ab-test-toolkit 0.0.15

Toolkit to simulate and analyze AB tests

ab-test-toolkit

Install

pip install ab_test_toolkit

imports

from ab_test_toolkit.generator import (
    generate_binary_data,
    generate_continuous_data,
    data_to_contingency,
    contingency_from_counts,
)
from ab_test_toolkit.power import (
    simulate_power_binary,
    sample_size_binary,
    simulate_power_continuous,
    sample_size_continuous,
)
from ab_test_toolkit.plotting import (
    plot_power,
    plot_distribution,
    plot_betas,
    plot_binary_power,
)

from ab_test_toolkit.analyze import p_value_binary

Binary target (e.g. conversion rate experiments)

Sample size:

We can calculate the sample size required with the function “sample_size_binary”. Input needed is:

• Conversion rate control: cr0

• Conversion rate variant for minimal detectable effect: cr1 (for example, if we have a conversion rate of 1% and want to detect an effect of at least 20% relate, we would set cr0=0.010 and cr1=0.012)

• Significance threshold: alpha. Usually set to 0.05, this defines our tolerance for falsely detecting an effect if in reality there is none (alpha=0.05 means that in 5% of the cases we will detect an effect even though the samples for control and variant are drawn from the exact same distribution).

• Statistical power. Usually set to 0.8. This means that if the effect is the minimal effect specified above, we have an 80% probability of identifying it at statistically significant (and hence 20% of not idenfitying it).

• one_sided: If the test is one-sided (one_sided=True) or if it is two-sided (one_sided=False). As a rule of thumb, if there are very strong reasons to believe that the variant cannot be inferior to the control, we can use a one sided test. In case of doubts, using a two sided test is better.

let us calculate the sample size for the following example:

n_sample = sample_size_binary(
    cr0=0.01,
    cr1=0.012,
    alpha=0.05,
    power=0.8,
    one_sided=True,
)
print(f"Required sample size per variant is {int(n_sample)}.")

Required sample size per variant is 33560.

n_sample_two_sided = sample_size_binary(
    cr0=0.01,
    cr1=0.012,
    alpha=0.05,
    power=0.8,
    one_sided=False,
)
print(
    f"For the two-sided experiment, required sample size per variant is {int(n_sample_two_sided)}."
)

For the two-sided experiment, required sample size per variant is 42606.

Power simulations

What happens if we use a smaller sample size? And how can we understand the sample size?

Let us analyze the statistical power with synthethic data. We can do this with the simulate_power_binary function. We are using some default argument here, see this page for more information.

# simulation = simulate_power_binary()

Note: The simulation object return the total sample size, so we need to split it per variant.

# simulation

Finally, we can plot the results (note: the plot function show the sample size per variant):

# plot_power(
#     simulation,
#     added_lines=[{"sample_size": sample_size_binary(), "label": "Chi2"}],
# )

Compute p-value

n0 = 5000
n1 = 5100
c0 = 450
c1 = 495
df_c = contingency_from_counts(n0, c0, n1, c1)
df_c

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; } </style>

	users	converted	not_converted	cvr
group
0	5000	450	4550	0.090000
1	5100	495	4605	0.097059

p_value_binary(df_c)

0.11824221841149218

The problem of peaking

wip

Contunious target (e.g. average)

Here we assume normally distributed data (which usually holds due to the central limit theorem).

Sample size

We can calculate the sample size required with the function “sample_size_continuous”. Input needed is:

• mu1: Mean of the control group

• mu2: Mean of the variant group assuming minimal detectable effect (e.g. if the mean it 5, and we want to detect an effect as small as 0.05, mu1=5.00 and mu2=5.05)

• sigma: Standard deviation (we assume the same for variant and control, should be estimated from historical data)

• alpha, power, one_sided: as in the binary case

Let us calculate an example:

n_sample = sample_size_continuous(
    mu1=5.0, mu2=5.05, sigma=1, alpha=0.05, power=0.8, one_sided=True
)
print(f"Required sample size per variant is {int(n_sample)}.")

Let us also do some simulations. These show results for the t-test as well as bayesian testing (only 1-sided).

# simulation = simulate_power_continuous()

# plot_power(
#     simulation,
#     added_lines=[
#         {"sample_size": continuous_sample_size(), "label": "Formula"}
#     ],
# )

Data Generators

We can also use the data generators for example data to analyze or visualuze as if they were experiments.

Distribution without effect:

df_continuous = generate_continuous_data(effect=0)
# plot_distribution(df_continuous)

Distribution with effect:

df_continuous = generate_continuous_data(effect=1)
# plot_distribution(df_continuous)

Visualizations

Plot beta distributions for a contingency table:

df = generate_binary_data()
df_contingency = data_to_contingency(df)
# fig = plot_betas(df_contingency, xmin=0, xmax=0.04)

False positives

# simulation = simulate_power_binary(cr0=0.01, cr1=0.01, one_sided=False)

# plot_power(simulation, is_effect=False)

# simulation = simulate_power_binary(cr0=0.01, cr1=0.01, one_sided=True)
# plot_power(simulation, is_effect=False)