Weak Instrument Bias

The primary diagnostic for weak instruments is the first-stage F-statistic, which tests the joint strength of the genetic variant-exposure associations. A value below 10 signals dangerously weak instruments. This threshold originates from the Staiger-Stock rule: when F < 10, the bias of two-stage least squares (2SLS) can exceed 10% of the ordinary least squares bias. In Mendelian randomization, the mean F-statistic across all variants is calculated as F = (R² × (N - 1 - k)) / (k × (1 - R²)), where R² is the proportion of exposure variance explained, N is sample size, and k is the number of instruments. Even with F > 10, bias persists—the rule merely caps it at an acceptable level.

In two-sample Mendelian randomization, where variant-exposure and variant-outcome associations come from independent datasets, weak instrument bias pulls the causal estimate toward the confounded observational association. This occurs because the first-stage regression coefficients are estimated with error, and the second stage inherits this noise. The bias is proportional to 1 / F and is exacerbated when instruments are many but individually weak. Critically, this means weak instruments can make a null causal effect appear significant if the observational association is non-null, producing false positive causal claims that are especially dangerous in drug target validation contexts.

In one-sample Mendelian randomization, where variant-exposure and variant-outcome associations come from the same dataset, weak instrument bias pulls the causal estimate toward the null. This is a finite-sample bias arising from the correlation between the first-stage estimation error and the structural error term. The 2SLS estimator is biased in the direction of the ordinary least squares estimator, but since one-sample settings often involve overfitting—the same data used to select instruments also estimates their effects—the net result is attenuation toward zero. This can mask genuine causal effects, leading to false negative conclusions in clinical applications.

Weak instrument bias is amplified by winner's curse, a phenomenon where genetic variant-exposure associations discovered in genome-wide association studies (GWAS) are systematically overestimated in the discovery sample. When these inflated effect sizes are used as instruments, the true first-stage strength is lower than estimated, making instruments even weaker than they appear. This is especially problematic when instruments are selected from the same GWAS used for MR analysis. Solutions include:

• Using independent replication samples for instrument selection
• Applying shrinkage corrections like the James-Stein estimator
• Employing split-sample designs where discovery and estimation are separated

Several statistical methods are specifically designed to be robust to weak instrument bias when the F-statistic is low:

• Limited Information Maximum Likelihood (LIML): A single-equation estimator that is median-unbiased even with weak instruments, unlike 2SLS which is mean-biased
• Anderson-Rubin Test: A test of the causal null hypothesis that is exact and robust to arbitrarily weak instruments, though it loses power with many instruments
• Conditional Likelihood Ratio (CLR) Test: Combines the power advantages of LIML with the robustness of Anderson-Rubin, recommended when instruments are weak
• Bayesian Methods: Priors on the concentration parameter can shrink estimates appropriately when instruments are weak, producing better-calibrated credible intervals

Weak instrument bias has direct implications for study design in Mendelian randomization:

• Sample size planning: The required sample size to achieve F > 10 depends on the variance explained by instruments. For polygenic exposures where each variant explains <0.1% of variance, tens of thousands of participants may be insufficient
• Instrument selection strategy: Using genome-wide significant variants (p < 5×10⁻⁸) is standard, but liberal thresholds (p < 5×10⁻⁶) increase instrument count at the cost of weaker individual strength
• Reporting standards: The STROBE-MR guidelines require reporting the mean F-statistic and the proportion of variance explained (R²) for all instruments
• Sensitivity analyses: Always report results from multiple robust methods (LIML, MR-Egger, weighted median) to assess sensitivity to weak instrument assumptions

A composite instrument constructed by aggregating multiple genetic variants into a single weighted score. Each individual's allele score equals the sum of risk-increasing alleles weighted by their estimated effect sizes on the exposure. Using an allele score rather than individual SNPs can mitigate weak instrument bias by:

• Increasing the proportion of exposure variance explained
• Reducing the number of instruments, limiting overfitting
• Providing a single, stronger instrument for two-stage estimation However, unweighted or externally weighted scores may still suffer from bias if the weights are imprecisely estimated.

A phenomenon where genetic variant-exposure associations estimated in discovery GWAS are systematically overestimated due to selection on statistical significance. This inflates the apparent instrument strength and introduces bias into two-sample Mendelian randomization:

• Discovery sample: Variants selected for surpassing the genome-wide significance threshold (p < 5×10⁻⁸)
• Replication sample: Effect sizes regress toward the mean When instruments are selected from the same dataset used to estimate exposure associations, winner's curse creates a positive feedback loop that exacerbates weak instrument bias. Using independent samples for instrument selection and effect estimation mitigates this distortion.