hypothesize 0.1.dev21

A Python package for comparing groups and measuring associations using robust statistics.

Hypothesize

A Python package for comparing groups and measuring associations using robust statistics.

This package is a port of Rand R. Wilcox's R library WRS. The functions in this repository, as well as the issues of robustness, are described in his book "Introduction to Robust Estimation and Hypothesis Testing".

Please visit the Hypothesize documentation site.

:warning: This repository is still in the early stages of development

Installation

The Hypothesize package is available on PyPI. To install it, simply type:

pip install hypothesize

Expand the following topics to see examples

How to compare two groups

Load data from a CSV or create some example data

from hypothesize.utilities import create_example_data

df=create_example_data(design_values=2)

df.head()

	cell_1	cell_2
0	0.0446518	0.90675
1	0.763458	0.291555
2	0.71039	0.59828
3	0.175208	0.268073
4	0.957819	0.222688

Import the desired function and pass in the data for each group

• This example uses the bootstrapped-t method with 20% trimmed means
• The output is a dictionary containing the results (95% confidence interval, p_value, test statistics, etc...)

from hypothesize.compare_groups_with_single_factor import yuenbt

results=yuenbt(df.cell_1, df.cell_2)

print(results['ci'])

[-0.3115715617702292, 0.10636703554225341]

How to compare groups in a factorial design

Load data from a CSV or create some example data

from hypothesize.utilities import create_example_data

df=create_example_data(design_values=[2,3])

df.head() 

	cell_1_1	cell_1_2	cell_1_3	cell_2_1	cell_2_2	cell_2_3
0	0.0446518	0.90675	0.795696	0.519486	0.333636	0.232153
1	0.763458	0.291555	0.84158	0.0339891	0.511235	0.732503
2	0.71039	0.59828	0.110407	0.898072	0.769496	0.0484005
3	0.175208	0.268073	0.888728	0.287442	0.100153	0.210394
4	0.957819	0.222688	0.834161	0.599158	0.655308	0.203486

Import the desired function and pass in the data

• This example uses a 2-by-3 design
• One approach is to use a set of linear contrasts that will test all main effects and interactions
• Then, the bootstrap-t method and the 20% trimmed mean can be used
• The results are a dictionary of DataFrames that contain various statistics for each factor and the interactions

from hypothesize.compare_groups_with_two_factors import bwmcp

results=bwmcp(J=2, K=3, x=df)

results['factor_A']

	con_num	psihat	se	test	crit_value	p_value
0	0	0.0393584	0.169849	0.231726	3.35959	0.941569

results['factor_B']

	con_num	psihat	se	test	crit_value	p_value
0	0	-0.104506	0.126135	-0.828529	2.4329	0.452421
1	1	-0.0931364	0.151841	-0.613382	2.4329	0.552588
2	2	0.01137	0.135392	0.0839783	2.4329	0.923205

results['factor_AB']

	con_num	psihat	se	test	crit_value	p_value
0	0	-0.100698	0.126135	-0.798336	2.3771	0.410684
1	1	-0.037972	0.151841	-0.250078	2.3771	0.804674
2	2	0.0627261	0.135392	0.463291	2.3771	0.659432

How to compute a robust correlation

Load data from a CSV or create some example data

from hypothesize.utilities import create_example_data

df=create_example_data(design_values=2)

df.head() 

	cell_1	cell_2
0	0.0446518	0.90675
1	0.763458	0.291555
2	0.71039	0.59828
3	0.175208	0.268073
4	0.957819	0.222688

Import the desired function and pass in the data for each group

• One approach is to winsorize the x and y data
• A heteroscedastic method for testing zero correlation is also provided in this package but not shown here
• Please see the function corb which uses the percentile bootstrap to compute a 1-alpha CI and p_value for any correlation
• The output is a dictionary containing various statistics (the winsorized correlation, winsorized covariance, etc...)

from hypothesize.measuring_associations import wincor

results=wincor(df.cell_1, df.cell_2)

print(results['wcor'])

-0.05690314435050796