DCT: A Conditional Independence Test in the Presence of Discretization

This repository contains implementation for paper : A Conditional Independence Test in the Presence of Discretization [arXiv]

DCT is a conditional independence test specifically designed for the scenario in which only discretized versions of variables are available. Specifically, DCT tries to recover the covariance matrix $\Sigma$ of the original continuous variables and constructs the relationship $\hat{\Sigma} - \Sigma$, which corresponds to the independence relationship. Correspondingly, DCT uses nodewise regression to construct $\hat{\Omega} - \Omega$, the statistical inference of conditional independence relationship.

How to Install Required Packages

run the code

conda env create -f environment.yml

Then you will have a conda environment named 'causal'. You can further activate the environment by running

conda activate causal

How to Use

We provide two examples of running the test in example_to_use.ipynb and running the PC algorithm with DCT as the test in example_to_use_pc.ipynb.

Our core algorithm is implemented at causal_learn.causallearn.utils.DisTestUtil.py.

Faster and faster

We will soon release the faster implementation of DCT in binary cases.