21

Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect
Page 2: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Bayesian AB Testing

Chris Campbell

Page 3: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Bayesian AB Testing

• Use prior information• Prevent peeking problem• Reduce impact of multiple comparisons• Probability of effect more interpretable than threshold

Page 4: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Section title

Page 5: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Search Item Click AB Test

5

Page 6: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Search Item Click

Section title

6

Page 7: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

7

Search Item Clicks

Page 8: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Paired t-test

• Sample size

• Minimum detectable effect

Page 9: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

power.t.test(n = 28, sd = 895, power = 0.9)

delta = 789.5388

9

Power Calculation

Page 10: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Bayesian Estimation Supersedes the t-Test

Page 11: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

library(BEST)BESTmcmc(clk$A - clk$B,

priors = list(muM = 0, muSD = 1400))

11

BEST

Page 12: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

BEST

Posterior Probability

= Prior

Probability +

New Evidence

Page 13: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Probability of Direction

13

Page 14: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

82%

14

Probability of Direction

Page 15: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Highest Density Interval

Page 16: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Highest Density Interval

89% of densityCredible Interval

Page 17: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Region of Practical Equivalence

Page 18: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

ROPE

Overlap of Credible Interval with ROPE

Page 19: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

library(rstanarm)stan_glm(

clicks ~ group +weekday,

data = clk)

19

Bayesian GLM

Page 20: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Better control of additional explanatory variables

Bayesian GLM

Page 21: Bayesian AB Testing · 2019-08-29 · Bayesian AB Testing • Use prior information • Prevent peeking problem • Reduce impact of multiple comparisons • Probability of effect

Summary

• Large quantity of modelling methods available• Intuitive interpretation

21