Abstract:The aim of this paper is to make clear and precise the relationship between the Rubin causal model (RCM) and structural causal model (SCM) frameworks for causal inference. Adopting a neutral logical perspective, and drawing on previous work, we show what is required for an RCM to be representable by an SCM. A key result then shows that every RCM -- including those that violate algebraic principles implied by the SCM framework -- emerges as an abstraction of some representable RCM. Finally, we illustrate the power of this conciliatory perspective by pinpointing an important role for SCM principles in classic applications of RCMs; conversely, we offer a characterization of the algebraic constraints implied by a graph, helping to substantiate further comparisons between the two frameworks.

What problem does this paper attempt to address?

The paper aims to clearly and precisely elucidate the relationship between the Rubin Causal Model (RCM) and the Structural Causal Model (SCM) frameworks for causal inference. Specifically, the authors adopt a neutral logical perspective and, based on previous research, demonstrate how RCM can be represented in the form of SCM. A key result in the paper shows that every RCM, including those that violate the algebraic principles implied by the SCM framework, can be viewed as an abstract form of some representable RCM. Furthermore, the authors facilitate the comparison between the two frameworks by highlighting the importance of SCM principles in classical RCM applications and providing a method to characterize the algebraic constraints implied by graphs. The paper first explores the representational relationship between RCM and SCM, particularly how SCM can represent RCM under the satisfaction of two key principles (compositionality and invertibility). Subsequently, the authors discuss situations where these principles may fail and attribute such failures to the neglect of low-level causal details when abstracting variables. An important theoretical result in the paper (Theorem 1) supports this intuition, stating that every RCM is a constructive abstraction of some RCM that satisfies compositionality and invertibility. The paper also introduces a neutral formal language for reasoning about counterfactual probabilities, which aids in further comparing the two frameworks. For this general language, the authors provide completeness results for all RCMs (Theorem 2) and for RCMs that meet specific conditions (Corollary 2). Additionally, the authors partially address an open question on how to characterize the algebraic constraints implied by specific graphs (Theorem 3), which helps incorporate graphical assumptions into this neutral general language. Overall, through a series of new technical results, the paper clarifies how RCM and SCM are theoretically related while also pointing out the different perspectives the two frameworks take in addressing causal inference problems.

Comparing Causal Frameworks: Potential Outcomes, Structural Models, Graphs, and Abstractions

Beyond Structural Causal Models: Causal Constraints Models

Rubin causal model

Abstraction between Structural Causal Models: A Review of Definitions and Properties

Foundations of Structural Causal Models with Cycles and Latent Variables

Standardizing Structural Causal Models

A Unifying Framework for Causal Analysis in Set-Theoretic Multimethod Research

Causal Inference Grounded in Causal Diagrams: Benefits, Limitations, and Opportunities for the Information Systems Field

Distribution-consistency Structural Causal Models

A Survey of Causal Inference Frameworks

How SCM Causal Diagrams Differ from PLS-PM and CB-SEM Path Diagrams

Credible causal inference beyond toy models

CKH: Causal Knowledge Hierarchy for Estimating Structural Causal Models from Data and Priors

Causal Modeling in Multi-Context Systems: Distinguishing Multiple Context-Specific Causal Graphs which Account for Observational Support

Reflections stimulated by the comments of Shadish (2010) and West and Thoemmes (2010).

Aligning Graphical and Functional Causal Abstractions

Constraint-based Causal Discovery for Non-Linear Structural Causal Models with Cycles and Latent Confounders

Modeling Latent Selection with Structural Causal Models

Causal modelling without introducing counterfactuals or abstract distributions

Causal Imputation for Counterfactual SCMs: Bridging Graphs and Latent Factor Models

Learning Causal Abstractions of Linear Structural Causal Models