Statistical Methods in Online A/B Testing
9
© Georgi Z. Georgiev. All Rights Reserved.
Practitioners and academics alike were battling over which approach is
best overall, or squabbling over the merits of particular applications -
frequentist inference vs. decision-theoretic vs. Bayesian approaches. To
make matters even more confusing, there seemed to be noticeable
schisms within each school of thought.
Most confusing of all, statistics as such turned out to be very context-
dependent - it meant different things in different scientific and business
fields. Practitioners in those fields had, over time, developed somewhat
separate branches of statistics. Therefore, statistics would mean
something different for you depending on whether you come from
physics, medicine, social studies, econometrics, environmental studies, or
industrial quality control.
It was simply a nightmare attempting to navigate this fractured jungle of
jargon, conflicting stances, and math-heavy explanations. Yet, I
persevered! And through painstaking reading, practice,
implementing/coding methods, and countless simulation runs, I was able
to garner a good enough understanding of the matter to begin writing
methodological white papers and in-depth articles, to start delivering
lectures and courses on statistics in A/B testing, and to become a
developer of statistical tools.
From my current position, I see both the immense value of statistical
methods applied to business risk management, estimation and prediction
problems, and the immense harm done by improper applications or
misguided understanding of those same methods. Thus, in-depth
explanations of the practical application of statistical methods, as well as
common errors and how to avoid them, are key elements of this work.
Furthermore, in 2019 the difficulties that I went through are about as
severe as they were a few years before, despite the valiant efforts of some
in the statistics and A/B testing communities. Addressing common
mistakes, misconceptions, and misapplications of statistical methods is,
therefore, a central part of this work.