Can analysis for experiments that have 3 variations be more precise (~33.33% instead of 34/33/33%)? More generally, can experiments with odd-numbers of variations be more evenly split with greater decimal precision?
For a 3-way experiment, we currently split traffic equally at ~33.33% (if not to a greater decimal precision) using an in-house enrollment system. However, when we use GrowthBook for analysis, GrowthBook does not accept values for traffic splits beyond the first two decimals (see image), so in a way, “even split” in GrowthBook is not really even. As a result, when we analyze our experiments, we often get a SRM Warning, esp since our tests have large sample sizes and p-values are a function of the sample size. While the SRM Warning is statistically true, we find the warning is to be conceptually misleading because the actual traffic split (of ~33.33%) is more even than GrowthBook’s expected “even split” of 34/33/33.
We’re concerned that this warning will limit our Data Science team’s willingness to trust GrowthBook for analyses of large experiments and for tests that have odd-numbers of variations. Would it be possible for greater decimal precision on traffic splits? Thanks for all of your hard work on developing a great product!
07/06/2022, 6:49 PM
Yes, this is something we're working on now to fix. Hope to land the new UI by end of this week