DNCIT provides statistical tests for conditional associations between an image $X$ and a scalar $Y$, given a vector-valued confounder $Z$, called deep nonparametric conditional independence test (CIT) (DNCIT), which were developed in the paper ‘Deep Nonparametric Conditional Independence Tests for Images’. The DNCIT maps the image through a embedding map onto a feature representation and applies existing nonparametric CITs to the feature representation and $Y$, given the confounder $Z$. The package consists of a wrapper for potential embedding maps and a selection of existing nonparametric CITs.

You can install the development version of DNCIT from GitHub with:

options(dncit_with_python_env = TRUE)
# install.packages("devtools")
devtools::install_github("MSimnach/DNCIT")

A more detailed instruction on the installation can be found in vignette("Installation"). To incorporate CITs and embedding maps originally implemented in python, the corresponding modules have to be installed in a virtual environment. Therefore, a virtual environment for the python-based CITs is automatically created when loading the package for the first time (by options(dncit_with_python_env = TRUE)), if there exists no virtual environment called ´r-cits´. The virtual environment is created with the following packages: tigramite, scikit_learn, fcit, open_clip. If you want to use other python-based CITs and embedding maps, you can install them in the virtual environment with the following command: reticulate::py_install(packages = “package”, envname = ‘r-cits’). If setting up a virtual environment is not possible, the installation will try to set up a Conda environment. You can find more to the environment setup in zzz.R.