Mendelian randomization (MR) acts as nature’s randomized controlled trial. By using genetic variants—specifically single nucleotide polymorphisms (SNPs)—as instrumental variables, researchers can estimate the causal effect of an exposure on an outcome without the typical confounding that plagues observational epidemiology.
However, this elegant framework relies on a fragile assumption known as the exclusion restriction. For an MR study to be valid, the chosen genetic variants must affect the outcome only through the exposure. If a genetic variant bypasses the exposure to alter the outcome directly through an independent pathway—a phenomenon known as horizontal pleiotropy—the causal inference collapses.
Because we cannot directly observe every biological pathway, we must stress-test our genetic instruments. Sensitivity analysis in MR is not just a final check; it is a multi-layered diagnostic suite designed to detect, quantify, and correct for violations of these core assumptions.
Triangulating Truth with Multiple Estimators
How do we know if a causal estimate is a genuine biological signal or an artifact of an invalid assumption? The foundational strategy of sensitivity analysis is estimator triangulation. As argued by Korf et al. in Causal clarity in statistical software, comparing estimates from multiple statistical methods that rely on fundamentally different modeling assumptions allows researchers to identify when core assumptions are violated. If different estimators yield wildly divergent results, it signals that the assumptions of at least one model have broken down.
The standard baseline for MR is the Inverse-Variance Weighted (IVW) method. IVW pools the causal estimates of individual SNPs, effectively taking a weighted average. However, IVW assumes that horizontal pleiotropy is either entirely absent or perfectly balanced (meaning the side-effects of different SNPs cancel each other out).
To challenge the IVW baseline, researchers deploy alternative estimators:
- The Weighted Median: This estimator remains valid even if up to 50% of the genetic instruments are invalid. Instead of averaging the effects, it takes the median, preventing a handful of highly pleiotropic SNPs from dragging the entire estimate astray.
- The Weighted Mode: This groups SNPs by their estimated causal effects and identifies the largest cluster of consistent variants, offering robustness even when the majority of instruments are invalid, provided the valid instruments share a common effect size.
By comparing the IVW baseline against these robust alternatives, studies like those by Yi et al. in Causal analysis between altered levels of interleukins and obstructive sleep apnea and Yujia Chen et al. in Periodontitis and the Risk of Heart Failure:a Meta-analysis and Mendelian Randomisation Study. can verify whether their causal conclusions hold under varying assumptions of instrument validity.
Quantifying Directional Pleiotropy via MR-Egger
What happens if horizontal pleiotropy is not random, but systematically pushes the outcome in a specific direction? This directional pleiotropy introduces severe bias, making an exposure look causally linked to an outcome when it is not.
MR-Egger regression provides an ingenious way to detect and adjust for this. In a standard MR analysis, we can plot each SNP's effect on the outcome against its effect on the exposure. If the exclusion restriction holds, a regression line fitted through these points should pass exactly through the origin (an intercept of zero). This makes intuitive sense: a genetic variant that has zero effect on the exposure should have zero effect on the outcome.
MR-Egger Intuition:
Causal Effect (Slope) & Directional Pleiotropy (Intercept)
Outcome Effect
^
| * (SNP 3)
| /
| * (SNP 2)
| /
+--|--* (SNP 1) <-- Intercept != 0 signals directional pleiotropy
|/
-----+--------------------> Exposure Effect
|
MR-Egger allows this intercept to float. If the intercept significantly deviates from zero, it mathematically proves the presence of directional pleiotropy. Crucially, the slope of the MR-Egger line provides a causal estimate that is automatically adjusted for this average pleiotropic bias. As demonstrated by Shen et al. in Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug‐Target Mendelian Randomization Study and Wang Hq et al. in Systemic Inflammatory Factors and Neuropsychiatric Disorders: A Bidirectional Mendelian Randomization Study, evaluating the MR-Egger intercept acts as a vital gatekeeper for causal claims.
Purging Outliers with MR-PRESSO
Sometimes, the issue is not a global trend of directional pleiotropy, but a few "bad apple" SNPs that distort the entire analysis. To find and remove these specific anomalies, researchers turn to MR-PRESSO (Mendelian Randomization Pleiotropy RESidual Sum and Outlier).
MR-PRESSO operates via a three-step mechanism:
- Detection: It calculates the residual sum of squares for the instruments to identify variants that lie unusually far from the expected regression line.
- Correction: It removes these outlier SNPs from the instrument pool and recalculates the causal estimate.
- Distortion Testing: It tests whether the causal estimate changed significantly after removing the outliers.
This iterative filtering process, utilized by Wu Zhang et al. in Complex Role of Circulating Triglycerides in Breast Cancer Onset and Survival: Insights From Two‐Sample Mendelian Randomization Study and Jun Zhang et al. in Causal inference between serum bilirubin levels and juvenile idiopathic arthritis‐associated uveitis: A bidirectional Mendelian randomization study, ensures that a handful of pleiotropic variants do not artificially drive a false-positive causal association.
Testing Heterogeneity and Leverage
Two final diagnostic pillars complete the sensitivity suite: Cochran's Q test and leave-one-out analysis.
Cochran's Q Test
Cochran's Q test measures heterogeneity—the variance among the causal estimates calculated from each individual SNP. If all SNPs are truly acting solely through the exposure pathway, their individual causal estimates should be statistically consistent, varying only due to random sampling error. High heterogeneity implies that different SNPs are operating via different biological mechanisms, a clear warning sign of horizontal pleiotropy. Authors such as Li et al. in Causal relationship between interleukin-6 levels and the risk of synovitis and tenosynovitis: A two-sample Mendelian randomization study and Jin et al. in Causality between Autism Spectrum Disorder and Telomere Length routinely use Cochran's Q to evaluate whether their instrument pool is sufficiently cohesive.
Leave-One-Out Analysis
A leave-one-out analysis systematically drops one SNP at a time and recalculates the entire causal effect using the remaining variants. If removing a single SNP causes the overall estimate to shift dramatically or lose statistical significance, the entire causal inference is fragile and overly reliant on that single variant's leverage. As shown by Jiang et al. in Association Between Unsaturated Fatty Acid Levels and Chronic Obstructive Pulmonary Disease: A Bidirectional Mendelian Randomization Study and Feifei Zhang et al. in Association of Smoking with Chronic Kidney Disease Stages 3 to 5: A Mendelian Randomization Study, passing a leave-one-out test confirms that the observed causal effect is a robust property distributed across the genetic instruments, rather than an artifact of a single influential variant.
Together, these sensitivity metrics form a rigorous, multi-layered defense system. By embedding these diagnostics directly into large-scale repositories like MRdb (Liu et al.), the field has shifted from simply reporting causal estimates to systematically stress-testing them against the hidden complexities of human genetics.
Key Foundational Work
Fasting Insulin and Risk of Overall and 14 Site-Specific Cancers: Evidence From Genetic Data (2022) — contributed 1 concept elements to this direction.
Genetic Predisposition to a Higher Whole Body Water Mass May Increase the Risk of Atrial Fibrillation: A Mendelian Randomization Study (2023) — contributed 1 concept elements to this direction.
Morning chronotype and digestive tract cancers: Mendelian randomization study (2022) — contributed 1 concept elements to this direction.