CausalInference.jl

A Julia package for causal inference, graphical models and structure learning. This package contains constraint-based and score-based algorithms for causal structure learning, as well as functionality to compute adjustment sets which can be used as control variables to compute causal effects by regression.

Introduction

The first aim of this package is to provide Julia implementations of popular algorithms for causal structure identification, the PC algorithm, the GES algorithm, an exact score-based method, and the FCI algorithm. The aim of these algorithms is to identify causal relationships in observational data alone, in circumstances where running experiments or A/B tests is impractical or even impossible. While identification of all causal relationships in observational data is not always possible, both algorithms clearly indicate which causal can and which cannot be determined from observational data. Secondly, similarly to DAGitty, covariate adjustment sets for estimating causal effects for example by regression can be computed.

Causal inference is by no means an easy subject. Readers without any prior exposure to these topics are encouraged to go over the following resources in order to get a basic idea of what's involved in causal inference:

• Causal Inference in Statistics: A Primer
• Review of Causal Discovery Methods Based on Graphical Models
• On Pearl’s Hierarchy and the Foundations of Causal Inference

There are also tutorials and examples linked in the navigation bar of this package.

Implementation Details

The PC algorithm was tested on random DAGs by comparing the result of the PC algorithm using the d-separation oracle with the CPDAG computed with Chickering's DAG->CPDAG conversion algorithm (implemented as dsep and cpdag in this package).

See the Library for other implemented functionality.

The algorithms use the SimpleGraph and SimpleDiGraph graph representation of the Julia package Graphs. Both types of graphs are represented by sorted adjacency lists (vectors of vectors in the Graphs implementation).

CPDAGs are just modeled as SimpleDiGraphs, where unoriented edges are represented by a forward and a backward directed edge.

The listing algorithms for adjustment sets are implemented from scratch using an memority efficient iterator protocol to handle large problems.

Performance

The speed of the PC and GES algorithm is comparable with the C++ code of the R package pcalg. The exact score-based algorithm scales to 20-25 variables on consumer hardware, which comes close to the theoretical limits of these approaches.

Plotting

Main package provides a text-based output describing all identified edges for PC and FCI algorithm (plot_pc_graph_text and plot_fci_graph_text, respectively).

In addition, additional plotting backends are supported with lazy code loading orchestrated by Requires.jl. Upon importing of TikzGraphs.jl, additional plotting methods plot_pc_graph_tikz and plot_fci_graph_tikz will be loaded (these are also aliased as plot_pc_graph and plot_fci_graph for backward compatibility). Similarly, upon importing of both GraphRecipes.jl and Plots.jl, additional plotting methods plot_pc_graph_recipes and plot_fci_graph_recipes will be loaded.

At the time of writing (December 2022), TikzGraphs.jl cannot be installed on ARM-based systems, so GraphRecipes.jl + Plots.jl is the recommended plotting backend in such cases.

Contribution

See issue #1 (Roadmap/Contribution) for questions and coordination of the development.

References

• P. Spirtes, C. Glymour, R. Scheines, R. Tillman: Automated search for causal relations: Theory and practice. Heuristics, Probability and Causality: A Tribute to Judea Pearl 2010
• P. Spirtes, C. Glymour, R. Scheines: Causation, Prediction, and Search. MIT Press 2000
• J. Zhang: On the completeness of orientation rules for causal discovery in the presence of latent confounders and selection bias. Artificial Intelligence 16-17 (2008), 1873-1896
• T. Richardson, P. Spirtes: Ancestral Graph Markov Models. The Annals of Statistics 30 (2002), 962-1030
• D. M. Chickering: Learning Equivalence Classes of Bayesian-Network Structures. Journal of Machine Learning Research 2 (2002), 445-498.
• D. Colombo, M. H. Maathuis: Order-Independent Constraint-Based Causal Structure Learning. Journal of Machine Learning Research 15 (2014), 3921-3962.
• B. van der Zander, M. Liśkiewicz, J. Textor: Separators and Adjustment Sets in Causal Graphs: Complete Criteria and an Algorithmic Framework. (https://doi.org/10.48550/arXiv.1803.00116)
• M. Schauer, M. Wienöbst: Causal structure learning with momentum: Sampling distributions over Markov Equivalence Classes of DAGs. (https://doi.org/10.48550/arXiv.2310.05655)