Pearl's Causal Hierarchy

• Pearl’s Causal Hierarchy is a three-level framework that categorizes causal questions into association, intervention, and counterfactual, driving applications in policy and image analysis.
• The framework employs Structural Causal Models to formalize causal reasoning through abduction, action, and prediction, ensuring clear pathways from observation to intervention.
• Its layered design underpins practical tools like do-calculus, d-separation, and mediation analysis, addressing challenges in causal discovery, external validity, and computational complexity.

Searching arXiv for recent and foundational papers on Pearl's Causal Hierarchy. Pearl’s Causal Hierarchy, also called the Ladder of Causation, is a three-level organization of causal questions into Association, Intervention, and Counterfactual. In the formulation emphasized by Pearl, the hierarchy is not merely taxonomic: it expresses a one-way informational ordering captured by the statement that “Questions at level $i$ can only be answered if information from level $i$ or higher is available” (Pearl, 2018). Association concerns statistical relations such as $P(y \mid x)$, intervention concerns post-manipulation quantities such as $P(y \mid \operatorname{do}(x))$, and counterfactuals concern alternate-world quantities such as $P(y_x \mid x', y')$ (Pearl, 2018). Subsequent work has recast this hierarchy in potential-outcomes terms, quantified the information needed to move between rungs, analyzed the computational complexity of reasoning in its languages, and used it as an evaluation lens in domains including causal discovery, image modeling, neuroscience, and vision-language reasoning (Wu et al., 28 Jan 2026, Emadi, 4 May 2026, Dörfler et al., 2024, Peyrard et al., 2020, Zečević et al., 2022, Jalaldoust et al., 10 Nov 2025, Hoang et al., 27 May 2026).

1. The three rungs and their asymmetry

Pearl’s hierarchy consists of three rungs, each associated with a characteristic query syntax and semantics (Pearl, 2018).

Level	Signature queries	Semantics
Association	$P(y \mid x)$, $P(X,Y)$, $E[Y \mid X=x]$	Belief updating from passive observation
Intervention	$P(y \mid \operatorname{do}(x))$, $E[Y \mid \operatorname{do}(X=x)]$	Outcomes under external manipulation
Counterfactual	$i$0, $i$1	Alternate worlds conditioned on factual evidence

At the first rung, association asks how observations of $i$2 change beliefs about $i$3. Typical examples include questions such as “What does a symptom tell me about a disease?” or “What does a survey tell us about election results?” In this regime, data are passive, no manipulation occurs, and tools such as Bayesian networks, deep learning, and other function-fitting approaches compute or approximate conditional probabilities, correlations, and conditional expectations (Pearl, 2018).

At the second rung, intervention asks what happens when a variable is forcibly set by an external action. The operation $i$4 denotes a “surgical” modification of the data-generating mechanism: it removes the usual causes of $i$5 and sets $i$6 by fiat. Questions such as “What if I take aspirin, will my headache be cured?” or “What if we ban cigarettes?” are of this form. A central distinction is that $i$7 generally differs from $i$8, because conditioning on $i$9 does not itself alter the mechanism that produced $P(y \mid x)$0 (Pearl, 2018).

At the third rung, counterfactuals ask what would have happened under alternative actions or conditions, often for a unit about which factual information is already known. The canonical form emphasized by Pearl is

$P(y \mid x)$1

the probability that outcome $P(y \mid x)$2 would have been $P(y \mid x)$3 had $P(y \mid x)$4 been set to $P(y \mid x)$5, given that in fact $P(y \mid x)$6 and $P(y \mid x)$7 were observed. This rung supports questions such as “Was it $P(y \mid x)$8 that caused $P(y \mid x)$9?” and “What if I had acted differently?” (Pearl, 2018).

The hierarchy is asymmetric. Association-level information cannot, in general, answer interventional queries; interventional information cannot, in general, answer counterfactual queries involving retrospection. Pearl accordingly emphasizes that counterfactual models subsume intervention and association, and intervention models subsume association, but not conversely (Pearl, 2018). A potential-outcomes reformulation makes the same point by classifying Layer 2 as functionals of marginal potential-outcomes distributions $P(y \mid \operatorname{do}(x))$0 and Layer 3 as functionals of joint cross-world objects such as $P(y \mid \operatorname{do}(x))$1 or nested counterfactuals like $P(y \mid \operatorname{do}(x))$2 (Wu et al., 28 Jan 2026).

2. Structural causal models and the semantics of the hierarchy

The hierarchy is formalized in Structural Causal Models (SCMs). In Pearl’s canonical notation,

$P(y \mid \operatorname{do}(x))$3

where $P(y \mid \operatorname{do}(x))$4 are endogenous variables, $P(y \mid \operatorname{do}(x))$5 are exogenous variables, $P(y \mid \operatorname{do}(x))$6 are structural assignments of the form

$P(y \mid \operatorname{do}(x))$7

and $P(y \mid \operatorname{do}(x))$8 is the joint distribution over exogenous variables (Pearl, 2018).

SCMs combine three ingredients. First, a causal DAG encodes assumptions about parent-child relations among endogenous variables. Second, structural equations specify how each variable is generated from its parents and exogenous disturbance. Third, interventional and counterfactual semantics specify how the model changes under interventions and how alternate worlds are compared (Pearl, 2018).

In intervention semantics, $P(y \mid \operatorname{do}(x))$9 replaces the structural equation for $P(y_x \mid x', y')$0 with $P(y_x \mid x', y')$1, thereby removing incoming edges into $P(y_x \mid x', y')$2. Potential outcomes such as $P(y_x \mid x', y')$3 are then the values of $P(y_x \mid x', y')$4 in this modified model (Pearl, 2018). In potential-outcomes language, this corresponds to treating Layer 2 quantities as marginals like $P(y_x \mid x', y')$5, $P(y_x \mid x', y')$6, or contrasts such as the average treatment effect, whereas Layer 3 involves joint or nested quantities not recoverable from marginals alone (Wu et al., 28 Jan 2026).

The observational semantics of an acyclic SCM induce factorizations of the form

$P(y_x \mid x', y')$7

under the usual Markov property assumptions. The interventional analogue is the truncated factorization in which the factors for intervened variables are removed and the intervened variables are fixed to their assigned values (Pearl, 2018, Sick et al., 20 Mar 2025).

Counterfactual semantics require more structure than either passive observation or experimentation. Pearl’s framework computes counterfactuals by abduction, action, and prediction: infer exogenous variables or their posterior distribution from factual evidence, modify the model with the desired intervention, then propagate the modified model while holding the abduced exogenous state fixed (Pearl, 2018). In continuous TRAM-DAGs, this procedure is implemented through bijective transformations $P(y_x \mid x', y')$8, so that exogenous values can be recovered pointwise and then reused under the modified intervention, enabling Layer 3 counterfactuals for the continuous case (Sick et al., 20 Mar 2025).

3. Identification, d-separation, and the causal inference engine

Pearl presents SCMs not only as semantics but as an “inference engine” that takes Assumptions, Queries, and Data, and returns an Estimand $P(y_x \mid x', y')$9, an Estimate $P(y \mid x)$0, and Fit indices (Pearl, 2018). The estimand is a formula reducing the target query to observable or experimentally available distributions under the supplied assumptions; the estimate is the numerical quantity computed from finite data; and fit indices encode testable implications, when such implications exist (Pearl, 2018).

A central graphical tool is d-separation, which links graph structure to conditional independencies in observational data. Pearl emphasizes two roles: transparency, because assumptions can be read from the graph, and testability, because some graphs imply conditional independencies that can be empirically checked (Pearl, 2018). This same logic has been extended to cluster DAGs, where relationships among clusters rather than individual variables are specified. In that setting, d-separation is proved sound and complete for probabilistic inference over clusters, and standard do-calculus and ID remain sound and complete for interventional identification at the cluster level (Anand et al., 2022).

The simplest identification formula in the hierarchy is the back-door adjustment. When a covariate set $P(y \mid x)$1 blocks all back-door paths from $P(y \mid x)$2 to $P(y \mid x)$3,

$P(y \mid x)$4

Pearl contrasts this with the purely associational expression

$P(y \mid x)$5

showing that intervention breaks the natural dependence of $P(y \mid x)$6 on $P(y \mid x)$7 (Pearl, 2018). When back-door adjustment is unavailable, do-calculus transforms interventional expressions into observational ones whenever the graph permits and declares failure otherwise. Pearl cites its completeness for identifiability in the causal hierarchy (Pearl, 2018).

Counterfactual computation cannot, in general, be recovered from observational data or even from purely interventional data that do not record the joint behavior of treated and untreated potential outcomes for the same units (Pearl, 2018). A potential-outcomes account sharpens this point: randomization identifies marginal distributions of $P(y \mid x)$8, but not the joint dependence between $P(y \mid x)$9 and $P(X,Y)$0, nor nested objects such as $P(X,Y)$1. Identification at Layer 3 therefore requires additional assumptions such as monotonicity, cross-world independence, association parameters, copulas, data fusion, or structural modeling (Wu et al., 28 Jan 2026).

4. The seven causal tasks beyond model-free prediction

Pearl organizes the practical significance of the hierarchy into seven “sparks” or “pillars,” each marking a task that model-free machine learning cannot generally perform but SCMs can (Pearl, 2018).

The first pillar is encoding causal assumptions through graphs, making assumptions explicit, transparent, and sometimes testable by d-separation. The second is do-calculus and the control of confounding, which enables estimation of $P(X,Y)$2 from observational data under graphical conditions. The third is the algorithmization of counterfactuals, allowing computation of quantities such as $P(X,Y)$3 and “causes of effects” for individuals (Pearl, 2018).

The fourth pillar is mediation analysis, where direct and indirect effects require interventional and counterfactual definitions. Pearl explicitly associates controlled direct effects and natural indirect effects with higher-rung reasoning because they partition effects into mechanisms rather than merely estimating aggregate treatment contrasts (Pearl, 2018). A potential-outcomes restatement classifies the controlled direct effect as Layer 2 but the natural direct and indirect effects as Layer 3 because they involve nested cross-world objects such as $P(X,Y)$4 (Wu et al., 28 Jan 2026).

The fifth pillar is external validity and transportability, where do-calculus is used to correct for domain shifts and sample-selection differences by modeling which mechanisms change and which remain invariant (Pearl, 2018). The sixth is missing data, where recoverability depends on modeling the missingness process causally rather than assuming generic missing-at-random conditions. The seventh is causal discovery, where observed independencies are mapped to equivalence classes of causal graphs consistent with the data (Pearl, 2018).

A related operationalization appears in “A Ladder of Causal Distances,” which defines rung-specific distances between causal models by the causal distributions they induce rather than by graph structure alone. The paper defines observational distance $P(X,Y)$5, interventional distance $P(X,Y)$6, and counterfactual distance $P(X,Y)$7, and proves inequalities

$P(X,Y)$8

and

$P(X,Y)$9

formalizing that higher-rung agreement implies lower-rung agreement (Peyrard et al., 2020).

5. Quantitative, computational, and logical extensions

Later work has made the hierarchy quantitative. “The Causal Description Gap” defines the description length of a query class as the Kolmogorov complexity of its answer oracle and studies the residual bits needed to specify higher-rung answers once lower-rung answers are known (Emadi, 4 May 2026). In that framework, observational knowledge can be extremely cheap while interventional or counterfactual knowledge remains expensive. The paper constructs binary acyclic SCMs for which the observational oracle has constant description length but the single-node interventional oracle has description length $E[Y \mid X=x]$0, and shows that even full hard-do knowledge can leave a $E[Y \mid X=x]$1 counterfactual description gap (Emadi, 4 May 2026). This suggests that Pearl’s asymmetry can be expressed not only as non-reducibility, but as a sharply quantified residual-information requirement.

A complementary line of work studies computational complexity. “From Probability to Counterfactuals” analyzes satisfiability problems across formal languages aligned with the hierarchy and proves a strict growth in complexity for languages with addition and marginalization: observational satisfiability is $E[Y \mid X=x]$2-complete, interventional satisfiability is PSPACE-complete, and counterfactual satisfiability is NEXP-complete (Dörfler et al., 2024). The same paper shows that for full polynomial languages with addition, multiplication, and marginalization, satisfiability at all three levels is succ$E[Y \mid X=x]$3-complete, resolving an open counterfactual case (Dörfler et al., 2024).

Parameterized refinements identify limited tractable regions. “Gateways to Tractability for Satisfiability in Pearl’s Causal Hierarchy” gives an FPT result for the probabilistic linear fragment with parameter $E[Y \mid X=x]$4 and an FPT result for the counterfactual linear fragment with parameter $E[Y \mid X=x]$5, but also proves hardness barriers such as NP-hardness of the causal linear fragment even when $E[Y \mid X=x]$6 and the primal graph consists of vertex-disjoint paths of length $E[Y \mid X=x]$7 (Ganian et al., 11 Nov 2025). Another extension constrains the model class directly: “Probabilistic and Causal Satisfiability: Constraining the Model” shows, for example, that with a fixed graph and compact marginalization, interventional base and linear satisfiability become NEXP-complete, while the polynomial fragment is Succinct-ETR-complete (Bläser et al., 28 Apr 2025).

These developments preserve Pearl’s basic ordering while changing the form in which it is expressed. In the original hierarchy, higher levels require stronger assumptions and richer models (Pearl, 2018). In the quantitative and logical extensions, they also require, in the worst case, additional bits, stronger structural constraints, or higher computational resources (Emadi, 4 May 2026, Dörfler et al., 2024, Ganian et al., 11 Nov 2025, Bläser et al., 28 Apr 2025).

6. Applications, reinterpretations, and contemporary debates

Recent work has used Pearl’s hierarchy as both a modeling framework and a diagnostic tool. In computer vision, “Pearl Causal Hierarchy on Image Data” maps Level 1 to ordinary observational image recognition, Level 2 to interventions on semantic content variables, and Level 3 to same-$E[Y \mid X=x]$8 counterfactual image edits that hold exogenous details fixed while changing content (Zečević et al., 2022). The paper stresses that many image-editing operations that appear “interventional” are more properly counterfactual because they retain the same background, pose, or viewpoint—precisely the same exogenous state required by $E[Y \mid X=x]$9 (Zečević et al., 2022).

In vision-language reasoning, “The Abstraction Gap in Vision-Language Causal Reasoning” uses Pearl’s hierarchy to design the CAGE benchmark and a dual-probe evaluation that separates plausible language from faithful causal structure. The benchmark spans Level 1 observational questions, Level 2 hypothetical actions, and Level 3 counterfactual scene alterations. The paper reports that seven of eight evaluated VLMs exhibit an Abstraction Gap above $P(y \mid \operatorname{do}(x))$0, with text scores around $P(y \mid \operatorname{do}(x))$1--$P(y \mid \operatorname{do}(x))$2 but chain scores below about $P(y \mid \operatorname{do}(x))$3, while one model achieves near-zero gap (Hoang et al., 27 May 2026). The result is not a reformulation of the hierarchy itself; rather, it operationalizes the distinction between sounding causal and generating explicit structure aligned with Levels 2 and 3 (Hoang et al., 27 May 2026).

The hierarchy has also been recast for abstraction and partial model specification. Neural causal abstractions define abstraction mappings that preserve Layer 1, Layer 2, and Layer 3 quantities under suitable invariance conditions, and use Neural Causal Models to learn such abstractions in practice (Xia et al., 2024). Cluster DAGs similarly allow valid L1–L3 reasoning when only cluster-level relations are known, proving soundness and completeness of d-separation, do-calculus, and ID at that level of specification (Anand et al., 2022).

In neuroscience, “A Causal Formulation of Spike-Wave Duality” explicitly invokes the Causal Hierarchy Theorem to argue that no amount of observational data can determine whether neural waves are causally efficacious or epiphenomenal. The paper defines epiphenomenality by the interventional invariance

$P(y \mid \operatorname{do}(x))$4

and derives a Rule-3-based graphical certificate for when variables can be removed from causal explanations without loss of causal explainability (Jalaldoust et al., 10 Nov 2025). This use is characteristic: the hierarchy separates predictive sufficiency from causal sufficiency, so that a variable may improve prediction at Level 1 while remaining irrelevant under intervention at Level 2 (Jalaldoust et al., 10 Nov 2025).

The hierarchy has also been restated in potential-outcomes language. That perspective classifies Layer 1 as functionals of the observed data distribution $P(y \mid \operatorname{do}(x))$5, Layer 2 as functionals of marginal distributions of $P(y \mid \operatorname{do}(x))$6, and Layer 3 as functionals of joint cross-world or individualized counterfactual quantities. It thereby clarifies why randomized experiments identify Layer 2 marginals but do not, by themselves, identify Layer 3 dependence structures such as probabilities of causation, treatment benefit and harm rates, natural direct and indirect effects, or individualized counterfactual outcomes (Wu et al., 28 Jan 2026).

Across these reinterpretations, the core thesis remains unchanged. Association supports prediction from passive observations, intervention supports policy evaluation and planning, and counterfactuals support explanation, attribution, and retrospection (Pearl, 2018). The later literature primarily deepens this thesis: by formalizing its semantics in alternative languages, by quantifying its information gaps, by characterizing its computational limits, and by using it to diagnose where current learning systems remain confined to Level 1 (Wu et al., 28 Jan 2026, Emadi, 4 May 2026, Dörfler et al., 2024, Hoang et al., 27 May 2026).