Abstract
A primary object of causal reasoning concerns what would happen to a system under certain interventions. Specifically, we are often interested in estimating the probability distribution of some random variables given that some other variables are set to certain values. Pearl’s do-calculus provides a powerful tool for this task when the exact causal structure is known. This paper generalizes do-calculus to work with Partial Ancestral Graphs (PAGs), which represent Markov equivalence classes of causal structures and are learnable from observational data.
Introduction
In causal inference, the goal is often to determine the effect of an intervention, denoted as $P(Y|do(X))$. Judea Pearl’s do-calculus provides the mathematical foundation for this, but it traditionally requires a fully specified Directed Acyclic Graph (DAG). However, in practice, observational data often cannot distinguish between multiple DAG structures—a phenomenon known as Markov equivalence. Assuming an incorrect DAG can lead to erroneous causal predictions. This research addresses this gap by performing causal inference using only the structural information that is testable from data, represented through Partial Ancestral Graphs (PAGs).
Methodology: From DAGs to PAGs
The methodology transitions from specific DAGs to more general "Ancestral Graphs" through a three-step escalation:
- Maximal Ancestral Graphs (MAGs): These represent a DAG while accounting for latent (unobserved) confounders without requiring explicit modeling.
- Partial Ancestral Graphs (PAGs): These represent an entire Markov equivalence class of MAGs. A PAG identifies the invariant features across all valid graphs in the class (e.g., certain causal directions or the presence of latent confounding).
- Generalized Rules: The paper redefines "graph surgeries" for these objects, introducing the distinction between visible and invisible edges. A visible directed edge is one where the graph structure guarantees no latent confounding exists, which is vital for identification.
The core contribution is Theorem 2, which provides three rules for PAG-based do-calculus. These rules are "unanimously" sound: if a rule applies in the PAG, it is guaranteed to be valid for every single DAG that could possibly fit the observed data.
Illustrative Application
To demonstrate the soundness of the generalized calculus, the paper employs a synthetic illustrative example. This model includes variables such as Income, Smoking, Lung Cancer, Genotype, and Parents' Smoking Habits.
In this scenario, the variable Profession is treated as a latent confounder. The author demonstrates that even with the "weaker" information provided by a PAG (where certain relations are ambiguous), it is often still possible to fully identify post-intervention distributions, such as $P(Lung\ Cancer | do(Smoking))$.
Key Insights and Results
- Testability: By utilizing PAGs, causal assumptions become empirically testable. If data contradicts the conditional independence relations in the PAG, the model is invalidated. This is a significant advancement over traditional do-calculus, which often relies on untestable expert-defined DAGs.
- Visibility as a Criterion: The research highlights that complete graph knowledge is unnecessary for prediction. Identification often hinges on whether specific edges are "visible," allowing them to be treated as direct causal links free of hidden "backdoor" paths.
- Soundness Proofs: The paper successfully proves that the three fundamental rules of do-calculus—ignoring observations, exchanging interventions for observations, and ignoring interventions—have valid, sound analogues in both MAGs and PAGs.
Limitations and Future Work
While the rules are sound (guaranteeing correct answers), they are not yet proven to be complete. There may be cases where a causal effect is theoretically identifiable from a PAG, but these three rules are insufficient to derive it. Additionally, finding a derivation in do-calculus remains a complex search problem. Future work aims to adapt more modern "algebraic" identification methods to PAGs to improve algorithmic efficiency.
Summary updated