Fixed-Effect Meta-Analysis of Genetic Association Estimates

Why Pool Genetic Association Estimates?

A genetic association may be estimated separately in several cohorts, each with its own sample size and precision. Pooling the estimates can summarize the evidence on a shared effect scale while giving more precise estimates greater influence. The result is useful only if the cohorts’ estimates refer to the same genetic effect and the model’s assumptions are credible.

Here, a cohort-specific estimate $\hat{\beta}_i$ is an estimated genetic effect in cohort $i$, and $SE_i$ is its standard error. For example, $\beta$ could be the mean change in a quantitative phenotype per copy of an effect allele. A fixed-effect analysis targets one common underlying effect, denoted $\beta$, for all included cohorts. The term “fixed-effect” describes this common-effect model; it does not mean that sampling estimates are identical.

Definition: In a fixed-effect meta-analysis of $K$ cohorts, the cohort-specific estimates are treated as independent, approximately unbiased estimates of the same underlying effect $\beta$, with sampling variances $SE_i^2$. The inverse-variance pooled estimate is $$ \hat{\beta}_{FE}=\frac{\sum_{i=1}^{K}w_i\hat{\beta}_i}{\sum_{i=1}^{K}w_i},\qquad w_i=\frac{1}{SE_i^2}. $$ Under the model, its standard error is $$ SE(\hat{\beta}_{FE})=\sqrt{\frac{1}{\sum_{i=1}^{K}w_i}}. $$

The formula requires that estimates be on a common scale and refer to the same effect allele and genetic contrast. For instance, estimates per copy of one allele cannot be pooled as written with estimates per copy of its opposite allele unless their directions are first made consistent. The cohorts must also be independent for the stated variance formula. Overlap or other dependence means the covariance between estimates matters, so the simple formula is not valid.

The standard errors must adequately represent the sampling uncertainty of the estimates. A common working model is $\hat{\beta}_i\sim N(\beta,SE_i^2)$, independently across cohorts. This normal model may be exact in an idealized setting or an approximation justified by large-sample inference. The usual confidence interval and test also rely on this approximation.

Why the Weights and Standard Error Have These Forms

The earlier material on inverse-variance weighting establishes the principle that estimates with smaller sampling variance receive greater weight. The fixed-effect calculation follows directly from the common-effect model. Suppose the independent cohort estimates have normal sampling distributions with variances $SE_i^2$. Apart from terms not involving $\beta$, the log likelihood is

\ell(\beta)=-\frac{1}{2}\sum_{i=1}^{K}\frac{(\hat{\beta}_i-\beta)^2}{SE_i^2}+\text{constant}.

Differentiating with respect to $\beta$ and setting the derivative to zero gives the likelihood estimate:

\frac{d\ell}{d\beta} =\sum_{i=1}^{K}\frac{\hat{\beta}_i-\beta}{SE_i^2}=0 \quad\Longrightarrow\quad \hat{\beta}_{FE}= \frac{\sum_{i=1}^{K}\hat{\beta}_i/SE_i^2} {\sum_{i=1}^{K}1/SE_i^2}.

Because the estimates are independent, the variance of their weighted sum is the sum of their weighted variances. With $W=\sum_i w_i$, the pooled estimate is $\sum_i(w_i/W)\hat{\beta}_i$, so

\operatorname{Var}(\hat{\beta}_{FE}) =\sum_{i=1}^{K}\left(\frac{w_i}{W}\right)^2SE_i^2 =\sum_{i=1}^{K}\frac{w_i}{W^2} =\frac{1}{W}.

The second equality uses $w_i=1/SE_i^2$, which implies $w_i^2SE_i^2=w_i$. Taking the square root yields the stated standard error. This derivation also shows why the formula is conditional on the common-effect and independence assumptions.

Calculating and Interpreting a Pooled Estimate

Worked Example: Pooling Three Cohorts

Suppose three independent hypothetical cohorts estimate the association between an allele and a standardized quantitative phenotype. Each estimate is the change in phenotype standard deviations per additional copy of the effect allele. The results are $\hat{\beta}_1=0.08$, $SE_1=0.04$; $\hat{\beta}_2=0.12$, $SE_2=0.03$; and $\hat{\beta}_3=0.04$, $SE_3=0.05$. Assume they estimate the same effect and are on the same allele and phenotype scale.

The inverse-variance weights are $w_i=1/SE_i^2$:

w_1=\frac{1}{0.04^2}=625,\qquad w_2=\frac{1}{0.03^2}=1111.111\ldots,\qquad w_3=\frac{1}{0.05^2}=400.

Thus, the total weight is $W=625+1111.111\ldots+400=2136.111\ldots$. The weighted sum of the estimates is

625(0.08)+1111.111\ldots(0.12)+400(0.04) =50+133.333\ldots+16=199.333\ldots.

Using these unrounded weights,

\hat{\beta}_{FE}=\frac{199.333\ldots}{2136.111\ldots} \approx 0.09332,\qquad SE(\hat{\beta}_{FE})=\sqrt{\frac{1}{2136.111\ldots}} \approx 0.02164.

An approximate 95% confidence interval is $\hat{\beta}_{FE}\pm1.96SE(\hat{\beta}_{FE})$, or $0.09332\pm1.96(0.02164)$, giving approximately $(0.05091,0.1357)$ standard deviations per effect-allele copy. Under the common-effect model and the normal approximation, the interval procedure has 95% coverage over repeated samples. The data are compatible with a positive common association on this scale.

To test $H_0:\beta=0$ against a two-sided alternative, the Wald statistic is $z=\hat{\beta}_{FE}/SE(\hat{\beta}_{FE})\approx0.09332/0.02164=4.313$. Its two-sided normal-reference p-value is approximately $1.6\times10^{-5}$. This p-value is calculated under the null model; it is not the probability that the null hypothesis is true.

Precision, Confidence Intervals, and the Meaning of the Common Effect

The estimated standard error describes sampling uncertainty in the pooled estimate under the model. It is not a measure of how similar the cohort estimates are to one another. A narrow interval can result when the total inverse-variance weight is large, but it does not establish that a single common effect is a suitable description of all cohorts.

A confidence interval should be interpreted in the units of the effect estimate. If $\beta$ is a log odds ratio per effect-allele copy, the pooled estimate and interval are on the log-odds scale; exponentiating the estimate and endpoints gives the corresponding odds ratio scale. If $\beta$ is a change in a quantitative phenotype measured in standard deviations, the estimate and interval remain in standard-deviation units per allele copy.

Worked Example: A Pooled Estimate Near Zero

Consider two independent hypothetical cohorts measuring an effect in standard-deviation units per effect-allele copy. Their estimates and standard errors are $\hat{\beta}_1=0.030$, $SE_1=\sqrt{0.001}$, and $\hat{\beta}_2=0.032$, $SE_2=\sqrt{0.008}$. Assume the estimates are unbiased for a common effect and satisfy the approximate normal model.

The weights are $w_1=1/0.001=1000$ and $w_2=1/0.008=125$. Hence $W=1125$, and the weighted sum is $1000(0.030)+125(0.032)=30+4=34$. Therefore,

\hat{\beta}_{FE}=\frac{34}{1125}\approx0.03022,\qquad SE(\hat{\beta}_{FE})=\sqrt{\frac{1}{1125}}\approx0.02981.

The 95% confidence interval is

0.03022\pm1.96(0.02981)\approx(-0.0282,0.0887)

standard deviations per effect-allele copy. The corresponding test statistic for $H_0:\beta=0$ is $z=0.03022/0.02981\approx1.014$, with a two-sided p-value of about 0.31. The interval includes zero, and these data do not provide strong evidence against a zero common effect. This conclusion does not demonstrate that the effect is exactly zero.

Assumptions, Diagnostics, and Alternatives

Before interpreting the pooled number, verify that each cohort contributes an estimate of the same target. Relevant checks include the effect-allele direction, genetic contrast, phenotype definition and units, and the association model used to obtain the estimates. Confirm that each reported standard error corresponds to its estimate and that cohorts do not share participants in a way that induces dependence. These checks concern the inputs to the calculation: a correct arithmetic result cannot repair incompatible estimates.

A useful first diagnostic is to compare the cohort estimates and their standard errors directly. A large difference between estimates relative to their reported uncertainty can signal that the common-effect assumption deserves scrutiny, although visual comparison alone is not a formal heterogeneity test. Formal tests and measures of between-cohort heterogeneity are separate tools. The pooled standard error above accounts for within-cohort sampling variance under a common effect; it does not add a component for variation in underlying effects.

Fixed-effect pooling also differs from taking an unweighted average. Equal weighting gives every cohort the same influence regardless of precision; inverse-variance weighting gives greater influence to estimates with smaller standard errors. A random-effects approach instead allows underlying effects to vary across cohorts and models that variation. It answers a different question and requires additional assumptions and estimation. The fixed-effect result should not be presented as an average effect across a distribution of cohort effects when its common-effect premise is implausible.

Even if all cohort estimates point in the same direction, their magnitudes may differ. Conversely, estimates can vary in direction because of sampling error even if a common effect is assumed. The fixed-effect model does not require the observed estimates to be identical; it requires the underlying target effect to be common, with observed differences attributed to sampling error. These distinctions matter especially when results are intended to generalize beyond the cohorts contributing data.

Key takeaway: Fixed-effect meta-analysis pools independent, comparable cohort estimates by inverse-variance weighting. Its standard error reflects sampling uncertainty under a shared underlying effect; it does not measure or resolve between-cohort variation.

Common Mistakes

Rounding weights too early. Keep full precision through the weighted sum and total weight, then round the pooled estimate and standard error. Premature rounding can change the last reported digit.
Calling the pooled value a simple average. It is a weighted average, and the weights depend on the squared standard errors.
Interpreting the standard error as heterogeneity. It quantifies sampling uncertainty in the pooled estimate under the model, not the spread of true cohort effects.
Treating a non-significant test as proof of no effect. A p-value above the chosen threshold means the data do not provide sufficient evidence against the specified null under the test assumptions; it does not establish that $\beta=0$.
Pooling estimates with mismatched scales or dependent samples using the simple formula. First ensure the estimates are comparable and independent, or use a method that correctly accounts for the mismatch or dependence.

Check Your Understanding

For each question, work through the calculation or interpretation before reading its solution.

Two independent cohorts estimate the same effect: $\hat{\beta}_1=0.10$, $SE_1=0.10$, and $\hat{\beta}_2=0.20$, $SE_2=0.20$. Find the fixed-effect estimate and standard error. Solution: The weights are $100$ and $25$, so $W=125$. The pooled estimate is $[100(0.10)+25(0.20)]/125=15/125=0.12$. Its standard error is $\sqrt{1/125}\approx0.08944$, in the same units as the cohort estimates.
Why does the first cohort in question 1 receive more weight? Solution: Its standard error is half as large, so its sampling variance is one quarter as large. Since inverse-variance weight is $1/SE_i^2$, its weight is four times the second cohort’s weight.
In words, what does a 95% confidence interval for a pooled effect describe? Solution: Under the model and repeated sampling, intervals constructed by this procedure cover the common underlying effect in 95% of repetitions. It does not mean that the particular computed interval has a 95% probability of containing a fixed effect.
Suppose the cohorts used opposite effect-allele directions but their estimates were pooled without changing signs. Which assumption is at risk? Solution: The estimates do not represent the same oriented genetic effect. They must be put on a consistent allele scale before the common-effect calculation is meaningful.
If two cohorts have substantial overlap in participants, can the displayed standard-error formula automatically be used? Solution: No. Their estimates may be correlated, while the formula assumes independence. The covariance must be accounted for using an appropriate method.