Inverse-Variance Weighting From Sampling Variance

Why Weight Effect Estimates by Their Precision?

Genetic association studies may estimate the same effect in several independent cohorts. Their estimates need not be equally precise: a large cohort or one with more informative genotype and outcome data will often have a smaller sampling variance. If the goal is to combine estimates of one common effect, giving more influence to the more precise estimates can reduce the variance of the combined estimate.

This tutorial derives that weighting rule from the sampling variances. It focuses on the calculation and its assumptions, not on a full meta-analysis framework. In particular, a weighted average is not automatically an estimate of one common biological effect: that interpretation requires the cohorts to target the same estimand.

Definition: Suppose $\hat{\beta}_1,\ldots,\hat{\beta}_K$ are independent estimates of the same effect $\beta$, with $E(\hat{\beta}_i)=\beta$ and sampling variance $\operatorname{Var}(\hat{\beta}_i)=v_i$, where each $v_i$ is known or estimated. The inverse-variance weighted estimate is $$ \hat{\beta}_{\mathrm{IVW}}=\frac{\sum_{i=1}^{K} w_i\hat{\beta}_i}{\sum_{i=1}^{K}w_i}, \qquad w_i=\frac{1}{v_i}. $$ The normalized weight for estimate $i$ is $a_i=w_i/\sum_{j=1}^{K}w_j$, so the same estimate can be written as $\hat{\beta}_{\mathrm{IVW}}=\sum_{i=1}^{K}a_i\hat{\beta}_i$.

The variance $v_i$ is the square of the standard error $SE_i$: $v_i=SE_i^2$. Thus inverse-variance weighting uses $w_i=1/SE_i^2$, not $1/SE_i$. The weights describe relative precision; they are not probabilities and need not sum to one. The normalized weights $a_i$, by contrast, do sum to one.

Why the Inverse-Variance Rule Appears

Consider a linear combination $\hat{\beta}_a=\sum_{i=1}^{K}a_i\hat{\beta}_i$. To estimate the shared effect without bias, require $\sum_i a_i=1$, because then $E(\hat{\beta}_a)=\sum_i a_i\beta=\beta$. Independence gives its variance as $\sum_i a_i^2v_i$. The question is therefore: among linear unbiased combinations, which weights make this variance as small as possible?

Derivation: Minimize $\sum_i a_i^2v_i$ subject to $\sum_i a_i=1$. Introduce a multiplier $\lambda$ and differentiate the constrained objective: $$ L(a_1,\ldots,a_K,\lambda)=\sum_{i=1}^{K}a_i^2v_i-\lambda\left(\sum_{i=1}^{K}a_i-1\right), \qquad \frac{\partial L}{\partial a_i}=2a_iv_i-\lambda=0. $$ Thus $a_i=\lambda/(2v_i)$, so $a_i$ is proportional to $1/v_i$. Enforcing $\sum_i a_i=1$ gives $$ a_i=\frac{1/v_i}{\sum_{j=1}^{K}1/v_j}. $$ Substituting these weights into the variance yields $$ \operatorname{Var}(\hat{\beta}_{\mathrm{IVW}})=\sum_{i=1}^{K}a_i^2v_i =\frac{1}{\sum_{i=1}^{K}1/v_i}. $$

This derivation establishes a minimum-variance result for linear unbiased combinations when the estimates are independent and target the same effect. It does not establish that the cohorts really have a common effect, nor does it protect against bias in the cohort-specific estimates. Those questions require scientific and design-based assessment.

Standard Error and Confidence Interval

When the variances are treated as known, the standard error of the weighted estimate is the square root of its variance. With estimated variances, the same formula is commonly used as a plug-in standard error. If the cohort estimates are normal, the weighted estimate is normal; with large-sample association estimates, a normal approximation is often used. A two-sided approximate 95% confidence interval is the estimate plus or minus 1.96 standard errors.

SE(\hat{\beta}_{\mathrm{IVW}})=\sqrt{\frac{1}{\sum_{i=1}^{K}1/v_i}}, \qquad \hat{\beta}_{\mathrm{IVW}}\ \pm\ 1.96\,SE(\hat{\beta}_{\mathrm{IVW}}).

The interval has the usual repeated-sampling interpretation under the model and approximation: across repeated applications under those conditions, about 95% of intervals constructed this way would cover the common effect. It is not a statement that the particular fixed effect has a 95% probability of lying in the calculated interval.

Worked Example: Combining Two Cohort Estimates

Suppose two hypothetical, independent cohorts estimate a per-allele log odds ratio for a binary outcome. Cohort A reports $\hat{\beta}_1=0.20$ with $SE_1=0.10$; cohort B reports $\hat{\beta}_2=0.08$ with $SE_2=0.04$. Assume both estimates target the same effect, use the same allele and scale, and have approximately normal sampling distributions.

First square each standard error to obtain the sampling variances: $v_1=0.10^2=0.0100$ and $v_2=0.04^2=0.0016$. The inverse-variance weights are $w_1=1/0.0100=100$ and $w_2=1/0.0016=625$. Cohort B receives more weight because its variance is smaller.

The weighted estimate is $$ \hat{\beta}_{\mathrm{IVW}} =\frac{100(0.20)+625(0.08)}{100+625} =\frac{20+50}{725} =0.09655. $$ The variance is $1/725=0.001379$, so the standard error is $\sqrt{0.001379}=0.03714$. The approximate 95% confidence interval is $$ 0.09655\pm1.96(0.03714) =0.09655\pm0.07280 =(0.02375,\ 0.1694). $$

On the log odds ratio scale, the combined estimate is 0.09655 per allele, with an approximate interval from 0.02375 to 0.1694. Under the common-effect and sampling assumptions, the interval is compatible with a positive association on this scale. It does not by itself establish that the effect is identical across cohorts or that the association is causal.

Checking How Precision Changes the Result

The next example shows that a less precise estimate can still contribute, but its influence is reduced. It also illustrates why weights are based on variances rather than on the standard errors themselves.

Worked Example: Three Estimates With Different Precision

Consider three hypothetical independent estimates of the same per-allele log odds ratio: $\hat{\beta}_1=0.12$ with $SE_1=0.06$, $\hat{\beta}_2=-0.02$ with $SE_2=0.03$, and $\hat{\beta}_3=0.05$ with $SE_3=0.10$. Assume aligned effect scales, a shared target effect, and suitable normal approximations.

Their variances are $0.0036$, $0.0009$, and $0.0100$. Thus the inverse-variance weights are $277.78$, $1111.11$, and $100$, respectively. The total weight is $1488.89$, and the weighted numerator is $$ 277.78(0.12)+1111.11(-0.02)+100(0.05) =33.333-22.222+5.000 =16.111. $$ Therefore, $$ \hat{\beta}_{\mathrm{IVW}}=\frac{16.111}{1488.89}=0.01082. $$

The standard error is $\sqrt{1/1488.89}=0.02592$. The 95% interval is $0.01082\pm1.96(0.02592)=0.01082\pm0.05080$, or $(-0.0400,\ 0.0616)$. This interval includes zero. The most precise estimate, the second one, has the greatest influence, even though its value is negative.

Weighting is not a procedure for choosing the estimate that agrees with a preferred direction. Every estimate contributes according to its variance. Large disagreement among estimates may be scientifically important; a narrow interval calculated under a shared-effect model does not make that disagreement disappear.

Worked Example: Why Equal Weights Can Be Less Precise

Suppose two hypothetical independent cohorts estimate a per-allele effect in the same units: $\hat{\beta}_1=0.10$, $SE_1=0.05$, and $\hat{\beta}_2=0.30$, $SE_2=0.15$. Assume a common target and normal sampling. Their variances are $0.0025$ and $0.0225$, giving inverse-variance weights $400$ and $44.44$.

The weighted estimate and its standard error are $$ \hat{\beta}_{\mathrm{IVW}}=\frac{400(0.10)+44.44(0.30)}{400+44.44} =\frac{53.333}{444.44}=0.1200, \qquad SE=\sqrt{\frac{1}{444.44}}=0.04743. $$ By contrast, the equal-weight mean is $ (0.10+0.30)/2=0.20$. Because the cohort estimates are independent, its variance is $$ \operatorname{Var}\left(\frac{\hat{\beta}_1+\hat{\beta}_2}{2}\right) =\frac{0.0025+0.0225}{4}=0.00625, $$ so its standard error is $\sqrt{0.00625}=0.07906$, larger than the weighted estimate's standard error. The equal-weight mean remains unbiased under the shared-effect assumption, but it is less precise here because it gives the noisier estimate the same influence as the more precise one.

Assumptions, Diagnostics, and Alternatives

Before calculating an inverse-variance average, check the conditions that give the formula its meaning. First, the estimates must be independent. If cohorts share participants or otherwise induce correlated estimation errors, the variance of a weighted sum includes covariance terms, and the independent-estimate formula does not apply. Second, the estimates must target the same parameter on the same scale. Differences in outcome definition, effect scale, ancestry composition, or allele coding can make a single common effect an inappropriate target. The effect estimates and their standard errors must also be valid for their stated models; severe model misspecification or poorly estimated standard errors is not repaired by weighting.

Useful checks include verifying the cohort-specific sample sizes and standard errors, confirming that each estimate uses the intended effect scale and reference allele, and examining whether the cohort estimates differ substantially relative to their uncertainty. A pronounced outlying estimate, direction reversals, or wide differences in estimates can signal sampling fluctuation, data problems, or genuine variation in effects. Such signs call for investigation, not automatic deletion or an unexamined claim of one shared effect. Formal heterogeneity methods are separate tools.

The main alternatives depend on the question. An equal-weight average is simple but generally not minimum-variance when sampling variances differ. Weighting by $1/SE_i$ is also not the derived rule: the variance of a linear combination depends on squared coefficients, which leads to weights proportional to $1/v_i=1/SE_i^2$. Combining p-values addresses evidence against null hypotheses, not the magnitude of a common effect; two studies can have similar p-values but different estimated effects and precision. Inverse-variance weighting instead combines effect estimates and carries their units into the result.

Common pitfall: “The standard error is twice as large, so this cohort should get half the weight.” The correction is that its variance is four times as large, so its inverse-variance weight is one quarter as large. For example, standard errors 0.04 and 0.08 imply weights $1/0.04^2=625$ and $1/0.08^2=156.25$, not weights in a 2:1 ratio.

The method has limits. Its minimum-variance justification applies to linear unbiased estimates of one common effect under independence and correctly specified variances. It does not establish common effects, correct selection bias, or remove confounding. When standard errors are estimated, the reported uncertainty is generally a plug-in approximation. If the cohorts do not share a meaningful estimand, a precise weighted average may still have no useful common-effect interpretation.

Check Your Understanding

Answer each question, then compare with the solutions below.

Two independent estimates of one common effect have standard errors 0.05 and 0.10. What is the ratio of their inverse-variance weights, first estimate to second?
For independent estimates with variances $v_1=0.04$ and $v_2=0.01$, calculate the normalized weights.
Why does the variance of the inverse-variance weighted estimate equal the reciprocal of the total inverse-variance weight?
Give two reasons why an inverse-variance average might not represent one common genetic association effect.

Solutions. (1) The weights are $1/0.05^2=400$ and $1/0.10^2=100$, so the ratio is 4:1. (2) The unnormalized weights are 25 and 100, with total 125; the normalized weights are $25/125=0.20$ and $100/125=0.80$. (3) With normalized weights $a_i=(1/v_i)/S$, where $S=\sum_i1/v_i$, independence gives $\operatorname{Var}(\sum_i a_i\hat{\beta}_i)=\sum_i a_i^2v_i=\sum_i(1/v_i)/S^2=S/S^2=1/S$. (4) Cohorts might target different outcome or effect definitions, or their association estimates might not be independent because participants overlap. Differences in allele coding or effect scale are also reasons to check comparability before interpreting a common estimate.