Orienting the causal relationship between imprecisely measured traits using GWAS summary data

GWAS
causal direction
causal inference
mendelian randomization
statistical methods
  • Core Problem: This research addresses how to determine the causal direction between two highly correlated traits (\(X \rightarrow Y\) vs. \(Y \rightarrow X\)) using easily accessible GWAS summary data.
  • Key Method (Steiger Filtering): The primary method is Steiger filtering, which tests if the genetic instrument explains more variance in the intended exposure than in the intended outcome. If the reverse is true, it suggests an incorrect causal direction or unmodeled pleiotropy.
  • Impact: The method has become a routine, mandatory step in Mendelian Randomization studies to validate that the genetic variants selected are genuine instruments for the exposure and not proxies for the outcome or unmodeled confounders, thus preventing reverse causation bias.

assumptions/limitations/bias

Published

23 January 2026

PubMed: 29149188 DOI: 10.1371/journal.pgen.1007081 Overview generated by: Gemini 2.5 Flash, 28/11/2025

Key Findings: Causal Direction between Traits

This paper addresses the common challenge in Mendelian Randomization (MR) and other observational studies: determining the direction of the causal relationship between two highly correlated traits (\(X\) and \(Y\)). The authors propose a novel method using GWAS summary data to test for the existence and direction of a causal effect, even when the traits are measured with error. The key finding is that under certain conditions, the causal direction can be inferred by comparing the strength of the genetic associations with the two traits.

Methods: MR Directionality Methods

The study proposes two primary methods for causal directionality testing, both rooted in the structure of MR and the concept of pleiotropy (where one gene affects multiple traits):

  1. Comparison of Variance Explained (\(R^2\)): This method suggests that if trait \(X\) is the cause of trait \(Y\), the proportion of variance explained in \(X\) by the genetic instruments (\(R^2_X\)) should be greater than the proportion of variance explained in \(Y\) by those same instruments (\(R^2_Y\)). This comparison allows for the testing of a directional hypothesis.
  2. Steiger Filtering: This method, which has become a foundational tool in MR, formally tests the assumption that the genetic instrument influences the exposure (\(X\)) more strongly than it influences the outcome (\(Y\)). This test is used to filter out genetic variants that appear to be stronger instruments for the outcome than for the exposure, suggesting a potential reverse causation (i.e., \(Y\) causes \(X\)) or unmodeled pleiotropy. Steiger filtering is particularly robust because it uses GWAS summary data and accounts for the fact that many traits are measured with error (imprecise measurement).

Results: Application and Robustness

The methods were tested using simulated data and applied to real-world traits with known causal direction.

  • Simulation Performance: Simulations demonstrated that the Steiger filtering approach was effective in correctly orienting the causal relationship under various scenarios, including the presence of measurement error in both traits.
  • Real-World Application: When applied to the causal relationship between height and educational attainment—where the causal direction is known to be primarily from height to educational attainment due to confounding factors—the methods successfully identified the correct direction.

Conclusions and Recommendations

The paper provides robust statistical methods for determining the causal direction between two traits using readily available GWAS summary statistics.

  • Utility in MR: The methods are critical for two-sample MR, where the causal direction is often assumed but may be incorrect. Steiger filtering has become a standard check to validate that the genetic instrument primarily targets the intended exposure.
  • Imprecise Measurement: A major strength is the ability to account for the impact of imprecise or confounded measurement of traits, a pervasive issue in observational epidemiology.
  • Causal Inference: The approach adds rigor to causal inference by allowing researchers to test, rather than simply assume, the direction of the relationship, greatly reducing the risk of making reverse causal inferences.