PyData Amsterdam 2017
A lot of theory is available on how the statistics of A/B testing could be improved using Bayesian statistics. In this talk I will discuss several theoretical problems and I will share my experiences on whether they actually impact A/B testing in practice. This will be demonstrated using hierarchical models build with pymc. Finally, I will share how I successfully implemented this into business.
I will first shortly discuss frequentist calculation of an A/B test, and three problems: the normal distribution instead of the beta distribution, multiple comparison problem and biased stopping times. Using these topics, I will shortly introduce Bayesian statistics and more specifically hierarchical Bayes, by using examples in pymc. I will then share whether these topics actually have direct implications for testing in practice and illustrate why several aspects hardly change the decisions made. I will then focus on one of the most important aspects from a business perspective: when to stop an insignificant test. I will present the stopping rule I currently use, explain how this works in practice and how this relates to solving the theoretical problems.