Causal Inference Without Structural Graphs

16 concepts4 papers1 proposal

For decades, the dominant paradigm in artificial intelligence and statistics has treated causality as a property derived from rigid geometric structures. In Judea Pearl’s structural causal models (SCMs), you draw a Directed Acyclic Graph (DAG) where arrows represent functional dependencies, and you define an "intervention" as a physical act that slices away these arrows to force a variable's value.

While highly successful, this graph-dependent view introduces severe limitations. It struggles to handle systems with cyclic feedback loops, systems where the exact graph topology is unknown, or scenarios requiring rich logical reasoning about beliefs, disjunctions, and counterfactuals.

A growing body of research addresses these limitations by decoupling causal reasoning from fixed structural graphs. By reframing causality through the lenses of information theory, modal logic, and possible-world semantics, these frameworks demonstrate that the core machinery of causal inference—such as the do-calculus and actual cause definitions—can operate completely graph-free.


Generalizing Statistical Interventions via Information Fields

To understand how we can perform causal inference without a graph, we must first understand what a graph actually represents: it is a visual shorthand for constraints on information flow. If there is no arrow from $X$ to $Y$, it means $Y$'s decision-making policy cannot see $X$.

In Generalized Do-Calculus Without Graphs, Benjamin Heymann et al. replace the explicit arrows of a DAG with Witsenhausen’s intrinsic model from stochastic control. Instead of variables organized by topological ordering, the system is built from $\sigma$-fields (information fields) that explicitly define what an agent or mechanism knows when making a decision.

Traditional DAG:       X ----> Y ----> Z  (Rigid topology)

Graph-Free Fields:    [Information Field X] ⊆ [Information Field Y]
                      Causal relations defined by policy dependencies, 
                      allowing for cycles and unknown structures.

By framing causality around informational relations and decision-making policies rather than fixed functional arrows, the authors generalize Pearl's do-calculus. They establish that foundational causal concepts like conditional blocking, directional separation ($d$-separation), and topological separation are mathematically equivalent within an information-theoretic setting. This allows researchers to compute the effects of interventions even in complex systems featuring cyclic dependencies or entirely unknown causal structures.

Taking a complementary logical approach, Yusuke Kawamoto et al. abstract graph-based causality into modal logic. In Formalizing Statistical Causality via Modal Logic, the authors introduce StaCL (Statistical Causality Language), which treats individual causal structures as "possible worlds" within a Kripke model.

Instead of proving that a statistical adjustment works for one specific graph, modal operators allow researchers to reason about classes of causal structures simultaneously. If a set of different topologies shares the same local characteristics (such as specific parent-child relationships), the logic can prove the validity of a causal inference method across all of them at once. This shifts the focus from a single, fixed graph to invariant topological properties.


Defining Actual Cause Without Interventions

The graph-dependent paradigm also dominates the definition of "actual cause"—the task of determining whether a specific event caused another (e.g., "Did the driver's texting cause the accident?"). The standard Halpern-Pearl definition relies on testing counterfactuals by intervening on structural equations.

However, Joseph Halpern et al. show in Causality Without Causal Models that actual causality does not inherently require structural equation models. The authors abstract the definition of causality so that it can be applied to any framework equipped with formal counterfactual semantics, such as philosophical possible-world models.

By decoupling actual cause from the mathematical straightjacket of SEMs, the authors unlock causal reasoning for far more expressive languages. For instance, standard causal graphs cannot easily evaluate whether a disjunction ("Event A or Event B") or an agent's internal belief system caused an outcome. By evaluating counterfactuals over generalized possible worlds rather than rigid graph mutations, actual causality can be rigorously analyzed in settings that include backtracking counterfactuals and complex logical formulas.

Structural Equation Approach (Rigid):
  Mutate equation: Y := f(X, U) becomes Y := x
  Check downstream effects on a fixed graph.

Possible-Worlds Approach (Flexible):
  Evaluate counterfactuals across a space of valid worlds.
  Supports complex logic: "If Agent believed A, or if B had occurred..."

Going a step further, Carlos Aguilera-Ventura et al. challenge the idea that physical "intervention" must be the foundational primitive of causal reasoning. In A Non-Interventionist Approach to Causal Reasoning Based on Lewisian Counterfactuals, the authors formalize actual cause using a non-interventionist framework rooted in David Lewis’s philosophy of counterfactuals.

The authors achieve this by designing a two-dimensional semantics that decomposes states into two components:

  1. Propositional valuations: The actual truth values of variables in a given state.
  2. Causal bases: The underlying laws or mechanisms that govern how those variables relate.

Instead of simulating an intervention by forcefully overriding a variable and breaking a causal arrow, this framework navigates alternative worlds by evaluating the closeness of their causal bases. This allows actual cause to be defined purely through the expressive power of counterfactual conditionals, proving that the concept of an external intervention is not a mandatory prerequisite for causal understanding.


Why It Works: The Core Intuition

The reason these graph-free approaches work is that graphs are a representation of causality, not causality itself.

At its core, causal inference relies on two things: knowing what information is available to a mechanism at the time of an event, and knowing how the system responds to hypothetical changes.

  1. Information over Topology: As demonstrated by the integration of Witsenhausen's intrinsic models, tracking $\sigma$-fields allows us to map out exactly who knows what, and when. This captures the exact same informational boundaries that $d$-separation captures on a graph, but without requiring an acyclic drawing.
  2. Logic over Equation Slicing: As shown by the possible-world and modal logic frameworks, an intervention is mathematically equivalent to shifting to an alternative state or world where certain facts are altered while preserving a core set of background laws (the causal basis).

By replacing the physical metaphor of "cutting graph edges" with the mathematical framework of information fields and modal logic, this research direction provides a more flexible, expressive, and mathematically unified foundation for causal AI.

Key Foundational Work

Generalized Do-Calculus Without Graphs (2020) — contributed 6 concept elements to this direction.
Formalizing Statistical Causality via Modal Logic (2022) — contributed 2 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_intervention_free_actual_cause_definitio_f8c311e7,
  title = {intervention-free actual cause definition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intervention-free-actual-cause-definition/rdr_f8c311e76531450c89c6fe6166b2e18d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_intervention_free_actual_cause_definitio_f8c311e7_v1,
  title = {intervention-free actual cause definition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intervention-free-actual-cause-definition/rdr_f8c311e76531450c89c6fe6166b2e18d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake