Sample code that seeks to discover underlying causality by Lingam from table data (consisting solely of continuous variables) generated by the following DAG with unobservable x6.

• lingam.BottomUpParceLiNGAM is sensitive to violating its assumption that there is no hidden variable that affects a non-exogenous variable.
• in this case, lingam.RCD (repetitive causal discovery) could properly identify the true DAG.
  • whether it works depends on the distribution of both error terms and hidden variable(s) as well as on their variances.
  • in particular, a gaussian error term makes it difficult for lingam.RCD to uncover causality.

• scr/
  • run_models.py contains functions that load dataset, estimate causality, save resulting DAG as png.
  • fns.py contains supplementary functions for it.
• graph/
  • pairplot_**.png is a pairplot for the dataset whose error terms follow ** distribution.
  • rcd_est_(distribution).png is the graphical counterpart of the estimated adjacency matrix given by model.fit.
  • rcd_est_(distribution)_dag(ranking)_freq=(frequency).png is one of DAGs estimated by the result of model.bootstrap.

• Windows 10 (or 11) + WSL2 + Ubuntu 20.04
  • python 3.10.7 (pyenv 2.3.2) + poetry (1.1.11)
• Ubuntu 22.04
  • python 3.11.0 (pyenv 2.3.7) + poetry (1.2.2)

Follow a typical routine to set up a virtual environment by pyenv + poetry.