A fine-grained look at causal effects in causal spaces

1. Problem Statement

Traditional causal inference frameworks (e.g., Structural Causal Models, Potential Outcomes) define causal effects at the level of random variables. While effective for tabular data where variables have clear semantic meanings (e.g., "drug dose" vs. "blood pressure"), this approach struggles in modern data domains like computer vision and natural language processing.

• The Issue: In these domains, raw variables (pixels, tokens) lack the semantic structure required to formulate meaningful causal questions. For instance, asking how a specific pixel causes an image classification is less meaningful than asking how the presence of a cat (an event) affects the classification.
• The Gap: Existing methods often require pre-defining abstraction maps to move from raw data to semantic variables. There is a lack of a rigorous, measure-theoretic framework that defines causal effects directly on events (subsets of the sample space) and $\sigma$-algebras, allowing for fine-grained analysis without committing to a specific representation map.

2. Methodology

The authors build their framework upon Causal Spaces, a recently introduced axiomatization of causality (Park et al., 2023) that generalizes SCMs and potential outcomes using measure theory.

Core Framework

• Causal Space: Defined as a quadruple $(\Omega, \mathcal{H}, P, \mathcal{K})$, where:
  • $(\Omega, \mathcal{H})$ is a product measurable space.
  • $P$ is the observational probability measure.
  • $\mathcal{K} = \{K_S\}_{S \subseteq T}$ is a collection of causal kernels (transition probability kernels). $K_S(\omega, \cdot)$ describes the distribution of the system after intervening on coordinates $S$ to fix them to values $\omega_S$.
• Interventions: An intervention on a set $U$ via a measure $Q$ yields a new measure $P^{do(U,Q)}$ defined by integrating the causal kernel against $Q$.

Proposed Definitions

The paper shifts the unit of analysis from variables to events ($A \in \mathcal{H}$) and $\sigma$-algebras ($\mathcal{F} \subseteq \mathcal{H}$).

1. Binary (Existence) Definitions:

   • Active Causal Effect: An outcome $\omega$ (or event $B$) has an active $U$-causal effect on event $A$ if the interventional probability $K_U(\omega, A)$ differs from the observational probability $P(A)$.
   • Conditional Causal Effect: Extends the above to conditioning on an event $G$ or a $\sigma$-algebra $\mathcal{G}$. It compares conditional measures $P_G(A)$ and $P^{do(U,\delta_\omega)}_G(A)$.
   • Post-Intervention Causal Effect: Characterizes whether an intervention on $U$ has an effect on $A$ after a prior intervention on $V$ has been applied. This captures mediation and context-dependent effects.
   • No/Dormant Effects: The paper also defines "No Causal Effect" (stronger than active) and "Dormant Causal Effect" (where an effect exists only under specific joint interventions, not single ones).

2. Quantifying Measures:

   • Mean Effect Score: For events, defined as $S(A) = f(P^{do}(A)) - f(P(A))$, where $f$ is a user-chosen scale function (e.g., linear or hyperbolic sine) to weight changes near boundaries (0 or 1) differently than changes near 0.5.
   • Maximum Effect Score: Captures the maximal deviation induced by an intervention within a specific set of outcomes.
   • Distributional Scores: For $\sigma$-algebras, the authors propose using a generic difference functional $D_F$ (valued in a normed vector space) to compare the intervention measure and observational measure. This allows for multi-dimensional summaries (e.g., differences in means, variances, or kernel mean embeddings).

3. Key Contributions

1. Event-Level Causality: The primary contribution is the formalization of causal effects at the level of events and $\sigma$-algebras. This allows for causal analysis on semantic concepts (e.g., "text is about politics") directly within the raw data space, bypassing the need for explicit abstraction functions.
2. Fine-Grained Granularity: The definitions distinguish between individual outcomes, sets of outcomes, and entire $\sigma$-algebras. This enables the identification of specific prompts or image regions that drive causal effects, rather than just stating that a variable "has an effect."
3. Theoretical Unification: The paper establishes rigorous links between causal effects and (conditional) independence under intervention measures. For example, it proves that "no active causal effect" is equivalent to independence between the intervened $\sigma$-algebra and the outcome under the interventional measure.
4. Recovery of Standard Metrics: The framework is shown to recover standard treatment effect metrics as special cases:
   • The Average Treatment Effect (ATE) is recovered by defining the difference measure $D_F$ as the difference in means between the control and treatment distributions.
   • Conditional Average Treatment Effects (CATE) and Distributional Treatment Effects (DTE) are similarly recoverable.
5. Post-Intervention Analysis: The introduction of post-intervention causal effects provides a formal tool to analyze causal chains and mediators without relying on structural equations.

4. Results and Illustrations

The paper validates its theory through a running example of Travel Insurance:

• Setup: Variables include Trip Danger ($N, L, H$), Insurance Status ($Y, N$), and Payment ($0, 30, 1000$).
• Binary Findings: Buying insurance ($Ins=Y$) has an active causal effect on the event "Paying 1000 CHF" (reducing probability from 0.00625 to 0). However, this effect is conditional: if the trip has no danger, buying insurance has no causal effect on the payment (both are 0). If the trip is highly dangerous, the effect is significant.
• Quantification:
  • Using a linear scale function, the mean effect score for buying insurance on the "Pay 1000" event is $-0.00625$.
  • Using a non-linear scale function (emphasizing boundary changes), the score is slightly larger in magnitude ($-0.00675$), demonstrating how the framework can prioritize rare events.
  • Distributional Analysis: When analyzing the $\sigma$-algebra of payments, the framework captures that buying insurance increases the mean payment (by 8.45 CHF) but drastically reduces the variance (by removing the risk of the 1000 CHF payout). This highlights the ability to capture multi-faceted causal impacts.

A secondary example involving Language Models demonstrates how the framework can define causal effects on semantic events (e.g., "text is political") without defining a specific variable for "politicalness," treating the prompt as an intervention on the output space.

5. Significance

• Bridging the Semantic Gap: This work provides a mathematical foundation for causal reasoning in high-dimensional, unstructured data (images, text) where traditional variable-based definitions fail. It allows researchers to ask causal questions about "concepts" rather than "pixels."
• Flexibility in Quantification: By introducing scale functions and generic difference functionals, the framework allows domain experts to tailor the definition of "effect strength" to their specific needs (e.g., prioritizing rare catastrophic events vs. average shifts).
• Theoretical Rigor: By grounding causality in measure theory and $\sigma$-algebras, the paper mirrors the development of probability theory (moving from events to random variables), offering a more fundamental and flexible axiomatization than Structural Causal Models.
• Future Directions: The paper explicitly notes that it focuses on population-level definitions. The next logical steps, as identified by the authors, are developing identification strategies and estimation procedures for these event-level quantities from finite samples.

In summary, Park and Zhou offer a paradigm shift from "variable-centric" to "event-centric" causal inference, providing a robust mathematical toolkit for analyzing causality in the complex, semantic-rich data environments of modern AI.