Abstract
We propose a new notion of causal contribution which describes the 'intrinsic' part of the contribution of a node on a target node in a DAG. We show that in some scenarios the existing causal quantification methods failed to capture this notion exactly. By recursively writing each node as a function of the upstream noise terms, we separate the intrinsic information added by each node from the one obtained from its ancestors. To interpret the intrinsic information as a causal contribution, we consider 'structure-preserving interventions' that randomize each node in a way that mimics the usual dependence on the parents and do not perturb the observed joint distribution. To get a measure that is invariant across arbitrary orderings of nodes we propose Shapley based symmetrization. We describe our contribution analysis for variance and entropy, but contributions for other target metrics can be defined analogously.
Introduction
In causal analysis, a fundamental challenge is determining "who is to blame" for the variation in a target variable. Standard methods, such as Pearl’s do-interventions, typically measure the effect of forcing a variable to a specific value. However, these approaches often struggle to distinguish between a node that is a primary source of variation and one that merely acts as a "wire," passing along variation from upstream ancestors.
This paper, presented at AISTATS 2024, introduces Intrinsic Causal Contribution (ICC). This metric separates a node's unique contribution to a target's uncertainty from the influence it inherits. By using "structure-preserving interventions" and Shapley values, the authors create a framework that identifies the root causes of variance or entropy in complex causal graphs without perturbing the observed joint distribution.
Methodology
The core of the framework relies on Structural Causal Models (SCMs), where every variable $X_j$ is a deterministic function of its parents and an independent "noise" term $N_j$. This noise term represents the node's intrinsic contribution—the part that cannot be explained by its ancestors.
Key components of the ICC approach include:
- Resolving Variables into Noises: The target node is expressed entirely as a function of all upstream noise terms: $X_n = F_n(N_1, \dots, N_n)$.
- Structure-Preserving Interventions: Instead of replacing a variable with a constant, the authors randomize the node’s mechanism by replacing the noise term $N_j$ with a randomized copy $N'_j$. This allows for measuring changes in target uncertainty without breaking the graph's overall structure.
- Shapley Symmetrization: To ensure the measure is invariant across arbitrary node orderings, the authors use Shapley values from cooperative game theory to fairly distribute the total uncertainty among all contributing nodes.
- Target Metrics: While generalizable, the framework focuses on quantifying contributions to Variance (for continuous data) and Shannon Entropy (for categorical data).
Key Insights: The "Wire" Property
A central concept in this research is the formalization of the "Wire" Property. The authors argue that a node acting as a perfect transmitter (a wire) should have zero intrinsic contribution. Standard causal measures often fail this intuitive test, attributing high influence to mediators simply because they are highly correlated with the target.
As shown in the figure above, ICC focuses on the mechanism of a node—how it reacts to its parents—rather than just its observed value. This is crucial for distinguishing between "inherited" and "intrinsic" effects. Furthermore, the authors demonstrate that in linear models, ICC simplifies to a standard Analysis of Variance (ANOVA) performed on noise terms, bridging modern causal inference with classical statistics.
Empirical Evaluation
The authors evaluated ICC using two real-world contexts:
- River Flows: Analyzing daily water flow records from five stations in England. ICC correctly identified that root (most upstream) nodes explained most of the variation at the target station. In contrast, standard "do-Shapley" methods incorrectly attributed high influence to the immediate upstream mediator, which was merely passing water through.
- Auto MPG: In this fuel efficiency dataset, ICC found that the number of cylinders accounted for 64% of the intrinsic contribution to fuel consumption variance. Factors like horsepower and weight, while highly correlated with MPG, were found to mostly inherit their influence from the initial design decision regarding the number of cylinders.
Limitations and Future Work
While powerful, ICC's accuracy is strictly tied to the quality of the underlying Structural Causal Model. Calculating exact Shapley values also presents computational challenges for large graphs, requiring sampling-based approximations. Future research may focus on improved estimation techniques for non-linear SCMs and applying ICC to "actual causation" to attribute blame for specific events rather than population-level variance.
Reference: Janzing, D., Blöbaum, P., et al. (2024). Quantifying intrinsic causal contributions via structure preserving interventions. Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS).
Summary updated