A/B testing

Statistical Thinking in Python (Part 2)

Justin Bois

Lecturer at the California Institute of Technology

Is your redesign effective?

ch4-1.002.png

Is your redesign effective?

ch4-1.003.png

Null hypothesis

The click-through rate is not affected by the redesign

Permutation test of clicks through

import numpy as np
# clickthrough_A, clickthrough_B: arr. of 1s and 0s

def diff_frac(data_A, data_B):
        frac_A = np.sum(data_A) / len(data_A)
        frac_B = np.sum(data_B) / len(data_B)
        return frac_B - frac_A

diff_frac_obs = diff_frac(clickthrough_A,
                          clickthrough_B)

Permutation test of clicks through

perm_replicates = np.empty(10000)
for i in range(10000):
     perm_replicates[i] = permutation_replicate(
                clickthrough_A, clickthrough_B, diff_frac)

p_value = np.sum(perm_replicates >= diff_frac_obs) / 10000

p_value

0.016

A/B test

Used by organizations to see if a strategy change gives a better result

Null hypothesis of an A/B test

The test statistic is impervious to the change

Let's practice!

Statistical Thinking in Python (Part 2)