Approaches to analyzing binary data for large-scale A/B testing.

An industry-academic collaboration was established to evaluate the choice of statistical test and study design for A/B testing in larger-scale industry experiments. Specifically, the standard approach at the industry partner was to apply a -test for all outcomes, both continuous and binary, and to apply naïve interim monitoring strategies that had not evaluated the potential implications on operating characteristics such as power and type I error rates. Although many papers have summarized the robustness of the -test, its performance for the A/B testing context of large-scale proportion data, with or without interim analyses, is needed. Investigating the effect of interim analyses on the robustness of the -test is important, because interim analyses rely on a fraction of the total sample size and one should ensure that desired properties are maintained when a -test is implemented not just at the end of the study, but for making interim decisions. Through simulation studies, the performance of the -test, Chi-squared test, and Chi-squared test with Yate's correction when applied to binary outcomes data is evaluated. Further, interim monitoring through a naïve approach with no correction for multiple testing versus the O'Brien-Fleming boundary are considered in designs that allow early termination for futility, difference, or both. Results indicate that the -test achieves similar power and type I error rates for binary outcomes data with the large sample sizes used in industrial A/B tests with and without interim monitoring, and naïve interim monitoring without corrections leads to poorly performing studies.