[Paper Review] Leveraging directed causal discovery to detect latent common causes

Ciarán M. Lee, Christopher D. Hart+2 more|arXiv (Cornell University)|Oct 22, 2019

Bayesian Modeling and Causal Inference27 references4 citations

TL;DR

This paper proposes a heuristic method to extend purely directed causal discovery algorithms—IGCI and KCDC—to detect latent common causes by analyzing the variance and mean of a novel test statistic Δ. The modified algorithms successfully identify common causes in synthetic and real-world data while preserving high accuracy in distinguishing directed causal relationships.

ABSTRACT

The discovery of causal relationships is a fundamental problem in science and medicine. In recent years, many elegant approaches to discovering causal relationships between two variables from observational data have been proposed. However, most of these deal only with purely directed causal relationships and cannot detect latent common causes. Here, we devise a general heuristic which takes a causal discovery algorithm that can only distinguish purely directed causal relations and modifies it to also detect latent common causes. We apply our method to two directed causal discovery algorithms, the Information Geometric Causal Inference of (Daniusis et al., 2010) and the Kernel Conditional Deviance for Causal Inference of (Mitrovic, Sejdinovic, & Teh, 2018), and extensively test on synthetic data -- detecting latent common causes in additive, multiplicative and complex noise regimes -- and on real data, where we are able to detect known common causes. In addition to detecting latent common causes, our experiments demonstrate that both the modified algorithms preserve the performance of the original in distinguishing directed causal relations.

Motivation & Objective

• To address the limitation of existing causal discovery algorithms that cannot detect latent common causes, which are common in real-world data such as medical records.
• To develop a general heuristic that enhances purely directed causal discovery methods to also detect latent confounders without requiring strong assumptions like additive noise or linearity.
• To preserve the original algorithms' performance in distinguishing directed causal relationships while adding the capability to detect common causes.
• To validate the method on diverse synthetic data regimes (additive, multiplicative, complex noise) and real-world datasets with known common causes.

Proposed method

• The method applies a heuristic thresholding strategy based on the mean and variance of a test statistic Δ, derived from the output of a base causal discovery algorithm.
• For each cause-effect pair, the algorithm computes Δ values across multiple runs to estimate the mean and variance of the distribution of causal direction scores.
• A high mean and low variance in Δ indicate a directed causal relationship; a low mean and high variance suggest a latent common cause.
• The approach is applied to two state-of-the-art algorithms: Information Geometric Causal Inference (IGCI) and Kernel Conditional Deviance for Causal Inference (KCDC).
• Thresholds for distinguishing directed from common cause structures are set based on domain knowledge and empirical analysis, following principles from prior work such as CAN and KCDC.
• The method does not require retraining or structural changes to the base algorithms, only post-processing of their outputs using the Δ statistic.

Experimental results

Research questions

• RQ1 Can a purely directed causal discovery algorithm be enhanced to detect latent common causes without altering its core structure?
• RQ2 Does the proposed heuristic maintain the original algorithm’s accuracy in distinguishing directed causal relationships when detecting common causes?
• RQ3 How well does the method perform across different noise regimes—additive, multiplicative, and complex—on synthetic data?
• RQ4 Can the method correctly identify known common causes in real-world datasets such as Breast Cancer Wisconsin and AutoMPG?
• RQ5 How do the mean and variance of the Δ statistic reliably distinguish between directed and common cause structures?

Key findings

• On the NLSchools dataset, modKCDC returned a mean Δ of 0.933 and variance of 0.007, indicating a directed causal relationship from socioeconomic status to language score, consistent with ground truth.
• On the Breast Cancer Wisconsin dataset, both modKCDC and modIGCI returned low mean Δ (0.491 and 0.308, respectively) with high variance (0.081 and 0.062), correctly identifying a common cause between perimeter and compactness.
• On the AutoMPG dataset, modKCDC returned mean Δ = 0.639 and variance = 0.086, indicating a common cause between acceleration and MPG, while modIGCI showed low mean (0.124) and very low variance (0.00062), also consistent with a common cause.
• The modified algorithms preserved high performance in distinguishing directed causal relationships, with modKCDC achieving 73% accuracy and modIGCI 74% on the Tübingen cause-effect pair dataset, matching the original algorithms.
• The heuristic thresholds for Δ were effective across diverse data types and noise regimes, demonstrating robustness without requiring extensive hyperparameter tuning.
• The method successfully identified latent common causes in real-world data where the underlying causal structure is known, validating its practical utility in medical and scientific applications.