מושגי ליבה
This work proposes a comprehensive framework for detecting and measuring confounding effects among variables, including separating observed and unobserved confounding, and assessing the relative strengths of confounding between different variable sets. The framework leverages data from multiple contexts where causal mechanisms of variables have shifted.
תקציר
The paper addresses the challenge of detecting and measuring confounding effects from observational data, which is crucial for causal inference tasks. Existing methods often assume causal sufficiency, disregarding the presence of unobserved confounding variables, which is an unrealistic and untestable assumption.
The authors propose three measures of confounding that utilize data from multiple contexts where causal mechanisms of variables have shifted. These measures can:
Detect and measure confounding among a set of variables.
Separate the effects of observed and unobserved confounding variables.
Understand the relative strengths of confounding bias between different sets of variables.
The measures are defined based on different properties of the causal generative process, such as directed information, mutual information, and conditional dependencies. The authors present theoretical analysis of the proposed measures, including their key properties like reflexivity, symmetry, positivity, and monotonicity.
Empirical results on synthetic datasets validate the correctness and effectiveness of the proposed framework. The authors also discuss the limitations, such as the need for a large number of contexts to evaluate the measures and the potential for different measures to yield different results for the same confounded pair of variables.