[Paper Review] Distinguishing Cause from Effect Based on Exogeneity

Kun Zhang, Jiji Zhang+1 more|arXiv (Cornell University)|Apr 22, 2015

Bayesian Modeling and Causal Inference20 references13 citations

TL;DR

This paper proposes a nonparametric causal discovery method that distinguishes cause from effect by testing exogeneity of the cause for parameters in the conditional distribution of the effect. Using bootstrap-based inference on marginal and conditional distributions, the method identifies causal direction without requiring structural constraints or smoothness assumptions, and successfully detects confounders when present.

ABSTRACT

Recent developments in structural equation modeling have produced several methods that can usually distinguish cause from effect in the two-variable case. For that purpose, however, one has to impose substantial structural constraints or smoothness assumptions on the functional causal models. In this paper, we consider the problem of determining the causal direction from a related but different point of view, and propose a new framework for causal direction determination. We show that it is possible to perform causal inference based on the condition that the cause is "exogenous" for the parameters involved in the generating process from the cause to the effect. In this way, we avoid the structural constraints required by the SEM-based approaches. In particular, we exploit nonparametric methods to estimate marginal and conditional distributions, and propose a bootstrap-based approach to test for the exogeneity condition; the testing results indicate the causal direction between two variables. The proposed method is validated on both synthetic and real data.

Motivation & Objective

• To address the challenge of distinguishing cause from effect in two-variable systems without relying on strong structural assumptions common in structural equation models.
• To develop a causal inference framework grounded in the statistical concept of exogeneity, which captures the idea that the cause should be independent of the parameters governing the effect.
• To create a computationally feasible, nonparametric method for testing exogeneity that can be applied to real-world observational data.
• To enable detection of unobserved confounders by assessing whether exogeneity holds in both causal directions.
• To provide a robust alternative to existing methods that require smoothness or parametric functional form assumptions.

Proposed method

• Proposes a framework where the cause is considered exogenous for the parameters of the conditional distribution of the effect, based on the idea that the data-generating process for the cause should be independent of the parameters of the effect mechanism.
• Employs nonparametric kernel density estimation to model the marginal distribution of the cause and the conditional distribution of the effect given the cause.
• Uses a bootstrap resampling procedure to generate multiple synthetic datasets from the empirical distribution to assess the sampling distribution of the estimated parameters.
• Tests for statistical independence between the estimated parameters of the conditional model (e.g., regression coefficients or density parameters) and the parameters of the marginal model of the cause.
• Applies a bootstrap-based hypothesis test to evaluate whether the cause is exogenous for the effect’s parameters; rejection of the null implies the direction is likely causal.
• Uses the test outcome to infer causal direction: if exogeneity holds in one direction but not the other, the former is inferred as the causal direction.

Experimental results

Research questions

• RQ1 Can exogeneity of the cause for the parameters of the conditional distribution serve as a valid criterion for distinguishing cause from effect?
• RQ2 Can a nonparametric bootstrap-based test detect causal direction without assuming specific functional forms or smoothness in the data-generating process?
• RQ3 To what extent can this method detect the presence of unobserved confounders in cause-effect relationships?
• RQ4 How does the performance of this exogeneity-based method compare to existing methods like IGCI, ANM, and GPI in both synthetic and real-world data?
• RQ5 Does the method maintain robustness when the data-generating process includes nonlinearities or confounding variables?

Key findings

• The proposed bootstrap-based exogeneity test correctly identified the causal direction in 41 out of 57 real-world cause-effect pairs (72% accuracy), comparable to other leading methods.
• On synthetic data with a confounder, the method successfully detected the presence of confounding when the confounder's effect was strong, as indicated by low p-values in both causal directions.
• In the linear confounder case (α=0, b=0), the method outperformed IGCI and ANM when β was high, indicating better detection of causal direction under strong confounding.
• In the nonlinear confounder case (α=0.2, b=0.3), the method maintained strong performance and was particularly effective at identifying the correct causal direction when the confounder had a significant impact.
• The method detected potential confounders in 20 out of 77 real datasets, as indicated by p-values < 0.01 in both directions, which aligned with visual inspection of complex dependence patterns.
• The approach does not require smoothness or parametric assumptions on the functional form, making it applicable in broader settings than SEM-based methods.