Efficient and unbiased estimation in adaptive platform trials

Before a new therapy is recommended for clinical practice, it will usually have been tested in a randomised clinical trial (RCT). An RCT is an experiment that randomly allocates consented participants to the experimental therapy and to the control, which is the current standard of care. Traditionally RCTs include a control and a single experimental therapy, and a single analysis is performed after the target number of participants has been recruited. However, as was the case during COVID-19 pandemic, multiple experimental therapies may become available simultaneously in which case an efficient design is to have a single RCT that allocates consented participants to the control and the available multiple experimental treatments. Because a single control arm is used for all the experimental treatments, this saves time and other resources compared to having separate RCTs corresponding to different experimental treatments. To enable making important clinical decisions as quickly as possible, for example as was desired during the pandemic because there was no existing efficacious treatment, it is beneficial to include multiple interim analyses to enable dropping early from the RCT the experimental treatments that are not promising or to conclude early that some of the experimental treatments are superior to the control. Also, new experimental treatments may become available while others are still being tested and it is efficient to add them to an existing RCT. New innovative trial designs referred to as adaptive platform trials incorporate these efficiency aspects. They are efficient multi-arm multi-stage RCTs in which a number of experimental therapies are assessed. They include interim analyses, giving the opportunity to stop the trial early with a positive result or due to futility, to drop poorly performing treatments, or add new ones to the trial. They have been used to test new therapies including in a number of COVID-19 RCTs in the UK.

Whenever a statistical analysis is performed, there is a chance to make an incorrect conclusion. With platform trials, there are multiple instances to make an incorrect conclusion. There are multiple interim analyses and in each, an incorrect conclusion can be made. Also, during interim analyses, the multiple experimental treatments in the trial may be compared to select those that continue with further testing and the selection may be by chance. Consequently, appropriate analysis needs to adjust for the number of interim analyses and decisions made at interim analysis (adaptations) so that the trial's results can be interpreted with confidence.

The aim of this project is to derive formulas to summarise the results of a platform trial while adjusting for the trial adaptations during interim analyses. We will focus on deriving formulas that quantify the magnitude of the clinical benefits of experimental treatments over the control, commonly referred to as point and interval estimators. It is important estimates are unbiased to avoid erroneously recommending inferior treatments for clinical practice. The existing formulas for computing estimates following platform trials do not adjust for trial adaptations and so may give biased estimates.

Deriving adjusted estimators is complex. We will build on estimators that have been derived for much simpler setting referred to as phase II/III RCTs. We will also consider several settings encountered in real platform trials such as different ways of measuring a treatment effect and different adaptations and so it will be a big programme of work.

The expected output from the project is that it will be clear how to obtain unbiased estimates following platform trials. This will contribute to the increase in uptake of platform trials. Consequently, better therapies will become available to those who need them more quickly compared to using traditional RCTs.

Adaptive platform trials are efficient multi-arm multi-stage randomised controlled clinical trials (RCTs) in which a number of experimental therapies are assessed. They include interim analyses, giving the opportunity to stop the trial early with a positive result or due to futility, to drop poorly performing treatments, or add new ones to the trial. The efficiency of adaptive platform trials helps produce early answers to important clinical research questions. Although they have been available and in use for some time, they have particularly come of age in the last few years in the design of studies in COVID-19, where the uncertainty over natural disease progression and available treatments and the need to answer research questions as quickly as possible to inform important treatment decisions has made such approaches particularly attractive. Appropriate analysis needs to adjust for trial adaptations. The other complexity is that for experimental treatments added after the trial has started, some of the control arm study participants are non-concurrent because they were recruited before the treatments were added to the trial. Whilst comparing an experimental treatment to a control, it is common to perform an analysis that includes concurrent controls only. In such a case, we will derive new adjusted point and interval estimators by extending adjusted estimators that exist for the simpler setting of phase II/III RCTs. In cases where the treatment effect of the control over the trial time is not expected to change, it is more efficient to include non-concurrent controls in the analysis. A recent network meta-analysis approach includes non-concurrent controls but does not adjust for trial adaptations. We will extend this approach to develop a network meta-analysis approach that adjusts trial for adaptations in order to obtain new adjusted point and interval estimators.