Abstract
Traditional Structural Causal Models (SCMs) and the Potential Outcomes (PO) framework rely on a deterministic consistency rule: if a treatment $T=t$ is observed alongside outcome $Y=y$, then the potential outcome $Y(t)$ must be exactly $y$. This paper identifies a "degenerative counterfactual problem" arising from this rule, where transient noise is erroneously treated as a fixed individual trait. We propose Distribution-consistency Structural Causal Models (DiscoSCMs), which relax this constraint to a distributional equivalence ($Y(t) \stackrel{d}{=} Y$). By decoupling stable unit traits from exogenous noise, DiscoSCMs enable more nuanced counterfactual reasoning, particularly in industrial applications like personalized marketing and medicine.
Introduction
In causal inference, the ability to reason about "what if" scenarios—counterfactuals—is paramount. However, the foundational consistency rule creates a logical trap. Consider a student with average ability who, through sheer luck, achieves a perfect test score. Under standard SCMs, if we were to intervene and have that student retake the test under identical conditions, the model insists they must achieve a perfect score again. It treats "luck" as a deterministic, inherent property of the individual.
This determinism leads to the "degenerative counterfactual problem," making it difficult to model the joint distribution of counterfactuals. In real-world settings, such as personalized incentives on digital platforms, this lack of flexibility prevents models from distinguishing between a user's stable characteristics and transient environmental noise, leading to sub-optimal decision-making.
The DiscoSCM Framework
To address these limitations, the authors introduce the Distribution-consistency assumption. Instead of forcing the potential outcome to equal the observed outcome, they assume they are equivalent in distribution.
Decoupling Randomness
The key technical innovation in DiscoSCMs is the explicit separation of two types of randomness:
- Unit Selection Variable ($U$): Represents the individual's stable identity, traits, or latent characteristics.
- Exogenous Noise Variable ($E$): Represents uncontrollable, transient factors (e.g., "good fortune" or measurement error).
By disentangling these variables, DiscoSCMs allow for a "new draw" of noise in counterfactual scenarios, aligning better with scientific intuition and the principle of regression to the mean.
The Probability of Consistency (PC)
The paper introduces a new identifiable causal parameter: the Probability of Consistency (PC). This metric measures the likelihood that an individual would achieve the same outcome if a scenario were repeated. While standard models fix PC at 1 (absolute determinism), DiscoSCMs allow PC to vary between 0 and 1, providing a quantitative measure of real-world uncertainty at the individual level.
Theoretical Contributions and Identifiability
The authors re-evaluate the "Ladder of Causation" within this new framework. They prove that for Layer 1 (Association) and Layer 2 (Intervention), DiscoSCMs are backward compatible with traditional SCMs, yielding identical results. However, at Layer 3 (Counterfactuals), DiscoSCMs offer a significantly more expressive lens, allowing for probabilistic rather than deterministic individual-level predictions.
Furthermore, the paper establishes several theorems (Theorems 5-8) demonstrating that individual-level and population-level counterfactual quantities are identifiable using propensity scores and observed data. This ensures that the framework is not just a theoretical curiosity but a practical tool for data-driven inference.
Applications and Results
The framework was validated through synthetic simulations involving 200 units across distinct subpopulations. These simulations demonstrated that DiscoSCMs correctly calculate individual treatment effects (ITE) in scenarios where traditional models produce degenerate results.
In industrial contexts, such as the Kuaishou video platform, this framework provides a more robust basis for "uplift modeling." By distinguishing between a user's inherent propensity to engage and the transient factors influencing a single session, platforms can optimize incentive distributions (like coupons or virtual coins) more effectively under budget constraints.
Conclusion and Future Work
DiscoSCMs offer a significant refinement to causal modeling by acknowledging the stochastic nature of individual outcomes. By replacing deterministic consistency with distributional consistency, the framework resolves the degenerative counterfactual problem and provides a clearer distinction between "luck" and "ability."
Future research will focus on integrating DiscoSCMs with deep learning architectures to learn unobservable causal representations and applying the framework to complex, multi-variable industrial datasets to further quantify ROI improvements in personalized decision-making systems.
Summary updated