Principles of New Trial Designs and their Practical Implications
3. Possible Approaches in Adaptive Design
3.4. Example 4: Responsive-Adaptive Randomisation
In this design, patients are hierarchically randomised based on their biomarker profile into one of four treatment arms. They start with a set of initial randomisation probabilities and then based on the observed eight-week outcome, new randomisation probabilities are derived. This maximises the chance that the patient receives the treatment that is most effective for him/her.
To be able to make use of response-adaptive randomisation, a few pre-requirements are necessary:
First, such a trial requires a fast dataflow. Data on the eight-week endpoint needs to be rapidly available in the data centre so that they can update the randomisation probabilities guiding new participants entering the trial. This is not always straightforward in the context of a large multicentre trial.
This example had a short endpoint (eight-week). It would not work so well if a Progression Free Survival (PFS) at six months was needed to guide randomisation of new participants.
It is difficult to interpret results beyond estimation. It is difficult to perform comparisons when no longer working with the classical concept of two independent samples. In addition, it can’t be conclusive for an arm that was prematurely terminated and for which little data is available.
When investigators start to realise that more participants are being randomised into a particular treatment arm, recruitment patterns may change during the course of a trial. This introduces operational bias, e.g. sicker participants could enrol earlier and healthier ones could decide to wait for a higher chance of receiving better treatment. In such a setting, blinding is essential but may not always be feasible.