Unifying Disparate Frameworks for Causal Inference

44 concepts39 papers1 proposal

Causal inference is currently undergoing a quiet foundational crisis. If you ask an econometrician, a computer scientist, a biostatistician, and a philosopher to define a "cause," you will receive four mathematically distinct answers. Econometrics favors the Potential Outcomes framework; computer science relies on Structural Causal Models (SCMs) and Directed Acyclic Graphs (DAGs); biostatistics often uses specialized instrumental variable methods; and philosophy and formal methods alternate between regularity-based logic and dynamical systems.

This methodological fragmentation means that insights discovered in one academic silo cannot easily be translated into another. An algorithm designed to learn a causal graph cannot ingest an economic model built on potential outcomes, and a logical verification tool in AI safety cannot readily parse statistical confounding.

To resolve this, a major research direction has emerged to build unified causal frameworks. Rather than proposing yet another competing model, this work constructs mathematical and algorithmic "Rosetta Stones" that bridge existing paradigms, proving they are often different facets of the same underlying theory.

Bridging Potential Outcomes and Structural Models

The most prominent divide in causal inference is between the Rubin Causal Model (RCM)—which frames causality around potential outcomes (what would have happened to individual $i$ under treatment $t$)—and Pearl's Structural Causal Models (SCMs), which use systems of equations and graphs.

To reconcile them, researchers have turned to the concept of causal abstraction. In a foundational analysis, the authors of a recent comparative study established a formal logical language showing that RCMs and SCMs are not contradictory. Instead, any potential outcomes model can be mathematically understood as a simplified, zoomed-out abstraction of an underlying SCM.

   [ High-Resolution SCM ]  --> Every mechanism and variable mapped explicitly
             |
             v  (Causal Abstraction Mapping)
             |
   [ Low-Resolution RCM ]  --> Coarsened into individual potential outcomes

To understand why this works, imagine a high-resolution map showing every street and traffic light in a city (the SCM), and a subway map showing only the connections between major stations (the RCM). Fabio Massimo Zennaro formalized this relationship by separating causal abstractions into structural (graph-based) and distributional (outcome-based) layers. By defining exact mathematical constraints (such as functoriality) on how these layers interact, the author demonstrated how detailed micro-level mechanisms can be systematically coarsened into macro-level potential outcomes without losing the validity of the causal interventions.

Category Theory as a Universal Translation Layer

When moving beyond standard statistical models to complex engineering systems, standard graphs break down. Biological networks have feedback loops, and physical systems operate in continuous time via differential equations. To unify these fragmented languages, researchers are leveraging category theory and topos theory to create a shared mathematical interface.

Sridhar Mahadevan introduced Topos Causal Models (TCMs) to generalize standard SCMs. A topos is a mathematical space that acts as a generalized universe for set theory, where the internal logic is fundamentally intuitionistic (non-Boolean). Why does this matter? In standard causal graphs, an intervention is a binary "forcing" of a variable ($do(X=x)$). In a topos, causal mechanisms are treated as sheaves or geometric objects. This allows local causal mechanisms to be glued together into global systems, providing a single standard that accommodates both discrete graphical interventions and continuous-time processes.

Using a similar geometric intuition, Gabriele D'Acunto et al. applied sheaf theory to solve the problem of multi-agent causal translation. When two AI agents observe the same system but use different variable sets or structural assumptions, they cannot directly share insights. By modeling causal knowledge as a sheaf over a topological space of observations, the authors created a translation layer that allows agents to map their localized causal understandings into a coherent global model.

For practical engineering applications, Joseph Halpern et al. developed Generalized Structural-Equations Models (GSEMs). Instead of forcing fields like biology or control theory to rewrite their domain-specific equations into Pearlian structural equations, GSEMs shift the focus to an abstract "interface" of interventions and outcomes. This high-level abstraction allows researchers to apply uniform definitions of actual causality, responsibility, and blame across entirely different computational substrates, from ordinary differential equations (ODEs) to hybrid automata.

Dissolving the Boundaries Between Graphs and Algebra

In everyday practice, causal inference often forces a choice between graphical methods (like the do-calculus) and algebraic methods (like proxy variables or matrix equations).

Sanghack Lee et al. dissolved this boundary by characterizing graphical constraints directly as matrix equations. Their algorithm, ID-ME, utilizes the Moore-Penrose pseudoinverse to map graphical paths to algebraic properties. By doing so, they proved that complex scenarios previously deemed "unidentifiable" under pure do-calculus can be solved by exploiting the underlying linear algebraic structure of the data.

Graphical Constraints (DAGs)  \
                               ===> [ Matrix Equations & Pseudoinverses ] ===> Identifies previously unidentifiable effects
Algebraic Proxies (Matrices)  /

Similarly, causal reasoning has historically required a separate algebraic toolkit from standard probability theory. To bridge this, John M. Winn introduced a graphical notation called "gates." By framing causal interventions as probabilistic constraints within a standard inference network, Winn demonstrated that Pearl’s do-calculus can be fully subsumed by general-purpose probabilistic message-passing algorithms.

On the statistical side, the field has been flooded with ad-hoc matching and weighting estimators. Nathan Kallus provided a unifying mathematical skeleton for these methods by reframing matching as a functional optimization problem. The author proved that disparate algorithms are actually specific instances of a single broader principle: minimizing the worst-case bias in a specific functional space.

Reconciling Logic and Structure in Actual Causation

While statistical frameworks focus on population-level effects (e.g., "Does smoking cause cancer?"), the study of actual causation focuses on specific events (e.g., "Did this short circuit cause this fire?"). Philosophy has long relied on the NESS (Necessary Element of a Sufficient Set) definition, while computer science relies on structural equations.

Sander Beckers bridged this long-standing divide by introducing Counterfactual NESS (CNESS). By formalizing the regularity-based NESS definition within the mathematical framework of SCMs, the author created a hybrid approach that resolves classic preemption paradoxes (where multiple causes compete to produce an effect) without requiring complex temporal or probabilistic additions.

Furthermore, Sander Beckers et al. leveraged Causal Probabilistic (CP) logic to create a parameterized framework. By mapping competing philosophical definitions of actual causation into this single logical language, the authors enabled direct, objective comparison between theories and extended them to handle non-deterministic scenarios and causation by omission.

In the context of Explainable AI (XAI), Tiago de Lima et al. unified logic-based abductive explanations with intervention-based counterfactual reasoning. By representing causal information as propositional "causal bases," the authors provided a unified language for formally verifying AI decisions, reducing complex causal queries to Quantified Boolean Formulas (QBF).

Unifying the Practical Pipeline

The drive toward unification is also transforming the practical machine learning pipeline, where tasks like causal discovery (finding the graph) and causal representation learning (extracting latent variables from raw pixels) have traditionally been treated as isolated problems.

Patrik Reizinger et al. exposed a fundamental duality between these two domains. They introduced the Identifiable Exchangeable Mechanisms (IEM) framework, proving that both causal structure learning and latent representation learning operate on the exact same underlying principle: exploiting non-i.i.d. (exchangeable) multi-environment data. Whether you are trying to find the causal arrow between two visible variables or trying to disentangle latent factors like "position" from raw pixels, the mathematical guarantees of identifiability are identical.

Finally, this theoretical convergence is addressing the "infrastructure gap" highlighted by Iavor Bojinov et al. and Thuc Duy Le et al., who noted that methodological advancements are frequently trapped in language-specific or discipline-specific codebases. To operationalize these unified theories, open-source toolboxes like causal-learn (Yujia Zheng et al.) and Causal Learner (Kui Yu et al.) have emerged. By integrating data generation, structure learning, and evaluation into unified software platforms, these toolkits allow practitioners to seamlessly deploy cross-disciplinary methods without jumping between incompatible programming environments.

Key Foundational Work

Causal Theories of Actions Revisited (2011) — contributed 1 concept elements to this direction.
Directed information and Pearl's causal calculus (2011) — contributed 1 concept elements to this direction.
Causality with Gates (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_framework_fragmentation_9c792f9d,
  title = {causal framework fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-framework-fragmentation/rdr_9c792f9d22ba45c0b0ecac0e5eb816ad},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_framework_fragmentation_9c792f9d_v1,
  title = {causal framework fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-framework-fragmentation/rdr_9c792f9d22ba45c0b0ecac0e5eb816ad/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake