Riko Kelter
Institute of Medical Statistics and Computational Biology
Faculty of Medicine, University of Cologne
Cologne, Germany
The bfbin2arm package implements Bayes-factor-based
power and sample size calculations for binomial endpoints, with a focus
on early-phase clinical trials, in particular, phase II trials. The
central idea is to replace Monte Carlo simulation by fast numerical
calculations of design operating characteristics, both in fixed-sample
and two-stage settings. The underlying statistical theory is developed
in (Kelter and Pawel 2025a), extended to
the single-arm two-stage optimal setting by (Kelter and Pawel 2025b), and further developed
to the two-arm single-stage setting by (Kelter
2026).
The package covers:
In a typical single-arm phase II proof-of-concept trial with a binary endpoint, we test the null hypothesis
\[ H_0 : p \le p_0 \]
against the alternative
\[ H_1 : p > p_0, \]
where \(p_0\) is the response probability of a standard therapy or historical control. Alternatively, a two-sided test of \(H_0:p=p_0\) versus \(H_1:p \neq p_0\) can be carried out.
The package uses the Bayes factor \(BF_{01}\) as the central measure of evidence, with the convention that small values indicate evidence against \(H_0\) and large values indicate evidence in favour of \(H_0\). A two-stage design introduces a single interim analysis at \(n_1\), where one can stop early for futility if the interim Bayes factor provides sufficiently strong evidence in favour of \(H_0\).
The underlying hypotheses in the two-arm setting are described in the associated vignette in detail. For brevity, we only detail some basics about the single-arm case in this overview.
Two types of priors play a key role:
The package allows separate specification of design priors under \(H_0\) and \(H_1\) via Beta distributions (truncated in the directional setting). This separation is important because the prior used to quantify evidence in the test (Bayes factor) need not coincide with the prior used for planning.
The package supports several calibration modes that determine which operating characteristics must satisfy user-specified targets:
These modes are available for single-arm two-stage designs via the
function design_singlearm_bf() and for underlying two-stage
calibration via optimal_twostage_singlearm_bf().
This vignette serves as an entry point and does not include code. The package currently implements single-arm and two-arm designs, where the former only assumes the presence of a treatment group, and the latter an additional control group. Also, for both single- and two-arm designs, there are fixed-sample or one-stage designs which do not allow to stop the trial early after an interim analysis, and two-stage designs. Two-stage designs allow to stop the trial early (for futility), when the data show sufficient evidence in favour of the null hypothesis of no effect.
The following vignettes provide detailed tutorials for single-arm designs with executable examples:
Calibration of Bayesian one-stage designs for single-arm phase II trials with binary endpoints This vignette is the starting point and serves as the simplest introduction detailing the power and sample size calculations for Bayes factors in one-stage (fixed-sample) single-arm phase II trials with binary endpoints. No interim analysis and no control group are assumed.
Optimal Bayesian calibration for single-arm two-stage
Bayes factor designs with binary endpoints
Explains how to construct optimal two-stage designs where type-I error
and power are calibrated in a purely Bayesian sense. No control group is
assumed but an interim analysis is introduced into the trial
design.
Optimal frequentist calibration for single-arm two-stage
Bayes factor designs with binary endpoints
Explains how to construct optimal two-stage designs where type-I error
and power are calibrated in a purely frequentist sense compared to the
optimal Bayesian calibration in point 2. No control group is assumed but
an interim analysis is introduced into the trial design.
Optimal hybrid calibration for single-arm two-stage Bayes
factor designs with binary endpoints
Explains how to combine a prior-predictive Bayesian notion of power with
a frequentist interpretation of type-I error, which matches regulatory
expectations for frequentist calibration while preserving Bayesian
planning. Again, no control group is assumed but an interim analysis is
introduced into the trial design.
Optimal full calibration for single-arm two-stage Bayes
factor designs with binary endpoints
Explains how to simultaneously enforce Bayesian and frequentist
constraints, resulting in designs that satisfy both perspectives at
once. This is the strongest form of calibrating a design. No control
group is assumed but an interim analysis is introduced into the trial
design.
Each of these vignettes assumes familiarity with the basic single-arm phase II setup and with the terminology introduced above. All of these vignettes treat the single-arm case, where only a treatment group but no control group is available.
Two-arm designs are also available in the package, and two vignettes detail the process of calculating an optimal design in this setting:
Bayesian calibration of two-arm one-stage Bayes factor designs with binary endpoints Explains how to calibrate a two-arm phase II design with binary endpoints, where no interim analysis is carried out. Thus, this equals a fixed-sample standard power calculation from a Bayesian point of view when both a treatment and control group are available.
Optimal Bayesian calibration of two-arm two-stage Bayes factor designs with binary endpoints Explains how to calibrate a two-arm phase II trial with binary endpoints, where now an interim analysis should be carried out which allows to stop the trial early for futility. Again, a treatment and control group are assumed.