When running a randomized clinical trial, if there are factors which are known during the study design phase to influence study results (e.g., gender or baseline disease severity), it may be advantageous to utilize a stratified randomization which ensures each prognostic factor is balanced between treatment arms. This can sometime lead to the temptation of stratifying by a large number of factors, but as the number of strata increases the chance of maintaining the desired treatment balance decreases (as it is not always feasible to find subjects for each combination of prognostic factors). Therefore, when wanting to utilize a stratified randomization, one must ask themselves “How many strata are feasible for my study?”

As with most things, there is no one size fits all response as the answer depends on the trial size, the number of study arms, the number of prognostic factors under consideration, and the number of levels for each prognostic factor. Luckily, there are some rules of thumb to keep in mind when determining the answer for your study.

1)  A helpful formula₁ that can guide decision making is:

•  Note: This formula is derived with the intention of keeping the chances of a stratum having fewer than the desired minimum of subjects to <1%.

2)  The minimum number of subjects per stratum should be a multiple of the number of treatment arms in the study. A minimum of 10 – 20 subjects, when feasible, should be considered as the absolute minimum per stratum.

3)  The number of strata in a randomization scheme increases quickly as additional stratification variables are added. Therefore, when in doubt as to whether or not a study can support additional stratification variables, it is best to lean towards fewer prognostic factors to ensure an adequate number of subjects for each stratum.

•  Example: A study stratifying on gender (male/female) has two strata. Meanwhile, a study stratifying on gender and baseline disease severity (mild/moderate/severe) would have 2 x 3 = 6 stratum.

4)  If your study has more prognostic factors/stratum levels than can be supported by the overall sample size, determine which factors are most important to have treatment balance and account for those.

Reference:
[1] Silcocks, P. How many strata in an RCT? A flexible approach. Br J Cancer 106, 1259–1261 (2012). https://doi.org/10.1038/bjc.2012.84

Scott Mollan, Associate Director, Biostatistics, has over 17 years of experience in clinical and non-clinical statistics across the CRO & pharmaceutical industry. Having led studies of all phases (pilot, pivotal, post-market, phases I-IV) and assisting clients during both the pre-submission phase and FDA approval via the NDA/PMA process (2 NDAs/10 PMAs), he has a wealth of experience to draw upon to support clients. Armed with graduate degrees in business and statistics, Mr. Mollan has been able to leverage his understanding of the clinical trial process via a diverse range of indications to publish on the medical device trial process, lung cancer diagnostics, and women’s health while similarly offering industry presentations on missing data analysis strategies and the use of adaptive trial designs within medical devices studies.