[Feature request] Better support for exact binomial tests (one- and two-sample)

[wolbersm]

My understanding is that currently rpact only provides limited support for single-arm trials with binomial endpoints getSampleSizeRates(..., groups = 1, normalApproximation = F).

The request is to enhance support for single arm trials and extend support to two-arm trials.

Requests for enhancements/clarifications for single arm trials:

• Due to the discreteness of the binomial distribution, the targeted type I error cannot be exactly achieved and the actual type I error varies with sample size. As a consequence, power does not increase monotonically with sample size. To account for this, rpact appears to calculate sample size conservatively, i.e. it chooses the sample size n such that the targeted power is not only achieved for this sample size $n$ but also for all sample sizes $>n$. It would be great if a this was clarified in the help and perhaps a non-conservative calculation also supported (as in gsDesign::nBinomial1Sample which has an argument conservative).
• rpact reports critical values on the treatment effect scale based on a normal approximation. This is somewhat confusing here because it does not match the exact critical value based on the binomial distribution.
• For designs with interim analyses, rpact uses a multivariate normal approximation to the joint distribution of test statistics across interims. Again, this implies that the calculation of stopping probabilities does not match the exact binomial calculations.

Requests for two arm trials:

• It would be great to add some support for two-arm trials (especially for the case without interim analyses), e.g. along the lines of what's implemented in the Exact package.

Thanks a lot for considering!

[fpahlke]

rpact 4.2.0: Argument conservative added to getSampleSizeRates() function

[gwassmer]

Exact critical values for one-sample test and request
"It would be great to add some support for two-arm trials (especially for the case without interim analyses), e.g. along the lines of what's implemented in the Exact package."
will be reconsidered