Safe Testing for Large-Scale Experimentation Platforms

Abstract: In the past two decades, AB testing has proliferated to optimise products in digital domains. Traditional AB tests use fixed-horizon testing, determining the sample size of the experiment and continuing until the experiment has concluded. However, due to the feedback provided by modern data infrastructure, experimenters may take incorrect decisions based on preliminary results of the test. For this reason, anytime-valid inference (AVI) is seeing increased adoption as the modern experimenters method for rapid decision making in the world of data streaming. This work focuses on Safe Testing, a novel framework for experimentation that enables continuous analysis without elevating the risk of incorrect conclusions. There exist safe testing equivalents of many common statistical tests, including the z-test, the t-test, and the proportion test. We compare the efficacy of safe tests against classical tests and another method for AVI, the mixture sequential probability ratio test (mSPRT). Comparisons are conducted first on simulation and then by real-world data from a large technology company, Vinted, a large European online marketplace for second-hand clothing. Our findings indicate that safe tests require fewer samples to detect significant effects, encouraging its potential for broader adoption.