Timetrend.fun.In previous version, the power function was only for disconjunctive power and FWER. In this version, the conjunctive power conjuncativepower_or_FWER and marginal power (per-hypothesis type I error) perHtypeIerror_marginalpowerfunc can also be calculated for the used of design evaluation and power optimisation in asymmetric boundary cutoff screening. The function to calculate disconjunctive power is renamed as disconjunctivepowerfunc.
In previous version, the hypothesis testing during the interim analysis is \(H_0\): \(\pi_k = \pi_0\); \(H_1\): \(\pi_k \neq \pi_0\). Now we add one side test to only conclude for superiority: \(H_0\): \(\pi_k \leq \pi_0\); \(H_1\): \(\pi_k > \pi_0\).
In previous version, the cutoff value can only be tuned for symmetric boundary. In this version, asymmetric boundary cutoff screening is added so that user can decide whether to use a more aggressive efficacy boundary with a more aggressive futility boundary (or a more conservative efficacy boundary with more conservative futility boundary). The asymmetric boundary screening process first select the type I error rate contour equal to the target, then pick a cutoff pair (eff, fut) that maximizes the power.
In previous version, the Trippa’s adaptive randomisation approach used fixed hyperparameters suggested in their paper. In this version, the hyperparameters a and b are tuned to optimise power when type I error (FWER) is controlled at target. This idea can be found on Trippa’s paper (2012). We used a gaussian process model to smooth the contour plot. The pair (a, b) with highest predicted power will be exploit for hyperparameter screening until there are enough point overlapped on the contour plot. In addition, the hyperparameter point (a, b) can be pick in a crude way to save computation time via setting the maximum number of points to be looked at.
This release fix one command in fix effect model analysis of main effect that is confusing.
resultrtostats. The calculation of “stats4” was the sum of main effect and interaction effect when using the model with interaction term (main * stage_continuous / main * stage_discrete). The sum was used to calculate the posterior probability of superior or inferior. However, “stats4” only represents the mean values of main effect in the final output matrix. The wrong value was given to “stats4”. This is the same for “stats5” which is the standard errors of the main effect in the modelThis release fix two bugs.
GP.optim where the formula of information weighed randomisation is wrong.demo_Cutoffscreening where the nextcutoff vector for sample may have only one element. This will lead to the error in function sample when you only want to sample one value greater than 1. The argument ‘ntrials’ in each example should be large (> 100) instead 2 to make the example more like an actual simulation example. I use ntrials = 2 in the example to speed up the check process.This release fix one bug reported by the cran team.
GP.optim where the nextcutoff vector for sample may have only one element. This will lead to the error in function sample when you only want to sample one value greater than 1.This release implements additional method of cutoff screening using Bayesian optimization.
Add a new Demo function called demo_Cutoffscreening_GP. The function indicates how to use Gaussian process-based Bayesian optimisation algorithm to calibrate the critical value for stopping boundary under the null scenario to control the Type I error rate or FWER before trial simulation.
Add a function GP_optim which returns a cutoff value for the next evaluation during the cutoff screening process. The function is used in demo demo_Cutoffscreening_GP.
I fixed a bug where the fix ratio allocation method code is not suitable for the multi-arm trial
New code: rpk[-1] = rep(1 / (armleft - 1 + Fixratiocontrol), armleft - 1)
#-max.ar check if (1 - max.ar > 1/K){ stop(“Error: The lower allocation ratio should be at least 1/K. Please check the number of arm at the beginning and the max.ar”) }
Another change is the randomisation ratio adjustment in the Simulation_AdaptiveRandomisationmethodRatioCalc.R.
I modified the randomisation ratio adjustment command for Thall’s approach.
randomprob[as.numeric(colnames(rpk))] = randomprob[as.numeric(colnames(rpk))] + rpk
randomprob[as.numeric(colnames(rpk))] = randomprob[as.numeric(colnames(rpk))]+rpk