eitanhemed/robusta

Statistical hypothesis testing in Python, using R

robusta's Issues

Inferring paired/unpaired t-test from formula

Issue
When fitting a two-samples t-test using the formula interface, it is required to also provide the paired kwarg, although it should be inferred from the formula.

How to reproduce

import robusta as rst
m = rst.groupwise.groupwise_models.T2SamplesModel(data=rst.load_dataset('sleep'), formula='extra~group|ID')
m.fit()

Output
NotImplementedError: Conversion 'py2rpy' not defined for objects of type '<class 'NoneType'>'

Expected Output

Incorrect effect size using groupwise.T2Samples

Issue
Cohen's-d effect size is miscalculated, resulting in larger effect size estimates.

How to reproduce

import robusta as rst
print(rst.groupwise.T2Samples(
    data=rst.load_dataset('sleep'), independent='group', 
    dependent='extra', subject='ID', paired=True, tail='less').report_table().round(2))

t	df	p-value	Cohen-d Low	Cohen-d	Cohen-d High
-4.062	9.0	0.001	-4.986	-2.708	-0.351

Expected Output
In R:

res = data.frame(tidy(t.test(extra ~ group, data=sleep, paired=T, alternative='less')
res$statistic / sqrt(res$parameter) # d = t / sqrt(N) ~ -1.35

Proposed Solution
The problem arises from rst.groupwise.results.TTestResults, where the usage of psych.t2d to calculate effect size is misguided. We should opt to use the 'n1andn2(degrees of freedom based on each group size) kwargs rather thann` (total degrees of freedom in sample), to get the right values.