AA test

An A/A test is a variation of an A/B test, the peculiarity of which is that the original is compared with itself, as opposed to an A/B test, which compares samples before and after exposure.

0. Import Libraries

import numpy as np
import pandas as pd

from lightautoml.addons.hypex import AATest
from lightautoml.addons.hypex.utils.tutorial_data_creation import create_test_data

pd.options.display.float_format = '{:,.2f}'.format

np.random.seed(42)  # needed to create example data

def show_result(result):
    for k, v in result.items():
        print(k)
        display(v)
        print()

1. Create or upload your dataset

In this case we will create random dataset with known effect size

If you have your own dataset, go to the part 2

data = create_test_data(rs=52, na_step=10, nan_cols=['age', 'gender'])
data

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry
0	0	0	0	488.00	414.44	NaN	M	E-commerce
1	1	8	1	512.50	462.22	26.00	NaN	E-commerce
2	2	7	1	483.00	479.44	25.00	M	Logistics
3	3	0	0	501.50	424.33	39.00	M	E-commerce
4	4	1	1	543.00	514.56	18.00	F	E-commerce
...	...	...	...	...	...	...	...	...
9995	9995	10	1	538.50	450.44	42.00	M	Logistics
9996	9996	0	0	500.50	430.89	26.00	F	Logistics
9997	9997	3	1	473.00	534.11	22.00	F	E-commerce
9998	9998	2	1	495.00	523.22	67.00	F	E-commerce
9999	9999	7	1	508.00	475.89	38.00	F	E-commerce

10000 rows × 8 columns

2. AATest

2.0 Initialize parameters

info_col used to define informative attributes that should NOT be part of testing, such as user_id and signup_month

info_cols = ['user_id', 'signup_month']
target = ['post_spends', 'pre_spends']

2.1 Simple AA-test

This is the easiest way to initialize and calculate metrics on a AA-test (default - on 2000 iterations)Use it when you are clear about each attribute or if you don’t have any additional task conditions (like grouping)

You can also add some extra arguments to the process(): * plot_set - types of plot, that you want to show (“hist”, “cumulative”, “percentile”) * figsize - size of figure for plots * alpha - value to change the transparency of the histogram plot * bins - generic bin parameter that can be the name of a reference rule, the number of bins, or the breaks of the bins * title_size - size of title for plots

experiment = AATest(info_cols=info_cols, target_fields=target)

results = experiment.process(data, iterations=2000)

show_result(results)

experiments

	random_state	post_spends a mean	post_spends b mean	post_spends ab delta	post_spends ab delta %	post_spends t-test p-value	post_spends ks-test p-value	post_spends t-test passed	post_spends ks-test passed	pre_spends a mean	...	pre_spends ks-test passed	control %	test %	control size	test size	t-test mean p-value	ks-test mean p-value	t-test passed %	ks-test passed %	mean_tests_score
0	1	452.18	452.15	-0.03	-0.01	0.97	0.56	False	False	487.33	...	False	50.00	50.00	5000	5000	0.59	0.49	0.00	0.00	0.52
1	2	452.82	451.50	-1.32	-0.29	0.09	0.18	False	False	487.04	...	False	50.00	50.00	5000	5000	0.44	0.45	0.00	0.00	0.45
2	4	452.41	451.92	-0.50	-0.11	0.53	0.06	False	False	487.20	...	False	50.00	50.00	5000	5000	0.56	0.36	0.00	0.00	0.43
3	5	452.64	451.69	-0.96	-0.21	0.23	0.41	False	False	486.90	...	False	50.00	50.00	5000	5000	0.26	0.40	0.00	0.00	0.35
4	6	452.70	451.63	-1.07	-0.24	0.17	0.53	False	False	487.31	...	False	50.00	50.00	5000	5000	0.21	0.35	0.00	0.00	0.30
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1755	1993	452.29	452.04	-0.24	-0.05	0.76	0.95	False	False	486.96	...	False	50.00	50.00	5000	5000	0.61	0.71	0.00	0.00	0.68
1756	1994	452.56	451.77	-0.78	-0.17	0.32	0.11	False	False	487.12	...	False	50.00	50.00	5000	5000	0.61	0.21	0.00	0.00	0.34
1757	1995	452.30	452.03	-0.26	-0.06	0.74	0.91	False	False	486.94	...	False	50.00	50.00	5000	5000	0.57	0.89	0.00	0.00	0.79
1758	1996	451.89	452.44	0.55	0.12	0.48	0.78	False	False	487.30	...	False	50.00	50.00	5000	5000	0.38	0.86	0.00	0.00	0.70
1759	1997	452.52	451.81	-0.70	-0.16	0.37	0.10	False	False	487.10	...	False	50.00	50.00	5000	5000	0.66	0.40	0.00	0.00	0.49

1760 rows × 26 columns

aa_score

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.00	0.00	0.00	0.00
pre_spends	0.00	0.00	0.00	0.00
mean	0.00	0.00	0.00	0.00

split

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	0	0	0	488.00	414.44	NaN	M	E-commerce	test
1	1	8	1	512.50	462.22	26.00	NaN	E-commerce	test
2	4	1	1	543.00	514.56	18.00	F	E-commerce	test
3	5	6	1	486.50	486.56	44.00	M	E-commerce	test
4	8	4	1	465.50	506.00	66.00	M	Logistics	test
...	...	...	...	...	...	...	...	...	...
9995	9990	0	0	490.00	426.00	NaN	M	Logistics	control
9996	9992	0	0	491.50	424.00	29.00	M	E-commerce	control
9997	9996	0	0	500.50	430.89	26.00	F	Logistics	control
9998	9997	3	1	473.00	534.11	22.00	F	E-commerce	control
9999	9998	2	1	495.00	523.22	67.00	F	E-commerce	control

10000 rows × 9 columns

best_experiment_stat

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	452.22	452.11	-0.11	-0.02	0.89	1.00	False	False
pre_spends	487.09	487.10	0.00	0.00	0.99	1.00	False	False

split_stat

control %              50.00
test %                 50.00
control size            5000
test size               5000
t-test mean p-value     0.94
ks-test mean p-value    1.00
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.98
Name: 60, dtype: object

resume

	aa test passed	split is uniform
post_spends	not OK	OK
pre_spends	not OK	OK

results.keys()

dict_keys(['experiments', 'aa_score', 'split', 'best_experiment_stat', 'split_stat', 'resume'])

results is a dictionary with dataframes as values. * ‘split’ - result of separation, column ‘group’ contains values ‘test’ and ‘control’

* ‘resume’ - summary of all results

* ‘aa_score’ - score of T-test and Kolmogorov-Smirnov test * ‘experiments’ - is a table of results of experiments, which includes - means of all targets in a and b samples, - p_values of Student t-test and test Kolmogorova-Smirnova, - and results of tests (did data on the random_state passes the uniform test) * ‘best_experiment_stat’ - like previous point but only for the best experiment * ‘split_stat’ - metrics and statistics tests for result of split

results['aa_score']

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.00	0.00	0.00	0.00
pre_spends	0.00	0.00	0.00	0.00
mean	0.00	0.00	0.00	0.00

results['resume']

	aa test passed	split is uniform
post_spends	not OK	OK
pre_spends	not OK	OK

2.2 Single experiment

To get stable results lets fix random_state

random_state = 11

To perform single experiment you can use sampling_metrics()

experiment = AATest(info_cols=info_cols, target_fields=target)
metrics, dict_of_datas = experiment.sampling_metrics(data=data, random_state=random_state).values()

The results contains the same info as in multisampling, but on one experiment

metrics

{'random_state': 11,
 'post_spends a mean': 451.8546,
 'post_spends b mean': 452.4745111111112,
 'post_spends ab delta': 0.6199111111112074,
 'post_spends ab delta %': 0.13700464797208323,
 'post_spends t-test p-value': 0.43154056610193947,
 'post_spends ks-test p-value': 0.95721723072851,
 'post_spends t-test passed': False,
 'post_spends ks-test passed': False,
 'pre_spends a mean': 487.2131,
 'pre_spends b mean': 486.9744,
 'pre_spends ab delta': -0.23869999999999436,
 'pre_spends ab delta %': -0.04901695037766718,
 'pre_spends t-test p-value': 0.5271083329122467,
 'pre_spends ks-test p-value': 0.14861030130677552,
 'pre_spends t-test passed': False,
 'pre_spends ks-test passed': False,
 'control %': 50.0,
 'test %': 50.0,
 'control size': 5000,
 'test size': 5000,
 't-test mean p-value': 0.4793244495070931,
 'ks-test mean p-value': 0.5529137660176427,
 't-test passed %': 0.0,
 'ks-test passed %': 0.0,
 'mean_tests_score': 0.5283839938474595}

dict_of_datas[random_state]

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	1	8	1	512.50	462.22	26.00	NaN	E-commerce	test
1	2	7	1	483.00	479.44	25.00	M	Logistics	test
2	5	6	1	486.50	486.56	44.00	M	E-commerce	test
3	6	11	1	483.50	433.89	28.00	F	Logistics	test
4	11	4	1	498.50	516.89	58.00	NaN	E-commerce	test
...	...	...	...	...	...	...	...	...	...
9995	9986	0	0	494.00	432.11	39.00	M	Logistics	control
9996	9989	6	1	466.50	487.44	19.00	F	E-commerce	control
9997	9991	0	0	482.50	421.89	43.00	NaN	Logistics	control
9998	9995	10	1	538.50	450.44	42.00	M	Logistics	control
9999	9998	2	1	495.00	523.22	67.00	F	E-commerce	control

10000 rows × 9 columns

results = experiment.experiment_result_transform(pd.Series(metrics))

results.keys()

dict_keys(['best_experiment_stat', 'best_split_stat'])

results['best_experiment_stat']

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	451.85	452.47	0.62	0.14	0.43	0.96	False	False
pre_spends	487.21	486.97	-0.24	-0.05	0.53	0.15	False	False

results['best_split_stat']

control %              50.00
test %                 50.00
control size            5000
test size               5000
t-test mean p-value     0.48
ks-test mean p-value    0.55
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.53
dtype: object

2.3 AA-test with grouping

To perform experiment that separates samples by groups group_col can be used

info_cols = ['user_id', 'signup_month']
target = ['post_spends', 'pre_spends']

group_cols = 'industry'

experiment = AATest(info_cols=info_cols, target_fields=target, group_cols=group_cols)

results = experiment.process(data=data, iterations=2000)

The result is in the same format as without groups

In this regime groups equally divided on each sample (test and control):

results['split']['industry'].value_counts(normalize=True) * 100

industry
Logistics    50.15
E-commerce   49.85
Name: proportion, dtype: float64

results['split'].groupby(['industry', 'group'])[['user_id']].count()

		user_id
industry	group
E-commerce	control	2493
	test	2492
Logistics	control	2508
	test	2507

show_result(results)

experiments

	random_state	post_spends a mean	post_spends b mean	post_spends ab delta	post_spends ab delta %	post_spends t-test p-value	post_spends ks-test p-value	post_spends t-test passed	post_spends ks-test passed	pre_spends a mean	...	pre_spends ks-test passed	control %	test %	control size	test size	t-test mean p-value	ks-test mean p-value	t-test passed %	ks-test passed %	mean_tests_score
0	0	451.53	452.80	1.27	0.28	0.11	0.19	False	False	487.00	...	False	50.01	49.99	5001	4999	0.36	0.52	0.00	0.00	0.47
1	2	452.53	451.80	-0.73	-0.16	0.35	0.83	False	False	487.19	...	False	50.01	49.99	5001	4999	0.47	0.92	0.00	0.00	0.77
2	3	452.10	452.23	0.13	0.03	0.87	0.85	False	False	487.11	...	False	50.01	49.99	5001	4999	0.90	0.93	0.00	0.00	0.92
3	4	452.18	452.15	-0.03	-0.01	0.97	0.30	False	False	487.20	...	False	50.01	49.99	5001	4999	0.77	0.47	0.00	0.00	0.57
4	7	452.38	451.95	-0.42	-0.09	0.59	0.40	False	False	487.20	...	False	50.01	49.99	5001	4999	0.58	0.50	0.00	0.00	0.53
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1723	1995	452.70	451.63	-1.07	-0.24	0.18	0.14	False	False	487.40	...	False	50.01	49.99	5001	4999	0.14	0.16	0.00	0.00	0.16
1724	1996	452.36	451.96	-0.40	-0.09	0.61	0.81	False	False	487.08	...	False	50.01	49.99	5001	4999	0.78	0.69	0.00	0.00	0.72
1725	1997	452.08	452.25	0.18	0.04	0.82	0.59	False	False	487.04	...	False	50.01	49.99	5001	4999	0.79	0.66	0.00	0.00	0.70
1726	1998	451.96	452.36	0.40	0.09	0.61	0.44	False	False	486.85	...	False	50.01	49.99	5001	4999	0.41	0.54	0.00	0.00	0.50
1727	1999	452.38	451.95	-0.42	-0.09	0.59	0.54	False	False	487.21	...	False	50.01	49.99	5001	4999	0.57	0.62	0.00	0.00	0.60

1728 rows × 26 columns

aa_score

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.00	0.00	0.00	0.00
pre_spends	0.00	0.00	0.00	0.00
mean	0.00	0.00	0.00	0.00

split

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	0	0	0	488.00	414.44	NaN	M	E-commerce	test
1	2	7	1	483.00	479.44	25.00	M	Logistics	test
2	4	1	1	543.00	514.56	18.00	F	E-commerce	test
3	5	6	1	486.50	486.56	44.00	M	E-commerce	test
4	7	11	1	496.00	432.89	57.00	M	E-commerce	test
...	...	...	...	...	...	...	...	...	...
9995	9983	0	0	494.50	428.33	31.00	F	Logistics	control
9996	9984	0	0	460.00	417.11	56.00	M	Logistics	control
9997	9985	0	0	484.00	411.33	52.00	M	E-commerce	control
9998	9991	0	0	482.50	421.89	43.00	NaN	Logistics	control
9999	9994	0	0	486.00	423.78	69.00	F	Logistics	control

10000 rows × 9 columns

best_experiment_stat

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	452.18	452.15	-0.03	-0.01	0.97	0.99	False	False
pre_spends	487.08	487.11	0.03	0.01	0.93	1.00	False	False

split_stat

control %              50.01
test %                 49.99
control size            5001
test size               4999
t-test mean p-value     0.95
ks-test mean p-value    0.99
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.98
Name: 1395, dtype: object

resume

	aa test passed	split is uniform
post_spends	not OK	OK
pre_spends	not OK	OK

2.4 AA with optimize group

If you have many columns for grouping and don’t know which colun or columns will make best result, you can use parametr ``optimize_group=True``. AA-Test will choose optimal number and names of group columns.

You can use columns_labeling to automatically name columns as target and group.

experiment.columns_labeling(data)

{'target_field': ['treat', 'pre_spends', 'post_spends', 'age'],
 'group_col': ['gender', 'industry']}

results = experiment.process(data=data, optimize_groups=True, iterations=2000)

experiment.group_cols

['industry']

show_result(results)

experiments

	random_state	post_spends a mean	post_spends b mean	post_spends ab delta	post_spends ab delta %	post_spends t-test p-value	post_spends ks-test p-value	post_spends t-test passed	post_spends ks-test passed	pre_spends a mean	...	pre_spends ks-test passed	control %	test %	control size	test size	t-test mean p-value	ks-test mean p-value	t-test passed %	ks-test passed %	mean_tests_score
0	0	452.60	451.72	-0.88	-0.19	0.26	0.48	False	False	487.38	...	True	50.00	50.00	5000	5000	0.20	0.25	0.00	50.00	0.23
1	1	452.18	452.15	-0.03	-0.01	0.97	0.56	False	False	487.33	...	False	50.00	50.00	5000	5000	0.59	0.49	0.00	0.00	0.52
2	2	452.82	451.50	-1.32	-0.29	0.09	0.18	False	False	487.04	...	False	50.00	50.00	5000	5000	0.44	0.45	0.00	0.00	0.45
3	3	451.25	453.08	1.83	0.40	0.02	0.08	True	False	486.67	...	False	50.00	50.00	5000	5000	0.02	0.13	100.00	0.00	0.09
4	4	452.41	451.92	-0.50	-0.11	0.53	0.06	False	False	487.20	...	False	50.00	50.00	5000	5000	0.56	0.36	0.00	0.00	0.43
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1995	1995	452.30	452.03	-0.26	-0.06	0.74	0.91	False	False	486.94	...	False	50.00	50.00	5000	5000	0.57	0.89	0.00	0.00	0.79
1996	1996	451.89	452.44	0.55	0.12	0.48	0.78	False	False	487.30	...	False	50.00	50.00	5000	5000	0.38	0.86	0.00	0.00	0.70
1997	1997	452.52	451.81	-0.70	-0.16	0.37	0.10	False	False	487.10	...	False	50.00	50.00	5000	5000	0.66	0.40	0.00	0.00	0.49
1998	1998	452.27	452.06	-0.21	-0.05	0.79	0.86	False	False	486.73	...	True	50.00	50.00	5000	5000	0.42	0.45	0.00	50.00	0.44
1999	1999	451.47	452.86	1.38	0.31	0.08	0.02	False	True	486.75	...	False	50.00	50.00	5000	5000	0.07	0.36	0.00	50.00	0.26

2000 rows × 26 columns

aa_score

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.04	0.04	1.00	1.00
pre_spends	0.05	0.03	1.00	0.00
mean	0.04	0.04	1.00	0.50

split

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	0	0	0	488.00	414.44	NaN	M	E-commerce	test
1	1	8	1	512.50	462.22	26.00	NaN	E-commerce	test
2	4	1	1	543.00	514.56	18.00	F	E-commerce	test
3	5	6	1	486.50	486.56	44.00	M	E-commerce	test
4	8	4	1	465.50	506.00	66.00	M	Logistics	test
...	...	...	...	...	...	...	...	...	...
9995	9990	0	0	490.00	426.00	NaN	M	Logistics	control
9996	9992	0	0	491.50	424.00	29.00	M	E-commerce	control
9997	9996	0	0	500.50	430.89	26.00	F	Logistics	control
9998	9997	3	1	473.00	534.11	22.00	F	E-commerce	control
9999	9998	2	1	495.00	523.22	67.00	F	E-commerce	control

10000 rows × 9 columns

best_experiment_stat

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	452.22	452.11	-0.11	-0.02	0.89	1.00	False	False
pre_spends	487.09	487.10	0.00	0.00	0.99	1.00	False	False

split_stat

control %              50.00
test %                 50.00
control size            5000
test size               5000
t-test mean p-value     0.94
ks-test mean p-value    1.00
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.98
Name: 69, dtype: object

resume

	aa test passed	split is uniform
post_spends	OK	OK
pre_spends	OK	OK

2.5 AA test with quantization

If you want make one column as parameter for quantization, you may use ``quant_field``.

info_cols = ['user_id', 'signup_month']
target = ['post_spends', 'pre_spends']

group_cols = 'industry'
quant_field = 'gender'

experiment = AATest(info_cols=info_cols, target_fields=target, group_cols=group_cols, quant_field=quant_field)

result = experiment.process(data=data, iterations=2000)

result['split'].groupby(['gender', 'industry', 'group'])['user_id'].count()

gender  industry    group
F       E-commerce  test       2261
        Logistics   control    2305
M       E-commerce  control    2240
        Logistics   control    2194
Name: user_id, dtype: int64

show_result(result)

experiments

	random_state	post_spends a mean	post_spends b mean	post_spends ab delta	post_spends ab delta %	post_spends t-test p-value	post_spends ks-test p-value	post_spends t-test passed	post_spends ks-test passed	pre_spends a mean	...	pre_spends ks-test passed	control %	test %	control size	test size	t-test mean p-value	ks-test mean p-value	t-test passed %	ks-test passed %	mean_tests_score
0	0	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
1	2	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
2	7	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
3	8	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
4	15	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
643	1981	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
644	1982	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
645	1984	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
646	1988	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53
647	1998	452.05	452.46	0.41	0.09	0.64	0.91	False	False	486.90	...	False	72.23	27.77	7223	2777	0.37	0.62	0.00	0.00	0.53

648 rows × 26 columns

aa_score

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.00	0.00	0.00	0.00
pre_spends	0.00	0.00	0.00	0.00
mean	0.00	0.00	0.00	0.00

split

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	8193	0	0	494.50	427.11	40.00	F	E-commerce	test
1	8195	0	0	494.00	416.22	41.00	F	E-commerce	test
2	4	1	1	543.00	514.56	18.00	F	E-commerce	test
3	8203	0	0	472.50	412.67	52.00	F	E-commerce	test
4	8205	0	0	460.00	408.22	66.00	F	E-commerce	test
...	...	...	...	...	...	...	...	...	...
9995	9990	0	0	490.00	426.00	NaN	M	Logistics	control
9996	9992	0	0	491.50	424.00	29.00	M	E-commerce	control
9997	9994	0	0	486.00	423.78	69.00	F	Logistics	control
9998	9995	10	1	538.50	450.44	42.00	M	Logistics	control
9999	9996	0	0	500.50	430.89	26.00	F	Logistics	control

10000 rows × 9 columns

best_experiment_stat

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	452.05	452.46	0.41	0.09	0.64	0.91	False	False
pre_spends	486.90	487.60	0.71	0.15	0.09	0.32	False	False

split_stat

control %              72.23
test %                 27.77
control size            7223
test size               2777
t-test mean p-value     0.37
ks-test mean p-value    0.62
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.53
Name: 0, dtype: object

resume

	aa test passed	split is uniform
post_spends	not OK	OK
pre_spends	not OK	OK

2.6 Unbalanced AA test

If you want to perform AA test with unbalanced groups, you can use parametr ``test_size`` to define sizes of test group and control group

info_cols = ['user_id', 'signup_month']
target = ['post_spends', 'pre_spends']

group_cols = 'industry'

experiment = AATest(info_cols=info_cols, target_fields=target, group_cols=group_cols)

result = experiment.process(data=data, test_size=0.3, iterations=2000)

result['split']['group'].value_counts(normalize=True)

group
control   0.70
test      0.30
Name: proportion, dtype: float64

show_result(result)

experiments

	random_state	post_spends a mean	post_spends b mean	post_spends ab delta	post_spends ab delta %	post_spends t-test p-value	post_spends ks-test p-value	post_spends t-test passed	post_spends ks-test passed	pre_spends a mean	...	pre_spends ks-test passed	control %	test %	control size	test size	t-test mean p-value	ks-test mean p-value	t-test passed %	ks-test passed %	mean_tests_score
0	1	451.99	452.57	0.57	0.13	0.50	0.76	False	False	487.11	...	False	70.01	29.99	7001	2999	0.69	0.66	0.00	0.00	0.67
1	2	452.16	452.17	0.01	0.00	0.99	0.64	False	False	487.00	...	False	70.01	29.99	7001	2999	0.73	0.72	0.00	0.00	0.72
2	3	452.22	452.03	-0.20	-0.04	0.82	0.59	False	False	487.18	...	False	70.01	29.99	7001	2999	0.66	0.48	0.00	0.00	0.54
3	4	451.90	452.78	0.88	0.19	0.31	0.24	False	False	487.13	...	False	70.01	29.99	7001	2999	0.54	0.60	0.00	0.00	0.58
4	5	452.64	451.05	-1.59	-0.35	0.06	0.14	False	False	487.20	...	False	70.01	29.99	7001	2999	0.23	0.14	0.00	0.00	0.17
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1705	1995	452.12	452.28	0.16	0.04	0.85	0.63	False	False	487.10	...	False	70.01	29.99	7001	2999	0.89	0.64	0.00	0.00	0.73
1706	1996	452.00	452.54	0.54	0.12	0.53	0.85	False	False	486.94	...	False	70.01	29.99	7001	2999	0.37	0.72	0.00	0.00	0.60
1707	1997	452.29	451.88	-0.40	-0.09	0.64	0.60	False	False	487.04	...	False	70.01	29.99	7001	2999	0.65	0.44	0.00	0.00	0.51
1708	1998	451.78	453.07	1.29	0.28	0.13	0.33	False	False	487.12	...	False	70.01	29.99	7001	2999	0.47	0.66	0.00	0.00	0.60
1709	1999	452.25	451.96	-0.29	-0.06	0.74	0.77	False	False	487.17	...	False	70.01	29.99	7001	2999	0.64	0.71	0.00	0.00	0.69

1710 rows × 26 columns

aa_score

	t-test passed score	ks-test passed score	t-test aa passed	ks-test aa passed
post_spends	0.00	0.00	0.00	0.00
pre_spends	0.00	0.00	0.00	0.00
mean	0.00	0.00	0.00	0.00

split

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	0	0	0	488.00	414.44	NaN	M	E-commerce	test
1	8193	0	0	494.50	427.11	40.00	F	E-commerce	test
2	8192	0	0	487.50	436.78	20.00	M	E-commerce	test
3	2	7	1	483.00	479.44	25.00	M	Logistics	test
4	8200	5	1	486.00	495.00	NaN	M	Logistics	test
...	...	...	...	...	...	...	...	...	...
9995	9994	0	0	486.00	423.78	69.00	F	Logistics	control
9996	9995	10	1	538.50	450.44	42.00	M	Logistics	control
9997	9997	3	1	473.00	534.11	22.00	F	E-commerce	control
9998	9998	2	1	495.00	523.22	67.00	F	E-commerce	control
9999	9999	7	1	508.00	475.89	38.00	F	E-commerce	control

10000 rows × 9 columns

best_experiment_stat

	a mean	b mean	ab delta	ab delta %	t-test p-value	ks-test p-value	t-test passed	ks-test passed
post_spends	452.15	452.19	0.04	0.01	0.96	0.99	False	False
pre_spends	487.09	487.09	-0.00	-0.00	1.00	0.99	False	False

split_stat

control %              70.01
test %                 29.99
control size            7001
test size               2999
t-test mean p-value     0.98
ks-test mean p-value    0.99
t-test passed %         0.00
ks-test passed %        0.00
mean_tests_score        0.99
Name: 1472, dtype: object

resume

	aa test passed	split is uniform
post_spends	not OK	OK
pre_spends	not OK	OK

MDE

this is the boundary value of the effect, for which it makes sense to introduce some changes.

info_cols = ['user_id', 'signup_month']
target = ['post_spends', 'pre_spends']

group_cols = 'industry'
mde_target = 'post_spends'

experiment = AATest(info_cols=info_cols, target_fields=target, group_cols=group_cols)

Single experiment of data splitting for MDE calculation.

P.s. [None] is the number of random state. You can change it like sampling_metrics(data, random_state=42) and get result with [42] instead of [None]

splitted_data = experiment.sampling_metrics(data)['data_from_experiment'][None]
splitted_data

	user_id	signup_month	treat	pre_spends	post_spends	age	gender	industry	group
0	0	0	0	488.00	414.44	NaN	M	E-commerce	test
1	2	7	1	483.00	479.44	25.00	M	Logistics	test
2	5	6	1	486.50	486.56	44.00	M	E-commerce	test
3	7	11	1	496.00	432.89	57.00	M	E-commerce	test
4	9	4	1	470.00	512.11	54.00	M	Logistics	test
...	...	...	...	...	...	...	...	...	...
9995	9994	0	0	486.00	423.78	69.00	F	Logistics	control
9996	9995	10	1	538.50	450.44	42.00	M	Logistics	control
9997	9996	0	0	500.50	430.89	26.00	F	Logistics	control
9998	9997	3	1	473.00	534.11	22.00	F	E-commerce	control
9999	9999	7	1	508.00	475.89	38.00	F	E-commerce	control

10000 rows × 9 columns

splitted_data[mde_target].hist()

<Axes: >

You can evaluate minimum detectable effect for your data. This will be the smallest true effect obtained from the changes, which the statistical criterion will be able to detect with confidence

mde = experiment.calc_mde(data=splitted_data, group_field="group", target_field=mde_target)
mde

(0.88, 0.02)

You can also calculate the amount of data you need to have in order to determine the minimum effect of the test.

experiment.calc_sample_size(data=splitted_data, target_field=mde_target, mde=5)

1949.4372012485414

Chi2 Test

target = ['post_spends', 'pre_spends']
treated_field = 'treat'

experiment = AATest(target_fields=target)

experiment.calc_chi2(data, treated_field)

{'post_spends': 4.2708618195357307e-129, 'pre_spends': 0.3904626181767134}