Sensitivity designs for preventing bias replication in randomized clinical trials

It is common, after a trial is completed, to employ sensitivity analyses to test the extent to which the results depend on various assumptions or conventions. There is a comparable benefit to employing sensitivity designs when planning a trial, so that features that cannot be varied at the analysis stage can instead be varied (e.g., across centres of a multi-centre trial) during the design stage. Design features amenable to such variation include: (1) the specific randomization methods, (2) the duration of follow-up and (3) the use or non-use of a surrogate endpoint as a replacement for a clinical endpoint. Generally, all centres in a given trial, and all trials in a given program, will employ identical protocols. This means that all will be vulnerable to the same types of biases, meaning that a single bias can by itself render all results unreliable. But by varying the randomization techniques, duration and primary endpoint, one can vary also the biases to which the site-specific results are vulnerable. This means that, if a significant result is found, then one can state that either the treatment worked or there were numerous biases (not just one) at play. This of course makes the attribution of the results to the treatments much more plausible and makes the findings much more robust to violations of assumptions.