The Temptation of Overgeneralizing Response-adaptive Randomization

To the Editor—We read with interest the recent article by Proschan and Evans [1] on the use of response-adaptive randomization (RAR) and its potential problems; however, these problems are neither new nor applicable in general to all RAR procedures. As discussed by Robertson et al [2], there is a lingering debate around the value of RAR in clinical trials. While many of the points raised by Proschan and Evans are valid for some RAR procedures, their broad brush diminishes their arguments. This caveat is important to prevent depreciating other existing procedures that do not have these problems.

A discussion of many of the specific points of Proschan and Evans is presented by Robertson et al [2], which we briefly summarize for the points highlighted in the abstract of Proschan and Evans [1]:

1. Bias from temporal trends and inefficiency in treatment effect estimation to account for this: There are power-orientated RAR procedures that seem particularly robust to time trends (see Villar et al [3]). As well, a largely ignored but highly relevant issue to consider is the likely magnitude of a time trend in a specific trial setting.

2. Volatility in sample size distributions that can cause a nontrivial proportion of trials to assign more patients to an inferior arm: As shown in Table 1 of Robertson et al [2], different RAR procedures have very low probabilities of nontrivial sample size imbalances in the wrong direction.

3. The difficulty of validly analyzing results: Standard analysis methods can be used for a large class of RAR procedures (see, eg, the book by Hu and Rosenberger [4]).

More generally, any discussion around the value of a specific type of RAR needs careful consideration of the context for which it is being proposed. The benefits can in certain cases more than outweigh the downsides. For example, if adaptation allows for a faster rolling out of a superior treatment in the context of an infectious disease pandemic, could a potential loss of power or added complexity in the statistical analysis be more important than slowing down the infection rate?

Proschan and Evans use the infamous Extra Corporeal Membrane Oxygenation (ECMO) trial as the key example to discourage the use of RAR in practice. This trial, which took place 35 years ago, has been fully vetted, and advocates of RAR recognize its extreme deficiencies [5]. We would encourage the use of the early controversy around the ECMO trial to learn when and how RAR might be used appropriately, rather than dismissing the use of any kind of RAR on the basis of a single example. Otherwise, following the same line of thought, any failed example of a fixed randomized trial could be used as a reason not to randomize at all.

Finally, we agree there are still open methodological questions surrounding the use of RAR, but researchers are working on solving these problems [4, 6]. By ignoring existing methodological research and focusing on a limited array of RAR procedures, the authors have done a potential disservice to researchers and those who might benefit from the use of RAR in the future.