abtoolkit 1.3.2

Package with tools for AB testing

ABToolkit

Set of tools for AA and AB tests, sample size estimation, confidence intervals estimation. For continuous and discrete variables.

Install using pip:

pip install abtoolkit

Continuous variables analysis

Sample size estimation:

from abtoolkit.continuous.utils import estimate_sample_size_by_mde

estimate_sample_size_by_mde(
    std=variable.std(),
    alpha=alpha_level, 
    power=power, 
    mde=mde,
    alternative="two-sided"
)

AA and AB tests simulation:

Using abtoolkit.continuous.simulation.StatTestsSimulation class you can simulate and check different stat-test, compare them in terms of stat test power to choose the best test for your data. As result of simulation for each stat test you will get the 1-st Type error estimation with confidence interval, 2-nd Type error estimation with confidence interval and plot of p-value distribution for different tests.

from abtoolkit.continuous.simulation import StatTestsSimulation
simulation = StatTestsSimulation(
        variable,
        
        stattests_list=["ttest", "diff_ttest", "regression_test", "cuped_ttest", "did_regression_test",
                        "additional_vars_regression_test"],
                        
        alternative=alternative,
        experiments_num=experiments_num,
        treatment_sample_size=sample_size,
        treatment_split_proportion=0.5,
        mde=mde,
        alpha_level=alpha_level,

        previous_values=previous_value,
        cuped_covariant=previous_value,
        additional_vars=[previous_value],
    )
simulation.run()  # Run simulation
simulation.print_results()  # Print results of simulation
simulation.plot_p_values()  # Plot p-values distribution

Output:

Full example of usage you can find in examples/continuous_var_analysis.py script.

Next stat tests implemented for treatment effect estimation:

• T-Test - estimates treatment effect by comparing variables between treatment and control groups.
• Difference T-Test - estimates treatment effect by comparing difference between actual and previous values of variables in treatment and control groups.
• Regression Test - estimates treatment effect using linear regression by tested predicting variable. Fact of treatment represented in model as binary flag (treated or not). Weight for this flag show significant of treatment impact. y = bias + w * treated
• Regression Difference-in-Difference Test - estimates treatment effect using linear regression by predicting difference between treatment and control groups whist represented as difference between current variable value and previous period variable value (two differences). Weight for treated and current variable values shows significant of treatment. y = bias + w0 * treated + w1 * after + w2 * treated * after
• CUPED - estimates treatment effect by comparing variables between treatment and control groups and uses covariant to reduce variance and speedup test. y = y - Q * covariant, where Q = cov(y, covariant) / var(covariant). Cuped variable has same mean value (unbiased), but smaller variance, that speedup test.
• Regression with Additional Variables - estimates treatment effect using linear regression by predicting tested variable with additional variables, which describe part of main variable variance and speedup test. Fact of treatment represented in model as binary flag (treated or not). Weight for this flag show significant of treatment impact. y = bias + w0 * treated + w1 * additional_variable1 + w2 * additional_variable2 + ...

Discrete variables analysis

Sample size estimation:

from abtoolkit.discrete.utils import estimate_sample_size_by_mde

estimate_sample_size_by_mde(
    p, 
    sample_size, 
    alpha=0.05,
    alternative="two-sided"
)

AA and AB tests simulation:

from abtoolkit.discrete.simulation import StatTestsSimulation

sim = StatTestsSimulation(
    count=variable.sum(),
    objects_num=variable.count(),
    stattests_list=["conversion_ztest", "bayesian_test", "chi_square_test"],
    alternative=alternative,
    experiments_num=experiments_num,  # Run each stattest 10 times
    treatment_sample_size=sample_size,
    treatment_split_proportion=0.5,
    mde=mde,  # Trying to detect this effect (very big for our simulated data)
    alpha_level=alpha_level,  # Fix alpha level on 5%
    power=power,
    bayesian_prior_positives=1,
    bayesian_prior_negatives=1,
)
info = sim.run()  # Get dictionary with information about tests
sim.print_results()  # Print results of simulation
sim.plot_p_values()  # Plot p-values distribution

Output:

Next stat tests implemented for treatment effect estimation:

• Conversion Z-Test estimates treatment effect on conversion variable using z-test
• Bayesian Test estimates probability of difference between conversions according to prior knowledge
• Chi-Square Test estimates the significance of association between two categorical variables

Another tools

Central Limit Theorem check

Helps you check if your variable meets the Central Limit Theorem and what sample size you need for it to meet.

from abtoolkit.utils import check_clt
import numpy as np

var = np.random.chisquare(df=2, size=10000)
p_value = check_clt(var, do_plot_distribution=True)

---

You can find examples of toolkit usage in examples/ directory.