NOME Violations and SIMEX Correction in MR-Egger

Why Exposure-Association Uncertainty Matters

MR-Egger regression relates each variant’s estimated outcome association to its estimated exposure association. Its slope is used to estimate a causal effect, while its intercept represents average directional pleiotropy under the model’s assumptions. But the horizontal coordinate in this regression, the estimated SNP–exposure association, is itself measured with uncertainty. If that uncertainty is appreciable, the fitted slope can be pulled toward zero: a form of regression dilution.

The MR-Egger NOME assumption means “NO Measurement Error”: for practical purposes, uncertainty in the SNP–exposure estimates is negligible relative to their spread across variants. NOME is not the same as the InSIDE assumption. InSIDE concerns the relationship between instrument strength and direct effects; it does not guarantee that exposure associations are estimated precisely. \(I^2_{GX}\) helps diagnose the extent of potential dilution, and SIMEX provides a way to examine and correct for it under additional modeling assumptions.

The usual formulation also relies on appropriate instrument selection and harmonization, and commonly assumes independent variants and independent exposure- and outcome-association estimates. When variants are correlated or association estimates are dependent, covariance-aware methods are needed; the simple calculations below do not account for those dependencies. As in other MR analyses, satisfying NOME alone does not establish that the causal assumptions hold.

How Measurement Error Dilutes a Slope

Suppose, as a simple model, that an observed predictor \(X^*\) equals a true predictor \(X\) plus independent, mean-zero measurement error \(U\): \(X^*=X+U\). Suppose the outcome follows \(Y=\alpha+\theta X+\varepsilon\), with the outcome error independent of \(X\) and \(U\). The ordinary regression slope of \(Y\) on \(X^*\) is attenuated because the observed predictor’s variation includes noise that does not explain \(Y\).

The factor in parentheses is the reliability ratio, between zero and one when both variances are nonnegative and the true-predictor variance is positive. Thus, in this simple model, the expected fitted slope is closer to zero than \(\theta\). MR-Egger is a weighted regression across variant associations rather than this single-predictor model, so the equation is an intuition for the mechanism, not a complete derivation of all MR-Egger bias. In particular, the weighted regression and the pattern of exposure-association errors affect the amount of dilution.

Measuring Potential Dilution With \(I^2_{GX}\)

The \(I^2_{GX}\) statistic compares the observed dispersion of the SNP–exposure estimates with the dispersion expected from their sampling uncertainty. Here \(K\) is the number of variants, and the calculation uses their exposure-association standard errors. A value near one indicates that most of the observed spread is beyond sampling error; a low value indicates that sampling error may account for a substantial part of the spread.

The subtraction of \(K-1\) reflects the expected degrees of freedom for dispersion around an estimated weighted mean under a common-effect sampling model. The maximum with zero prevents a negative value when \(Q_{GX}\) is no greater than its degrees of freedom. \(I^2_{GX}\) is a diagnostic, not a test that establishes NOME, and it does not indicate whether InSIDE or the other MR assumptions are satisfied.

SIMEX: Add Error, Then Extrapolate It Away

SIMEX (simulation extrapolation) examines how an estimator changes as additional measurement error is introduced. For MR-Egger, the added error is applied to the estimated exposure associations, using their standard errors to set its scale. At each added-error level, the MR-Egger regression is refitted. The resulting average estimates are then modeled as a function of the added error and extrapolated to the level that would remove the original measurement error.

The extrapolation target is \(-1\) because the observed exposure-association estimates already contain their original sampling-error variance, corresponding to \(\lambda=0\). Adding error at level \(\lambda\) gives total variance proportional to \((1+\lambda)SE(\hat{\gamma}_j)^2\). Setting \(1+\lambda=0\) yields \(\lambda=-1\), the formal zero-measurement-error target. This target lies outside the simulated range, so the correction depends on extrapolation rather than direct observation.

Interpreting the Diagnostic and the Correction

\(I^2_{GX}\) and SIMEX address related but distinct questions. The statistic summarizes how much exposure-association spread appears to exceed sampling error. SIMEX uses the stated standard errors to simulate additional error and assess how the fitted MR-Egger coefficients respond. A low \(I^2_{GX}\) is a reason to examine dilution carefully; SIMEX is one sensitivity analysis for that concern. Neither replaces evaluation of InSIDE, instrument validity, population structure, sample overlap, or other sources of bias.

SIMEX results can be sensitive to the chosen extrapolation function and to simulation variability. Because the target \(\lambda=-1\) is outside the added-error levels, a curve that fits the simulated range well may still extrapolate poorly. A full analysis should report the uncorrected and corrected estimates, the simulation and curve-fitting choices, and uncertainty calculated with a method that accounts for the SIMEX procedure. Simply attaching the ordinary MR-Egger standard error to the corrected point estimate does not generally account for that uncertainty.

Common Pitfall

Wrong reasoning: “\(I^2_{GX}=0.90\), so NOME is proven, and the MR-Egger estimate is unbiased.” Correction: \(I^2_{GX}\) is a diagnostic for the relative contribution of exposure-association sampling error to observed dispersion. A high value suggests less dilution from this source; it does not prove NOME, InSIDE, or any other causal assumption. Similarly, a SIMEX-corrected slope is conditional on its error and extrapolation models. It is a sensitivity analysis, not a guarantee that all bias has been removed.