deepCausal

Non-Linear Structural Equation Modeling with Causal Dantzig in R.

This package extends traditional Structural Equation Models (SEMs) by incorporating neural networks to model non-linear relationships and uses a two-environment objective that balances pooled least-squares error and environment discrepancy.

Installation

This is an R package project. After cloning, you can build and install with your usual tooling (e.g., RStudio Build or devtools):

# install.packages("devtools")
devtools::document()  # if you use roxygen2
devtools::install()

Core API

• train_causal(): Trains a linear or neural model to minimize the objective.

• evaluate_causal_model(): computes MSE/RMSE on a test set.

• define_functional_forms(): Returns linear and neural_network predictors.

• pilot_systems(): Returns toy systems for simulation.

  See function documentation (?train_causal) for details and examples.

  Shiny App

A small Shiny prototype lives in systemsApp.R. It is excluded from the package build (.Rbuildignore) to keep the package lightweight. You can run it directly in an R session:

source("systemsApp.R")

Python Experiments

The Python experiment now lives outside this R package, in a sibling folder: ../deepCausal-python/. Run it from there, for example:

cd ../deepCausal-python
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python study01.py

The R package does not depend on Python.

Development

• Run R CMD check or devtools::check() to validate.
• Tests are recommended for obj_func, train_causal, and evaluate_causal_model.

License

MIT. See LICENSE.