearest to 5% using the proposed numerical approach.
TableÂ 6 shows the wellapproximated type I error rates and power in scenarios 15. The values of wellapproximated type I error rates and power were very similar with assumptions of not only \(\tilde{\sigma }^2 = 25\) and 50, but also \(\tilde{\sigma }^2 = 10\) in TableÂ 4. Therefore, we found that the value of the variance \(\tilde{\sigma }^2\) of the hyperprior parameter of \(\mu\) did not affect the operating characteristics of the proposed method.
Operating characteristics with weak informative prior distribution for \(p_0\)
We investigated the operating characteristics of the proposed method under the assumption that the prior distribution of \(p_0\) was weak informative (TableÂ 7). We assumed that \(p_0\) followed a beta distribution with parameters Beta(0.05,Â 0.95) (i.e., the effective sample size (ESS) is one) and a mean of 0.05. Other simulation settings were the same as those described in the Simulation settings section. In the following simulation studies, the values of \(c_1\) and \(c_2\) were selected to achieve a wellapproximated FWER of lower than and nearest to 5% using the proposed numerical approach.
TableÂ 7 shows the wellapproximated type I error rates and power in scenarios 15. We found that the power observed with weak informative prior distribution for \(p_0\) was lower than that observed with strong informative prior distribution for \(p_0\) in TableÂ 4; however, the power observed with weak informative prior distribution for \(p_0\) was still enough when using the BHMstrong method.
Operating characteristics with missspecification of prior distribution for \(p_0\)
We investigated the operating characteristics of the proposed method under the assumption that the prior distribution for \(p_0\) was missspecified (TableÂ 8). We set \(p_0\) to fit a beta distribution pattern with the following parameters: (i) Beta(0.1,Â 0.9) with a mean of 0.1, or (ii) Beta(0.2,Â 0.8) with a mean of 0.2; however, the true value of \(p_0\) is 0.05. Other simulation settings were the same as those described in the Simulation settings section. In the following simulation studies, the values of \(c_1\) and \(c_2\) were selected to achieve a wellapproximated FWER of lower than and nearest to 5% using the proposed numerical approach.
TableÂ 8 shows the wellapproximated type I error rates and power in scenarios 15. The values of wellapproximated type I error rates and power were very similar with assumptions of Beta(0.1,Â 0.9) and Beta(0.2,Â 0.8). The wellapproximated type I error rates and power for the BHMweak and the independent methods were similar to those under the prior distribution for \(p_0\) are correctly specified in TableÂ 4. Otherwise, the wellapproximated type I error rates and power for the BHMstrong method were slightly higher than those observed when correctly specifying the prior distribution for \(p_0\) in TableÂ 4. However, since the wellapproximated FWER of the BHMstrong method was controlled to a value lower than 5%, the proposed method can be used even if the prior distribution for \(p_0\) is missspecified.
Discussion and conclusion
Explicit control of FWER provides more reliable evidence of efficacy in the final analysis, even in basket trials, and improves decisionmaking for regulatory drug approvals. Existing Bayesian methods often control the empirical FWER using simulationbased procedures, but the FWER is not controlled when some assumptions are violated. In this study, we provided closedform equations for the type I and II error rates and FWER in the context of the proposed Bayesian twostage design and developed a numerical approach for determining the thresholds for controlling the wellapproximated FWER to the exact value. To achieve more accurate control of the wellapproximated FWER, the proposed method can also be used to adjust the threshold of \(c_2\) based on the actual sample size during the final analysis. The simulation studies demonstrated that the proposed design can be used to control the wellapproximated FWER below the target value even in situations where the number of enrolled patients differs among subpopulations.
This study showed that the proposed method could weakly control the FWER. However, from a regulatory perspective, strong control of the FWER is necessary, especially in confirmatory basket trials. Therefore, methodological studies of theoretical control of FWER at a target level are required.
In the Bayesian context, controlling the exact FWER is less fundamental when evaluating the therapeutic effect but should be wellcalibrated at the planning stage of trials [7,8,9,10,11,12,13]. Before beginning the trials, intensive simulation studies are often required to investigate the empirical FWER under various scenarios with respect to the prior distribution, assumed therapeutic effect, sample size, and other factors affecting the therapeutic effect. In contrast to such methods, the proposed method can easily obtain a wellapproximated FWER using the proposed numerical approach. Particularly, it is typically necessary to repeatedly calibrate the design parameters in simulation studies to determine the thresholds in most existing basket designs, whereas this is not the case in the proposed method. Although the proposed method focused on controlling the wellapproximated FWER, the subpopulationspecific type I error rates and power cannot be controlled at the desired level under a limited sample size. To overcome this issue, it may be better to control the false discovery rate. The extension of the proposed method for controlling the false discovery rate should be further evaluated.
We estimated the posterior probability using the independent and BHM methods in this study; however, other Bayesian dynamic borrowing information models for estimating the posterior probability, such as the extended BHM, the exchangeabilitynon exchangeability (EXNEX) approach, or Bayesian model averaging can also be used [5,6,7,8,9,10,11,12,13]. The calculation time of 10,000 simulations for BHM was approximately 24 h with a standard computer (e.g., Intel Core i7 1.1 GHz 1.61 GHz, 16 GB RAM). Therefore, it may take days to finish the calculations for basket trials with a larger number of subpopulations (e.g., \(\ge\)10 subpopulations) or when using other more computationally complicated Bayesian dynamic borrowing information models. In practice, the accrual number of patients is sometimes unable to reach the predefined value; therefore, existing basket designs sometimes cannot ensure the operating characteristics defined before beginning the trial. However, the proposed design enabled adjustment of the cutoff value to control the FWER at the target value based on the results in the final analysis. Additionally, the formulation concept for the type I error rates and the numerical calculation method for controlling the FWER are generalizable. Although the singlearm basket design was the proposed design in this study, we can extend the proposed design to the other types of basket designs such as randomized basket designs.
Availability of data and materials
Not applicable.
Abbreviations
 BHM:

Bayesian hierarchical model
 FWER:

Family wise error rate
References
Renfro LA, Sargent DJ. Statistical controversies in clinical research: basket trials, umbrella trials, and other master protocols: a review and examples. Ann Oncol. 2017;28:34â€“43. https://doi.org/10.1093/annonc/mdw413.
Woodcock J, LaVange LM. Master protocols to study multiple therapies, multiple diseases, or both. N Engl J Med. 2017;377:62â€“70. https://doi.org/10.1056/NEJMra1510062.
Hirakawa A, Asano J, Sato H, Teramukai S. Master protocol trials in oncology: Review and new trial designs. Contemp Clin Trials Commun. 2018;12:1â€“8. https://doi.org/10.1016/j.conctc.2018.08.009.
Thall PF, Wathen JK, Bekele BN, Champlin RE, Baker LH, Benjamin RS. Hierarchical Bayesian approaches to phase II trials in diseases with multiple subtypes. Stat Med. 2003;22:763â€“80. https://doi.org/10.1002/sim.1399.
Neuenschwander B, Wandel S, Roychoudhury S, Bailey S. Robust exchangeability designs for early phase clinical trials with multiple strata. Pharm Stat. 2016;15:123â€“34. https://doi.org/10.1002/pst.1730.
Liu R, Liu Z, Ghadessi M, Vonk R. Increasing the efficiency of oncology basket trials using a Bayesian approach. Contemp Clin Trials. 2017;63:67â€“73. https://doi.org/10.1016/j.cct.2017.06.009.
Chu Y, Yuan Y. A Bayesian basket trial design using a calibrated Bayesian hierarchical model. Clin Trials. 2018;15:149â€“58. https://doi.org/10.1177/1740774518755122.
Cunanan KM, Iasonos A, Shen R, GÃ¶nen M. Variance prior specification for a basket trial design using Bayesian hierarchical modeling. Clin Trials. 2019;16:142â€“53. https://doi.org/10.1177/1740774518812779.
Chen N, Lee JJ. Bayesian cluster hierarchical model for subgroup borrowing in the design and analysis of basket trials with binary endpoints. Stat Methods Med Res. 2020;29:2717â€“32. https://doi.org/10.1177/0962280220910186.
Asano J, Hirakawa A. A Bayesian basket trial design accounting for uncertainties of homogeneity and heterogeneity of treatment effect among subpopulations. Pharm Stat. 2020;19:975â€“1000. https://doi.org/10.1002/pst.2049.
Psioda MA, Xu J, Jiang Q, Ke C, Yang Z, Ibrahim JG. Bayesian adaptive basket trial design using model averaging. Biostatistics. 2021;22:19â€“34. https://doi.org/10.1093/biostatistics/kxz014.
Jin J, Riviere MK, Luo X, Dong Y. Bayesian methods for the analysis of earlyphase oncology basket trials with information borrowing across cancer types. Stat Med. 2020;39:3459â€“75.
Zheng H, Wason JM. Borrowing of information across patient subgroups in a basket trial based on distributional discrepancy. Biostatistics. 2022;23:120â€“35.
Kaizer AM, Koopmeiners JA, Kane MJ, Roychoudhury S, Hong DS, Hobbs BP. Basket designs: Statistical considerations for oncology trials. JCO Precis Oncol. 2019;3:1â€“9. https://doi.org/10.1200/PO.19.00194.
Collignon O, Gartner C, Haidich AB, Hemmings RJ, Hofner B, PÃ©tavy F, Posch M, Rantell K, Roes K, Schiel A. Current statistical considerations and regulatory perspectives on the planning of confirmatory basket, umbrella, and platform trials. Clin Pharmacol Ther. 2020;107:1059â€“67. https://doi.org/10.1002/cpt.1804.
Hirakawa A, Sato H, Igeta M, Fujikawa K, Daimon T, Teramukai S. Regulatory issues and the potential use of Bayesian approaches for early drug approval systems in Japan. Pharm Stat. 2022;21:691â€“5. https://doi.org/10.1002/pst.2192.
Shi H, Yin G. Control of type I error rates in Bayesian sequential designs. Bayesian Anal. 2019;14:399â€“425. https://doi.org/10.1214/18BA1109.
Thall PF, Simon R. Practical Bayesian guidelines for phase IIB clinical trials. Biometrics. 1994;50:337â€“49. https://doi.org/10.2307/2533377.
Hirakawa A, Nishikawa T, Yonemori K, Shibata T, Nakamura K, Ando M, Ueda T, Ozaki T, Tamura K, Kawai A, Fujiwara Y. Utility of Bayesian singlesubpopulation design in new drug application for rare cancers in Japan: a case study of phase 2 trial for sarcoma. Ther Innov Regul Sci. 2018;52:334â€“8. https://doi.org/10.1177/2168479017728989.
Acknowledgements
Not applicable.
Funding
Â This work was partially supported by Japan Society for the Promotion of Science grants 20K11703 (â€˜Grantinaid for scientific research Câ€™).
Author information
Authors and Affiliations
Contributions
J.A. did the programming and analysis. All authors (J.A., H.S., and A.H.) wrote the manuscript. All authors (J.A., H.S., and A.H.) were included in the finalization of the manuscript. The authors read and approved the final manuscript.
Corresponding authors
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
Â The authors have no conflicts of interest to declare. The views expressed in this article are those of the authors andÂ do not necessarily reflect the official views of Pharmaceuticals and Medical Devices Agency.
Additional information
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Asano, J., Sato, H. & Hirakawa, A. Practical basket design for binary outcomes with control of familywise error rate. BMC Med Res Methodol 23, 52 (2023). https://doi.org/10.1186/s12874023018721
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12874023018721