Inverse-Variance Weighted Estimation

Why Weight Variant-Specific Ratios?

A genetic variant can provide a separate estimate of the causal effect of an exposure on an outcome: its variant-outcome association divided by its variant-exposure association. The Summary-Data Wald Ratio gives this estimate for one variant. With several valid, independent variants, the estimates can be combined. They should not generally receive equal influence: a ratio estimated with less uncertainty contains more information about the common causal effect.

Inverse-variance weighted (IVW) estimation makes this principle explicit. It averages the variant-specific ratios, assigning each a weight equal to the reciprocal of its variance. The result is most straightforward when all variants estimate the same causal effect and their ratio estimates are statistically independent. Those are substantive and statistical assumptions, not consequences of applying a formula.

The units of each ratio are outcome units per exposure unit. For example, if the exposure association is measured in standard-deviation units and the outcome association in millimetres of mercury, the ratio and the combined estimate are in millimetres of mercury per exposure standard deviation. The weights have inverse squared units, so the weighted average retains the units of the ratios.

Model and Causal Assumptions

To interpret the combined value as a causal estimate, each selected variant must be a valid instrument for the exposure. In standard terms, a valid instrument is associated with the exposure, is independent of causes shared by exposure and outcome, and affects the outcome only through the exposure. In addition, the simple common-effect IVW model assumes that each valid variant identifies the same causal effect, denoted by \(\theta\). Under that model, the population associations satisfy \(\Gamma_j=\theta\gamma_j\), where \(\gamma_j\) and \(\Gamma_j\) are the corresponding true variant-exposure and variant-outcome associations.

For the usual independent-variant calculation, the variants must also be independent of one another in the relevant data, or sufficiently uncorrelated that their association estimates can be treated as independent. The ratio estimates must be approximately normally distributed around \(\theta\), with correctly estimated standard errors. In practice, this approximation is most defensible when the exposure-association estimates in the denominators are sufficiently precise. A weak or very imprecise denominator can make a ratio unstable and make a normal approximation unreliable.

A common summary-data implementation treats the exposure-association estimates as precise and uses the outcome-association standard errors to approximate the ratio variances. Specifically, if \(SE(\hat{\Gamma}_j)\) is the standard error of the variant-outcome association, the first-order approximation is \(s_j \approx SE(\hat{\Gamma}_j)/|\hat{\gamma}_j|\). This gives \(w_j \approx \hat{\gamma}_j^2/SE(\hat{\Gamma}_j)^2\). When a ratio standard error that accounts for uncertainty in both associations is available and appropriate, use that ratio standard error to define the weight instead. These formulations should not be mixed without checking which uncertainty each weight represents.

Deriving the Estimate and Its Uncertainty

Suppose the observed ratio estimates are independent and satisfy the approximate model \(\hat{R}_j\sim N(\theta,s_j^2)\), with known or estimated standard errors \(s_j\). The likelihood for \(\theta\), apart from factors not involving \(\theta\), is proportional to the exponential of minus one half the weighted sum of squared deviations. Maximizing this likelihood is equivalent to minimizing that sum.

Differentiate with respect to \(\theta\) and set the derivative equal to zero. Since each \(w_j\) is fixed with respect to \(\theta\), the resulting equation is \(\sum_j w_j(\hat{R}_j-\theta)=0\). Solving gives the weighted average in the definition. Under the model, its variance follows from independence:

Thus the estimated standard error is \(SE(\hat{\theta}_{\mathrm{IVW}})=1/\sqrt{\sum_j w_j}\). When the normal approximation is reasonable, a 95% confidence interval is \(\hat{\theta}_{\mathrm{IVW}}\pm1.96SE(\hat{\theta}_{\mathrm{IVW}})\). The interval describes uncertainty in the estimated common causal effect under the model and instrument assumptions; it does not give the probability that the fixed causal effect lies within this particular interval.

Using Variant Association Summary Data

The weighting scheme can be expressed directly in terms of the exposure and outcome associations. Under the first-order approximation \(s_j\approx SE(\hat{\Gamma}_j)/|\hat{\gamma}_j|\), the ratio weight is approximately \(\hat{\gamma}_j^2/SE(\hat{\Gamma}_j)^2\). Substituting the ratio \(\hat{\Gamma}_j/\hat{\gamma}_j\) into the weighted average gives the familiar summary-data formula:

This expression is a weighted regression of the estimated outcome associations on the estimated exposure associations, constrained to pass through zero, using inverse outcome-association variances as regression weights. Its equivalence to the weighted ratio average depends on the stated first-order approximation. It does not mean exposure-association uncertainty is always negligible; if that uncertainty is material, the approximation and resulting uncertainty calculation may be inadequate.

What the Weights Do—and Do Not—Mean

A weight measures the relative precision of a ratio under the model. It is not a measure of how biologically important a variant is, and a large weight does not establish that the variant is valid. When weights sum to \(W\), variant \(j\)'s share of the estimate is \(w_j/W\). Consequently, a precise ratio can dominate the combined estimate, while a highly uncertain ratio contributes relatively little.

The causal interpretation still requires valid instruments and a common effect. If a variant affects the outcome through a pathway other than the exposure, is related to shared causes of exposure and outcome, or is correlated with other included variants, the simple independent common-effect calculation may not represent the intended causal effect. Those conditions cannot be verified merely by observing a small standard error or calculating a weighted average.

Common Mistakes

• Giving every ratio equal weight. An ordinary average ignores differences in precision. IVW uses \(1/s_j^2\), so estimates with smaller standard errors receive more weight.
• Using standard errors rather than variances as weights. The weight is the reciprocal of the variance, \(1/s_j^2\), not \(1/s_j\).
• Interpreting the weighted result as automatically causal. The formula combines estimates; causal meaning depends on valid instruments and the common-effect model.
• Ignoring how ratio standard errors were obtained. The shortcut based on outcome-association standard errors treats exposure associations as sufficiently precise. It should not be used without considering denominator uncertainty.
• Reading a confidence interval as a probability statement about the fixed effect. A 95% confidence procedure has 95% long-run coverage under its assumptions; it does not assign a 95% probability to the parameter being in the observed interval.