rankFD for Python

This repository contains a pure Python implementation of the independent one-way calculation in R package rankFD 0.1.1. It estimates weighted or unweighted nonparametric relative effects and provides the Wald-type and ANOVA-type global tests.

🔗 日本語の解説

Installation

Install the latest release from PyPI:

python3 -m pip install rankFD

To install the current development version directly from GitHub:

python3 -m pip install "git+https://github.com/okumuralab/rankFD.git"

To install an editable copy for development:

git clone https://github.com/okumuralab/rankFD.git
cd rankFD
python3 -m pip install -e .

Python 3.10 or newer is required. NumPy and SciPy are installed automatically as dependencies.

Usage

from rankfd import rankFD

x = [1, 2, 3, 4, 5]
y = [2, 3, 4, 5, 6]
z = [7, 8, 8, 9, 10]

result = rankFD(x, y, z)
print(result)                  # SciPy-style concise result
print(result.statistic)        # ANOVA-type statistic
print(result.pvalue)           # ANOVA-type p-value
print(result.wald)
print(result.anova)
print(result.relative_effects)
print(result.confidence_interval)

groups = [x, y, z]
same_result = rankFD(*groups)

The R defaults are preserved:

• effect="unweighted" uses pseudo-ranks.
• hypothesis="H0F" tests equality of distribution functions.
• ci_method="logit" computes range-preserving pointwise intervals.

For the more general null hypothesis stated in relative effects, use hypothesis="H0p". For classical global ranks and sample-size-weighted effects, use effect="weighted".

With effect="weighted" and hypothesis="H0F", the nested result.kruskal agrees with SciPy's tie-corrected Kruskal-Wallis test. With the R default effect="unweighted", rankFD applies its pseudo-rank version instead, which can differ in unbalanced designs.

RankFDResult.statistic and .pvalue refer to the ANOVA-type test, which is the small-sample global procedure recommended by the rankFD methodology. The full results are in .anova and .wald. The returned object can also be unpacked as statistic, pvalue = rankFD(...).

Scope

The API intentionally follows SciPy's independent-sample style: rankFD(*groups). Inputs are one-dimensional numeric arrays. NaNs can be handled with nan_policy="propagate", "omit", or "raise".

This port implements the one-factor calculation. It does not parse R formulas, construct crossed multi-factor designs, perform multiple contrast procedures, or create plots.

Tests

Run the Python tests with:

python3 -m unittest discover -s tests -v

The cross-language suite invokes the installed R rankFD package as an oracle and is skipped automatically when R or the R package is unavailable.

References

Edgar Brunner, Frank Konietschke, Markus Pauly, and Madan L. Puri (2016). “Rank-Based Procedures in Factorial Designs: Hypotheses About Non-Parametric Treatment Effects.” https://doi.org/10.1111/rssb.12222

Edgar Brunner, Frank Konietschke, Arne C. Bathke, and Markus Pauly (2018). “Ranks and Pseudo-Ranks - Paradoxical Results of Rank Tests -.” https://arxiv.org/abs/1802.05650

Georg Zimmermann, Edgar Brunner, Werner Brannath, Martin Happ, and Arne C. Bathke (2021). “Pseudo-Ranks: The Better Way of Ranking?” https://doi.org/10.1080/00031305.2021.1972836

rankFD: Rank-Based Tests for General Factorial Designs. https://cran.r-project.org/package=rankFD