Question

Regarding permuted block randomization I read a paper that I cannot find now that reports some problems of doing permuted block randomization within each site, vs. just ensuring overall treatment frequency balance over all sites combined. Does anyone have input about whether stratifying by site should be done? An investigator is worried about there being 5 patients assigned to treatment B and zero to treatment A at a site. We also realize that for very small sites such as those randomizing 3 patients, it is impossible to balance. My on opinion is that I’d rather let things balance out when sites are combined, but there may be a concern at individual sites.

Responses

There are two downsides that immediately come to mind: 1) Overstratification is self-defeating. Consider the extreme case where each stratum enrolls one participant, such that any blocking strategy within strata is equivalent to simple randomization. 2) Since there is a limit to how many variables can be used in useful stratification (and there are some rules of thumb, e.g., https://pubmed.ncbi.nlm.nih.gov/3203527), it may be better to focus on clinical factors (e.g., age, disease severity) that generally have a stronger association with a clinical outcome than site. These are also often the same factors that are used in subgroup analyses, which will benefit from balance (in contrast with site, which is generally not medically/scientifically interesting for subgroup analyses).

Also, just to stir the pot a little, although the general recommendation is to adjust for stratification variables in analyses, not doing so is generally a little bit conservative (e.g., https://pubmed.ncbi.nlm.nih.gov/831755), meaning that it uadjusted analyses are probably just fine (even if a little less powerful).

– Matt Shotwell, Vanderbilt Department of Biostatistics

This paper talks about it a little bit. https://www.appliedclinicaltrialsonline.com/view/an-adaptive-block-randomization-method-when-stratifying-by-investigator-in-small-to-medium-sized-studies

Most trials that I work on are large (n>500), very few sites, and these sites are often chosen under the assumption that they will enroll a lot of patients and enrollment will be equal across sites. So when constructing the randomization list, I stratify it by site and it is often safe to expect overall balance as a result of the site balancing.

In my experience, DSMB boards have always wanted enrollment numbers by site and consider balance there to be very important. So for smaller trials, I end up stratifying by site but assume a smaller block size to increase the chances of site balance, understanding that there is a trade-off with concealment.

– Rameela Raman, Vanderbilt Department of Biostatistics