PACC Causal Discovery

• PACC Discovery is a framework that infers causal relationships with finite-sample guarantees by defining explicit error (ε) and confidence (δ) parameters.
• It adapts established methods such as propensity scores, IV/2SLS, and SCCS under a PAC-like paradigm with clear sample complexity and error bounds.
• The approach provides actionable guidance for fields like epidemiology, economics, and AI/ML by formalizing causal discovery as a decision process.

Probably Approximately Correct Causal (PACC) Discovery refers to a rigorous, resource-aware framework for inferring causal relationships from data with finite, rather than asymptotic, guarantees. Building on the conceptual foundation of Probably Approximately Correct (PAC) learning introduced by Valiant, PACC Discovery centers on sample and computational efficiency, framing causal discovery as a decision process that accepts high—though not perfect—accuracy given realistic resource limits. The framework defines formal error and confidence parameters (ε, δ), and stipulates that a causal discovery algorithm should, with probability at least 1 – δ, infer structural or treatment-causal properties up to an error ε, possibly utilizing diverse method classes such as propensity score, instrumental variables, and even self-controlled case series approaches (Wei et al., 25 Jul 2025).

1. Conceptual Foundations and Formal Framework

The central goal of PACC Discovery is to rationalize the design and evaluation of causal inference algorithms in the finite-sample regime. Unlike classical approaches that rely on large-sample or oracle asymptotic identifiability, PACC Discovery explicitly models the instance space $\mathcal{I}$—typically all possible variable assignments or observations—and a causal model $M$ as a probability distribution over $\mathcal{I}$.

A “causal concept” $c$ is operationalized as a binary property on the model (e.g., the existence of a directed edge, or a nonzero treatment effect), analyzed in terms of discriminability between model pairs that differ only with respect to property $c$. The family $\mathcal{F}_{\delta, c}$ collects all relevant such pairs. The haLLMark of the framework is the requirement that, for every pair in $\mathcal{F}_{\delta, c}$, a learning algorithm $L$—provided with $n = \text{poly}(1/\varepsilon, 1/\delta)$ examples—should correctly identify the presence or absence of $c$ with probability at least $1-\varepsilon$. This mirrors the PAC learning guarantee:

$$\mathbb{P}[\text{error}(L) \leq \varepsilon] \geq 1-\delta$$

By operating over model pairs and providing finite-sample guidance, PACC Discovery unifies the performance analysis of a wide range of causal frameworks.

2. Adaptation of Causal Methods into the PACC Paradigm

PACC Discovery’s generality makes it applicable to numerous established causal inference methods. The framework provides explicit sample complexity and error control for:

• Propensity Score and Covariate Adjustment: The method models data as arising from a joint distribution over covariates $X$, exposure $Z$, and outcome $Y$. The sample complexity for distinguishing causal from noncausal effects using regression-adjusted ATE estimation—when combined with rejection sampling for balancing—is shown to be polynomial in $1/\gamma$, where $\gamma$ is a function of effect size and desired error (Wei et al., 25 Jul 2025).
• Instrumental Variables (IV) / Two-Stage Least Squares (2SLS): Assuming a valid instrument $D$, the 2SLS estimator is shown to be PACC: it discriminates nonzero from zero causal effects in $O(1/(\delta^2 \varepsilon))$ samples, with both Type I and II error probabilities explicitly bounded using Chebyshev’s inequality.
• Self-Controlled Case Series (SCCS): For exposure-outcome analysis under strict assumptions (e.g., no unmeasured time-varying confounding), SCCS is formally analyzed for the first time in the PAC regime. Given $$O\left((1/\log^2(\delta)) \log(1/\varepsilon)\right)$$ examples, the method can reliably distinguish whether an exposure has a causal effect by hypothesis testing on estimated log-relative incidence parameters.

This approach circumvents the reliance on specific data generating process details, instead leveraging worst-case distinguishability in the context of model pairs.

3. Error Bounds, Confidence, and Sample Complexity

A distinctive feature of PACC Discovery is its explicit treatment of computational and sample efficiency:

• The framework mandates that all guarantees and decisions are realized with sample size $n$ that is polynomial in $1/\epsilon$ and $1/\delta$, and, where necessary, in effect size $\delta$ and other method-dependent parameters.
• This methodology delivers actionable finite-sample guidance for practitioners, in contrast to traditional methods that often focus exclusively on large-sample properties.
• Error bounds are provided not only on point estimation (e.g., ATE) but also on the probability of false discovery (both Type I and II errors), enabling confident causal claims at the desired error/coverage levels.

Classical probabilistic tools (Hoeffding, Chernoff, Chebyshev inequalities) are used for deviation bounds, and the sample size formulas are explicit and sharp for each method examined.

4. Family of Causal Concepts and Algorithmic Structure

The formalism introduces instance-based analysis: for a causal concept $c$, the relevant family $\mathcal{F}_{\delta, c}$ encompasses model pairs differing only in $c$. Algorithm 1 in (Wei et al., 25 Jul 2025) lays out a generic PACC discovery routine:

Step	Description
Input	$\varepsilon$, $\delta$, model family $\mathcal{F}_{\delta, c}$
Sample Acquisition	Draw $|S|$ samples from an unknown pair member
Model Evaluation	Score candidate models via likelihood/hypothesis tests
Decision	Output the model with maximal support in $S$
Guarantee	Correct identification with probability $\geq 1-\epsilon$

This structure enables the analysis of arbitrary causal properties (edges, effects, mediation, etc.) and aligns with a worst-case guarantee over plausible instantiations.

5. Theoretical Implications and Guarantees

PACC Discovery enables the formalization of guarantees for complex causal inference scenarios:

• For each analyzed method, finite-sample distinguishability is formally shown between models with and without specified causal features, conditional on relevant assumptions (e.g., instrument validity, confounding, positivity).
• The PAC framework allows explicit quantification of the relationship between effect size, sample requirements, and error tolerance, and supports practical guidance for paper design in observational and quasi-experimental settings.
• Notably, the framework provides the first formal PAC-type guarantee for informal causal discovery methods such as SCCS (Wei et al., 25 Jul 2025), previously lacking rigorous error/sample complexity analysis.

6. Practical Applications and Future Directions

The framework underpins applications in fields characterized by finite, noisy, resource-limited observational data:

• Epidemiology/Public Health: Enables principled sample size calculations for pharmacovigilance, vaccine safety studies, and post-market surveillance.
• Economics: Informs sample requirements in natural experiments and policy impact evaluations with instrumental variable strategies.
• AI/ML and Decision Sciences: Guides the design of interpretable, resource-aware causal analysis modules within autonomous agents.

Future research envisages:

• Extension to Bayesian PACC Discovery: maintaining rigorous error control while propagating prior/posterior uncertainty over broader model spaces.
• PACC analysis for more complex structural features: including mediation, effect modification, and high‐dimensional discovery.
• Integration with fairness and differential prediction assessments, e.g., “probably approximately fair” methods as analogs to probably approximately correct causality.

7. Summary Table: PACC Discovery Guarantees for Key Methods

Method	Assumptions	Sample Complexity Formula	Guarantee
Propensity Score	Overlap, no unmeasured confounding	$O(1/\gamma^3 \cdot \log(1/\gamma))$	PACC on ATE
IV/2SLS	Valid instrument, exclusion	$O(1/(\delta^2 \varepsilon))$	PACC on effect
SCCS	No time-varying confounders	$O((1/\log^2\delta) \log(1/\epsilon))$	PACC on exposure effect

This quantifies the finite-sample, error-controlled nature of PACC Discovery across a variety of established causal methods (Wei et al., 25 Jul 2025).

---

PACC Discovery modernizes the theory and practice of causal inference by supplying formally grounded, finite-sample guarantees for both widely used and emerging methods, marking a shift from reliance on asymptotic identifiability to actionable, resource-constrained reliability. Its formalization of causal discovery as a PAC learning problem delivers practical, interpretable, and robust guarantees suitable for real-world deployment in high-stakes applications.