Simulate one or many frequentist confidence clinical trials based on a specified set of parameters. From a two-arm, single-stage trial to a perpetually run Adaptive Platform Trial, this package offers vast flexibility to customize your trial and observe operational characterisitics over thousands of instances.
The analysis of trial results is carried out using the
confidenceCurves package (on CRAN and
GitHub)
and all details about frequentist confidence analysis may be found
there.
The complexity of the design of the trial being simulated varies from
a two-arm one-stage trial, to a Multi-Arm Multi-Stage
(MAMS) Trial through to an Adaptive Platform
Trial. The trial design features are specified by a parameter
list that is fed into the runSingleTrial function via the
inputs field. There is an example parameter list named
inputs which may be loaded by using:
data(inputs)Some examples of designs are as follows:
With confidenceSim you may design a randomised trial
comparing a single experimental treatment to control with a single
analysis point which is the primary analysis. Primary analysis of
treatment efficacy is based on confidence in treatment benefit. To
design a single-stage trial, the looks fields would be a
single number, respresenting the maximum sample size. The length of
looks represents the number of stages.
confidenceSim, decisions to stop early are made on the
basis of confidence analysis. To control for Type I error, thresholds
are derived with the getGSdesign function using an
alpha-spending function. However, there are numerous options for trial
monitoring. For example, by setting
multiarm.mode = 'MONITOR FUTILITY' the trial will only be
stopped early for futility, and a test for efficacy only undertaken in
the final, primary, analysis. In that case a fixed
alpha=0.05 may be set. Otherwise
multiarm.mode = 'CONFIDENCE-BASED' uses confidence
monitoring for futility, efficacy and inferiority (harm).
A multi-arm multi-stage trial involves multiple experimental
treatments compared to a single control, and multiple analysis points.
For example, consider a design evaluating efficacy of two experimental
treatments against a common control over two stages. At the end of the
first stage, only the best performing treatment will go on, with all
remaining patients assigned to either this arm or control. In this case,
all three response rates (control + 2 treatments) must be specified in
the resprate parameter in inputs, with control
as the first rate. Then, by setting
multiarm.mode='SELECT THE BEST', the interim analysis will
carry the best treatment forward to the final analysis. Even if the
second best arm exceeds the confidence thresholds, it will not carry
forward. In this R package, multiple arms are compared to control with a
pairwise analysis and we only control for
multiplicity across stages, not arms. The same confidence
threshold is applied to each pairwise analysis. In this case, each pair
is treated like a separate trial.
The perpetual setting for inputs designates
if you want to run the trial perpetually, like an adaptive platform
trial, adding new arms once old arms are dropped. If you want to run
perpetually, the simulations will take new arms from the
resprate field once another arm is dropped. Therefore,
include all available treatments in resprate and
specify how many will run at any given time with the
alloc.ratio field. For example, you may have six treatments
to use - in which case there should be seven different response rates in
resprate, but you only run three study arms - in which
case, alloc.ratio should be of length 3, like
alloc.ratio = c(1,1,1). As with MAMS, each arm goes through
its own trial, with comparison to concurrent controls
only. The runSingleTrial tracks the trial of each treatment
separately, knowing if an analysis point is an interim analysis or a
final analysis based on the number of interims and the number of
patients on the arm so far. The fields alloc.ratio and
looks determines the journey for each treatment, and the
fields resprate and ppm determine how long the
platform trial itself runs for, based on how many people are recruited
and how many treatments there are. When treatments run out, the
perpetual setting is turned off. When people run out, the
trial stops.
In runSingleTrial, the primary outcome may be binary,
ordinal or continuous.
If the primary outcome measure is binary, patient response is 0 or 1.
In this case, the point estimate can be specified in
estimator.type as either a risk difference (‘risk diff’),
odds ratio (‘odds ratio’), or risk ratio (‘risk ratio’). This
instruction is then passed to makeConfidenceCurves and
calculated there. To specify response rates for binary data into
inputs, the function expects a list with each item giving
the probability of a positive ‘1’ response. For example, a two arm trial
with binary outcome may have resprate = c(0.4, 0.6) with
probability of control response being 0.4 and treatment response being
0.6.
A few notes:
If a ratio estimate is used, it will be converted to the log scale
and the mean value for the results returned by
runSingleTrial is on the log scale. Further, it is very
important to note that it assumes that smaller is
better*. So that 0 is a better result than one. The
dir.benefit field for makeConfidenceCurves is
set to 0, meaning a point estimate of less than zero indicates benefit
to treatment over control.
For risk difference, the linear and not log scale is used. As compared with above, a point estimate of greater than zero indicates benefit to treatment over control.
If the primary outcome is ordinal, the point estimate is a
generalised odds ratio and is calculated by the genodds R
package. In this scheme, smaller is better. If the ordinal outcomes take
of values 1 through 6, it is assumed that 1 is the best outcome and 6 is
the worst. To specify response rates, resprate expects a
list of lists, with each list corresponding to a different arm. Each arm
should have the probability of response for each ordinal value. For a
two-arm trial predicting response on a 6-point ordinal outcome, the
structure is:
resprate = list(control = c(0.1, 0.1, 0.2, 0.5, 0.6, 0.6), treatment = c(0.1, 0.2, 0.3, 0.4, 0.5, 0.6))If the primary outcome is continuous, the point estimate is a
Difference of Means, where a positive value denotes benefit to treatment
over control. In this case, response rates should be fed to
resprate as a list of lists, which each list combining the
mean and standard deviation for each arm. For example:
resprate = list(control=c(0,1), treatment=c(0.2,1)