Heritable confounding in Mendelian randomization studies

This paper demonstrates that Mendelian randomization studies using genetic variants with small effect sizes are susceptible to heritable confounding that biases causal estimates in a consistent direction across methods, but shows that this bias can be mitigated through pre-estimation filtering or multivariable MR when potential confounders are identified.

The Gist

The Big Picture: Trying to Find the "Real" Cause

Imagine you are a detective trying to solve a mystery: Does high inflammation (measured by a protein called C-Reactive Protein, or CRP) actually cause Type 2 Diabetes?

In the real world, it's hard to tell cause from effect because everything is mixed up. Maybe people who eat a lot of junk food have both high inflammation and diabetes. If you just look at the data, it looks like inflammation causes diabetes, but really, the junk food is the culprit. This is called confounding.

To solve this, scientists use a method called Mendelian Randomization (MR). Think of MR as using genetic lottery tickets as a detective's tool.

• You are born with certain genetic "tickets" (variants) that naturally make your CRP levels higher or lower.
• Because these tickets are assigned at conception (like a lottery), they aren't influenced by your diet, lifestyle, or junk food.
• If people with "high CRP tickets" get diabetes more often, you can be more confident that CRP causes the diabetes, not the junk food.

The New Problem: The "Hidden Cousin"

This paper argues that while the genetic lottery ticket method is great, it has a new, sneaky flaw that is getting worse as science advances.

The Flaw: Heritable Confounding
Imagine your genetic ticket doesn't just control your CRP levels directly. Instead, imagine the ticket controls your weight (adiposity).

1. Being heavier makes your CRP go up.
2. Being heavier also makes you more likely to get diabetes.
3. So, your genetic ticket is actually pulling two strings at once: it makes you heavy, which raises CRP and causes diabetes.

In this scenario, the genetic ticket is not a pure test of CRP. It's a test of "being heavy." If you use this ticket to prove CRP causes diabetes, you are actually just proving that "being heavy causes diabetes," but you think you've proved it's the CRP. The paper calls this heritable confounding.

Why is this getting worse? (The "Deep Dive" Problem)

The paper explains that as scientists get bigger and bigger data sets (like looking at millions of people instead of thousands), they start finding weaker genetic tickets.

• The Strong Tickets: The obvious, powerful genetic tickets that directly control CRP are easy to find. These are usually "clean" and don't cause much trouble.
• The Weak Tickets: As we dig deeper, we find tiny, weak genetic signals. These are often "distal"—they are far away in the biological chain. They might control a tiny protein that controls another protein, which then controls your weight, which then controls CRP.

The Analogy:
Think of a Rube Goldberg machine.

• Strong tickets are the ball hitting the first domino. It's a direct line to the result.
• Weak tickets (found in huge studies) are the ball hitting a feather, which hits a fan, which blows a sail, which moves a boat, which hits the domino.
• The more steps in the chain, the more likely there is a "side path" (a confounder) that messes up the result.

The paper shows that the more genetic tickets we find, the more likely we are to pick the "weak, messy" ones that are actually linked to hidden confounders (like weight). This makes the results look more biased, not less.

The "Fake Confidence" Trap

The paper warns that when scientists use standard computer methods to analyze these messy tickets, the results often look very consistent and confident.

• It's like having a group of witnesses who all tell the same story, but they are all actually repeating the same lie because they are all influenced by the same hidden factor (the "heavy" gene).
• The paper shows that standard methods often fail to spot this, leading researchers to believe they have found a cause-and-effect relationship when they haven't.

The Solution: How to Fix the Detective Work

The authors suggest two simple ways to fix this, using the CRP and Diabetes example:

1. The "Multivariable" Approach (Checking the Side-Effects):
   Instead of just asking, "Does this ticket affect CRP?" ask, "Does this ticket also affect weight?"

   • If a ticket affects weight, and weight affects diabetes, we can use a statistical tool to "subtract" the weight effect. This isolates the pure CRP effect.
   • Result in the paper: When they did this for CRP, the link to diabetes disappeared. It turned out CRP wasn't the cause; weight was the real culprit.

2. The "Steiger Filter" (The Quality Control Check):
   This is a filter that throws away any genetic ticket that explains more variation in the confounder (weight) than it does in the exposure (CRP).

   • If a ticket is better at predicting weight than it is at predicting CRP, it's a "bad ticket" for this specific test. Throw it out.
   • Result in the paper: Removing these bad tickets also made the link between CRP and diabetes vanish.

The Takeaway

The paper concludes that Mendelian Randomization is not immune to bias. In fact, as we find more and more tiny genetic signals, we might be accidentally picking up more "confounded" signals.

• The Lesson: Just because a study uses genetics doesn't mean it's perfect. Researchers need to check if their genetic "tickets" are actually linked to other things (like weight, diet, or other diseases) that could be the real cause.
• The Specific Finding: In the case of C-Reactive Protein and Type 2 Diabetes, the apparent link was likely an illusion caused by adiposity (body fat). Once the researchers adjusted for this "heritable confounding," the evidence that CRP causes diabetes went away.

In short: Don't trust the genetic lottery just because it's random. You have to make sure the ticket you picked isn't actually a ticket for something else entirely.

Technical Summary

1. Problem Statement

Mendelian Randomization (MR) is a widely used epidemiological method that utilizes genetic variants as instrumental variables (IVs) to infer causal effects of exposures on outcomes. It relies on the assumption of gene-environment equivalence, specifically that genetic instruments influence the outcome solely through the exposure of interest (no horizontal pleiotropy).

The paper addresses a critical, often overlooked violation of this assumption: heritable confounding. This occurs when genetic instruments are associated with a heritable confounder that affects both the exposure and the outcome. The authors argue that as Genome-Wide Association Study (GWAS) sample sizes increase, researchers identify variants with smaller effect sizes. These smaller-effect variants are often more distal in the biological pathway (mediated through upstream traits) and are therefore more prone to being associated with confounders, leading to correlated pleiotropy. This bias can be more severe than the bias found in traditional linear regression and can mislead researchers even when multiple MR methods yield consistent results.

2. Methodology

The authors employed a multi-faceted approach combining theoretical derivation, network simulations, and an applied case study:

• Theoretical Bias Analysis:

  • They derived mathematical expressions comparing the bias in MR estimates (using a single instrument) versus linear regression estimates in the presence of a single unobserved heritable confounder.
  • They demonstrated that MR bias depends on the ratio of the pleiotropic effect (instrument $\to$ outcome via confounder) to the instrument's effect on the exposure.

• Network Simulations:

  • Using the R/dagitty package, they simulated random Directed Acyclic Graphs (DAGs) representing genotype-phenotype networks with varying numbers of traits (N=75 to 150) and edge densities.
  • They modeled the discovery of genetic instruments as a function of GWAS sample size (simulated by varying p-value thresholds).
  • Hypothesis: Variants with longer paths to the exposure (more distal) have smaller effect sizes and are more likely to act via confounders.

• Bias Simulation of MR Estimators:

  • They simulated data where an exposure and outcome are confounded by a heritable trait, with no true causal effect between them.
  • They tested standard univariable MR methods (Inverse Variance Weighted [IVW], MR-Egger, Weighted Median, Weighted Mode) against two proposed mitigation strategies:
    1. Multivariable MR (MVMR): Including the confounder as an additional exposure.
    2. Steiger Filtering: Removing SNPs that explain more variance in the confounder than in the exposure.
  • They also tested advanced methods designed for correlated pleiotropy: CAUSE and MR-cML.

• Applied Case Study (CRP $\to$ Type 2 Diabetes):

  • Exposure: C-Reactive Protein (CRP).
  • Outcome: Type 2 Diabetes (T2D).
  • Process:
    1. Identified 728 lead variants for CRP from a large GWAS.
    2. Performed a Phenome-Wide Association Study (PheWAS) using the GWAS Catalog to identify traits associated with these CRP instruments.
    3. Filtered for traits that also associated with T2D, identifying 235 potential heritable confounders.
    4. Applied Steiger filtering and MVMR to adjust for the top confounders (specifically adiposity-related traits and protein biomarkers).
    5. Compared results with CAUSE and MR-cML.

3. Key Contributions

• Identification of a Specific Bias Mechanism: The paper clarifies that "heritable confounding" is a distinct source of bias where genetic instruments are linked to upstream confounders, generating correlated pleiotropy.
• Sample Size Paradox: It highlights a counter-intuitive finding: larger GWAS sample sizes increase the risk of heritable confounding bias. As sample sizes grow, weaker variants (which are often more distal and pleiotropic) are discovered and included as instruments, violating IV assumptions more frequently than strong, proximal variants.
• Bias Magnitude: Theoretical and simulation results show that MR estimates biased by heritable confounding can be more biased than the corresponding observational linear regression estimates, potentially leading to false confidence in causal claims.
• Limitations of Current Robust Methods: The study demonstrates that standard "pleiotropy-robust" methods (MR-Egger, Weighted Median) and advanced Bayesian methods (CAUSE, MR-cML) fail to correct for bias when the majority of instruments are associated with a specific heritable confounder (correlated pleiotropy).
• Proposed Solutions: The authors advocate for the systematic use of Steiger filtering and Multivariable MR (MVMR) using known or suspected confounders as primary robustness checks.

4. Key Results

• Simulation Findings:

  • In scenarios where a high proportion of SNPs act via a confounder, all univariable MR methods (IVW, MR-Egger, etc.) produced biased estimates consistent with the confounded observational association.
  • MVMR and Steiger filtering successfully recovered the null causal effect (the true value in the simulation) when the confounder was known and measurable.
  • Methods like CAUSE and MR-cML failed to correct the bias in the presence of strong correlated pleiotropy, yielding estimates similar to the biased IVW results.

• Case Study (CRP $\to$ T2D) Findings:

  • Initial univariable MR suggested a causal effect of CRP on T2D.
  • PheWAS identified adiposity-related traits and protein biomarkers (linked to the APOE locus, rs429358) as major confounders.
  • Steiger filtering (removing variants associated more strongly with adiposity) and MVMR (adjusting for adiposity) both led to a substantial attenuation of the effect estimate, rendering the causal link between CRP and T2D non-significant (compatible with a null effect).
  • This suggests the original association was driven by heritable confounding (likely body fat influencing both CRP and T2D) rather than a direct causal effect of inflammation.

5. Significance and Implications

• Re-evaluation of MR Literature: The findings suggest that many existing MR studies, particularly those using large biobank data and weaker instruments, may have overestimated causal effects due to unaccounted heritable confounding.
• Methodological Shift: The paper argues that MR is not immune to confounding bias and requires active adjustment strategies similar to (but distinct from) linear regression.
• Gene-Environment Equivalence (GEE): The authors challenge the plausibility of GEE for many complex traits, noting that if genetic variants influence an exposure via upstream phenotypes (e.g., adiposity affecting CRP), the genetic manipulation does not mimic a specific intervention on the exposure alone.
• Practical Recommendations:
  • Researchers should systematically screen instruments for associations with known confounders (via PheWAS).
  • Steiger filtering and MVMR should be standard robustness checks.
  • Consistency across different MR methods does not guarantee validity if all methods share the same underlying bias from a dominant confounder.
  • Results should be interpreted within a "triangulation of evidence" framework, acknowledging that MR is susceptible to specific biases that differ from observational studies.

In conclusion, this paper provides a rigorous warning that the "gold standard" of MR is vulnerable to heritable confounding, particularly in the era of massive GWAS datasets, and offers practical, accessible tools to mitigate this threat.