Stattests#

Google Colab Download - Jupyter Notebook

The Stattests class comprise simple utilities for two-group statistical hypothesis testing.

Loading data#

Throughout this guide we use our demonstration simple_shop dataset. It has already been converted to Eventstream and assigned to stream variable. If you want to use your own dataset, upload it following this instruction.

import numpy as np
from retentioneering import datasets

stream = datasets.load_simple_shop()

General usage#

The primary way to use the Stattests is to call Eventstream.stattests() method. Beforehand, we need to set the following arguments: groups and func. The former defines two user groups to be compared, the latter – a metric of interest to be compared in these two groups.

For our first example, we will split users 50/50 based on the index:

data = stream.to_dataframe()
users = data['user_id'].unique()
index_separator = int(users.shape[0]/2)
user_groups = users[:index_separator], users[index_separator:]

print(user_groups[0])
print(user_groups[1])
array([219483890, 964964743, 629881394, ..., 901422808, 523047643,
       724268790])
array([315196393, 443659932, 865093748, ..., 965024600, 831491833,
       962761227])

Optionally, we can define the names of the groups to be displayed in the method output with the group_names argument.

Let us say we are interested in the proportion of cart events in a user’s path. So the func parameter will look like this:

def cart_share(df):
    return len(df[df['event'] == 'cart']) / len(df)

The interface of the func function expects its single argument to be a pandas.DataFrame that contains a single user path. The function output must be a either a scalar number or a string. For example, if we pick a some_user id and apply cart_share function to the corresponding trajectory, we get the metric value of a single user.

some_user = user_groups[0][378]
cart_share(data[data['user_id'] == some_user])
0.15384615384615385

Let us run the test that is defined by test argument. There is no need to specify a test hypothesis type - when applicable, the method computes the statistics for both one-sided hypothesis tests. Stattests outputs the statistic that could be significant, indicating which of the groups metric value could be greater:

stream.stattests(
    groups=user_groups,
    func=cart_share,
    group_names=['random_group_1', 'random_group_2'],
    test='ttest'
)
random_group_1 (mean ± SD): 0.075 ± 0.095, n = 1875
random_group_2 (mean ± SD): 0.078 ± 0.102, n = 1876
'random_group_1' is greater than 'random_group_2' with p-value: 0.21369
power of the test: 8.85%

The method outputs the test p-value, along with group statistics and an estimate of test power (which is a heuristic designed for t-test). As expected, we see that the p-value is too high to register a statistical difference.

Test power#

Changing the alpha parameter will influence estimated power of the test. For example, if we lower if to 0.01 (from the default 0.05), we would expect the power to also drop:

stream.stattests(
    groups=user_groups,
    func=cart_share,
    group_names=['random_group_1', 'random_group_2'],
    test='ttest',
    alpha=0.01
)
random_group_1 (mean ± SD): 0.075 ± 0.095, n = 1875
random_group_2 (mean ± SD): 0.078 ± 0.102, n = 1876
'random_group_1' is greater than 'random_group_2' with p-value: 0.21369
power of the test: 2.11%

Categorical variables#

We might be interested in testing for difference in a categorical variable - for instance, in an indicator variable that indicates whether a user entered cart state zero, one, two or more than two times. In such cases, a contingency table independence test could be suitable.

Let us check if the distribution of the mentioned variable differs between the users who checked:

  • product1 exclusively

  • product2 exclusively:

user_group_1 = set(data[data['event'] == 'product1']['user_id'])
user_group_2 = set(data[data['event'] == 'product2']['user_id'])

user_group_1 -= user_group_1 & user_group_2
user_group_2 -= user_group_1 & user_group_2
def cart_count(df):
    cart_count = len(df[df['event'] == 'cart'])
    if cart_count <= 2:
        return str(cart_count)
    return '>2'

some_user = user_groups[0][378]
cart_count(data[data['user_id'] == some_user])
'2'
some_user = user_groups[0][379]
cart_count(data[data['user_id'] == some_user])
'0'

To test the statistical difference between the distribution of 0, 1, 2, and >2 categories we apply chi2_contingency test.

stream.stattests(
    groups=(user_group_1, user_group_2),
    func=cart_count,
    group_names=('product_1_group', 'product_2_group'),
    test='chi2_contingency'
)
product_1_group (size): n = 580
product_2_group (size): n = 1430
Group difference test with p-value: 0.00000

In this case, the output contains only the group_names, group sizes and the resulting test statistics. We can see that the variable of interest indeed differs between the exclusive users of two products.