abtestwise 0.1.0

Lightweight toolkit for binary A/B experiment analysis using aggregate counts.

abtestwise

A lightweight Python toolkit for binary A/B experiment analysis using aggregate count data. Version 0.1 combines frequentist and Bayesian summaries for binary proportions.

Install

Install from PyPI:

pip install abtestwise

Development install

To work on the package locally (with the test dependencies):

pip install -e ".[dev]"

Quickstart

from abtestwise import BinaryABTest

test = BinaryABTest.from_counts(
    control_successes=120,
    control_total=1000,
    treatment_successes=145,
    treatment_total=1000,
    prior_alpha=1,
    prior_beta=1,
    n_simulations=100_000,
    credible_interval=0.95,
    seed=42,
)

result = test.run()

print(result.summary())
print(result.prob_lift_above(0.01))

prob_lift_above(0.01) gives the posterior probability that Treatment B improves the metric by more than 1 percentage point.

Do-no-harm checks

prob_no_harm(margin) gives the posterior probability that Treatment B is not worse than Control A by more than margin (in raw decimal units, so 0.005 means 0.5 percentage points). prob_harm_above(margin) is its complement.

result.prob_no_harm(0.005)     # P(lift >= -0.005): B is not worse by more than 0.5pp
result.prob_harm_above(0.005)  # P(lift <  -0.005): B is worse by more than 0.5pp

Raw result values are also available:

result.to_dict()

Plotting

import matplotlib.pyplot as plt

result.plot_lift_distribution()
result.plot_probability_bar()

plt.show()

The lift distribution plot shows posterior lift in percentage points.

The probability bar plot shows:

P(Treatment B > Control A)
P(Control A > Treatment B)

Groups and sign convention

In product A/B testing terms:

• Control (A) is the baseline group.
• Treatment (B) is the test group or variant B.
• Lift is always Treatment B - Control A.
• Positive lift means Treatment B is better than Control A.
• Negative lift means Control A is better than Treatment B.

Scope

Current package scope:

• Binary proportions only.
• Aggregate counts only.
• Two groups only.
• Frequentist: two-sided pooled two-proportion z-test.
• Bayesian: beta-binomial posterior simulation with default prior Beta(1, 1).
• Equal-tailed credible interval.
• Expected loss.
• Practical lift thresholds.
• Do-no-harm probabilities using a user-defined harm margin.
• Simple plots.

Development

Run tests with:

python -m pytest -q