Why Model Ancestry Effects as Random?
An association estimate can differ across ancestry groups because the underlying association effects differ, because estimates are noisy, or both. Fixed-effect trans-ancestry meta-analysis treats the underlying effect as common across groups. A random-effects model instead allows each ancestry-specific effect to vary around a mean. This is useful when the aim includes describing the distribution of effects or predicting the effect in a new population, rather than only estimating one shared effect.
The model does not make ancestry groups interchangeable, nor does it explain why effects differ. It summarizes effect variation on a specified scale. It also does not correct biased ancestry-specific estimates: the association analyses producing those estimates must still be appropriately designed and adjusted.
The normal distribution for \(\beta_j\) is a working model for variation across the population of settings to which inference is intended to apply. In a particular study, ancestry groups are not necessarily a random sample from a well-defined biological superpopulation. Therefore, interpreting \(\mu\) and predicting for a new population require a defensible target population and a reasonable exchangeability assumption: the new population's effect should be generated by the same distribution as the modeled effects.
The effects \(y_j\) must estimate a comparable quantity on a common scale. For example, estimates on the same log-odds scale can be combined if the association models and effect definitions are comparable. Alleles and effect directions must also be aligned. Substantial overlap between cohorts violates the independence assumption unless the covariance between estimates is incorporated; ignoring positive covariance can make the combined estimate look more precise than it is.
Estimation of the Mean and Heterogeneity
If \(\tau^2\) were known, each estimate would receive a weight inversely proportional to its total variance, \(s_j^2+\tau^2\). The estimated mean is the weighted average of the \(y_j\), with a standard error based on the sum of these weights. In practice, \(\tau^2\) must be estimated from the observed differences among the estimates.
The truncation at zero reflects that a variance cannot be negative. When \(Q\leq k-1\), this method estimates \(\tau^2\) as zero; it does not prove that all true effects are identical. The DerSimonian–Laird estimator is a method-of-moments estimator, not a guarantee that the variance has been estimated precisely. Restricted maximum likelihood (REML) is a common alternative, particularly when the number of groups is small or the standard errors differ substantially. Different estimators can give meaningfully different results with few groups.
Why the Variance Estimate Uses \(Q\)
The rationale for the DerSimonian–Laird estimate can be seen from the expected value of \(Q\). Write \(w_j=1/s_j^2\) and \(W=\sum_jw_j\). Under the model, the \(y_j\) have variance \(s_j^2+\tau^2\). The weighted residual sum of squares around the weighted mean has expectation $$ E(Q)=(k-1)+\tau^2\left(W-\frac{\sum_j w_j^2}{W}\right) =(k-1)+\tau^2C. $$ The \(k-1\) term is the expected contribution from sampling variation when the true effects are identical. The extra term is the expected contribution from between-group variation. Solving this expectation equation for \(\tau^2\) gives \((Q-(k-1))/C\); taking the maximum with zero enforces the variance boundary. This is a moment-based justification, not an exact claim that \(Q\) follows a simple distribution under every design.
Worked Example: Fitting a Random-Effects Model
Suppose four hypothetical ancestry-specific association estimates, all on the same effect scale, are \(0.10,\ 0.30,\ 0.50,\ 0.70\), each with standard error \(0.20\). Assume independent estimates and comparable association targets. The common standard error gives \(w_j=1/0.20^2=25\), and the fixed-effect weighted mean is $$ \hat{\mu}_{F}=\frac{25(0.10+0.30+0.50+0.70)}{100}=0.40. $$
The squared deviations from \(0.40\) sum to \(0.09+0.01+0.01+0.09=0.20\), so $$ Q=25(0.20)=5.00,\qquad df=k-1=3. $$ Also, \(C=100-(4\times25^2)/100=75\). Thus $$ \hat{\tau}^2_{\mathrm{DL}}=\frac{5.00-3}{75}=0.02667. $$ There is no need to truncate this positive estimate. The estimated between-group standard deviation is \(\hat{\tau}=\sqrt{0.02667}=0.1633\), in the same units as the effects.
Each random-effects weight is \(1/(0.20^2+0.02667)=15.00\). The weighted mean remains \(0.40\), because the standard errors are equal, and $$ SE(\hat{\mu})=\sqrt{\frac{1}{4(15.00)}}=0.1291. $$ The conventional normal-approximation 95% confidence interval for the mean effect is \(0.40\pm1.96(0.1291)\), or \([0.147,\ 0.653]\). Under this model and approximation, it is an interval for \(\mu\), the mean of the effect distribution—not an interval intended to contain 95% of individual ancestry-specific effects.
Prediction for a New Population
A confidence interval for \(\mu\) addresses uncertainty in the average effect. A prediction interval addresses where the true effect \(\beta_{\mathrm{new}}\) in a new, exchangeable population might lie. The latter includes both uncertainty in estimating the mean and actual variation of effects around that mean. A commonly used approximate interval is
Here \(t_{k-2,\,0.975}\) is the 97.5th percentile of a Student \(t\) distribution with \(k-2\) degrees of freedom. This is a plug-in approximation: it uses an estimated \(\tau^2\), and it does not fully represent uncertainty in that estimate. With very few groups, the \(t\) multiplier can be large and the interval can be unstable. Prediction intervals based on REML or other approaches, including appropriate bootstrap methods, may be preferable in some analyses.
Worked Example: Interpreting a Prediction Interval
Continue with the four hypothetical groups above. The prediction standard error is $$ \sqrt{0.02667+0.1291^2} =\sqrt{0.02667+0.01667} =\sqrt{0.04334}=0.2082. $$ With \(k-2=2\) degrees of freedom, \(t_{2,0.975}=4.303\). The approximate 95% prediction interval is therefore $$ 0.40\pm4.303(0.2082)=0.40\pm0.896, $$ or \([-0.496,\ 1.296]\) effect units.
This interval is much wider than the confidence interval for \(\mu\). Under the model and its exchangeability assumptions, it describes uncertainty about the true effect in one new population. It does not mean that there is a 95% probability that a fixed effect lies in this particular realized interval, nor does it describe a new noisy study estimate; predicting a future estimate would also include that study's sampling variance. The wide interval here reflects both estimated heterogeneity and the very small number of groups.
Worked Example: When the Estimated Heterogeneity Is Zero
Consider three hypothetical independent groups with estimates \(0.20,\ 0.24,\ 0.16\), each having standard error \(0.10\), on a common effect scale. The fixed-effect mean is \(0.20\), each fixed-effect weight is \(100\), and $$ Q=100\{(0.20-0.20)^2+(0.24-0.20)^2+(0.16-0.20)^2\} =100(0.0032)=0.320. $$ Since \(Q<k-1=2\), the DerSimonian–Laird estimate is \(\hat{\tau}^2_{\mathrm{DL}}=0\).
The random-effects weights are then \(100\), so \(\hat{\mu}=0.20\) and \(SE(\hat{\mu})=\sqrt{1/300}=0.05774\). A normal-approximation 95% confidence interval for the mean is \(0.20\pm1.96(0.05774)=[0.087,\ 0.313]\). The estimate at the zero boundary means this method detected no excess dispersion beyond sampling error; it is not evidence that heterogeneity is certainly absent. With only three groups, the data have limited ability to estimate between-group variation.
Diagnostics, Alternatives, and Limitations
Before interpreting a fitted model, check that the estimates refer to the same effect and that their directions are harmonized. Review whether standard errors and estimates are plausible, whether any one group dominates the weighted mean, and whether conclusions change under a reasonable alternative variance estimator. A forest plot of estimates and confidence intervals can reveal extreme values or patterns that a single variance parameter conceals. Large differences may reflect a mixture of populations or systematic bias, rather than draws from one normal distribution.
A random-effects model differs from a fixed-effect model in its target and weighting. The fixed-effect model assumes one common true effect; it estimates that common effect and generally gives relatively more weight to precise groups. The random-effects model estimates the mean of a distribution of effects and adds \(\hat{\tau}^2\) to each group's sampling variance. This generally reduces the dominance of the most precise groups. If heterogeneity is truly negligible, the two approaches can be similar. A random-effects model should not be selected merely because estimates differ: the scientific target and assumptions determine which model is meaningful.
Important limitations include imprecise estimation of \(\tau^2\) when \(k\) is small, sensitivity to unusual groups, and the normality and exchangeability assumptions used to describe effects. A single \(\tau^2\) cannot explain the sources of heterogeneity. Nor does a prediction interval establish that effects are transportable to a specific population whose relationship to the modeled groups is unknown. If effects are correlated, the independent-estimate formulas above are inappropriate unless the covariance is handled in the model.
The central distinction is between precision of the mean and variability of effects. A large study can estimate \(\mu\) precisely while \(\tau\) remains substantial, so a narrow confidence interval for the mean can coexist with a wide prediction interval. Reporting both \(\hat{\mu}\) with its uncertainty and \(\hat{\tau}\) or a prediction interval gives a more informative account of cross-ancestry results than reporting a pooled estimate alone.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In the model, what do \(\mu\) and \(\tau^2\) describe, and what units does \(\tau\) have?
- Three groups have estimates \(0.30,\ 0.30,\ 0.30\), each with standard error \(0.10\). What are \(Q\) and the DerSimonian–Laird estimate of \(\tau^2\)?
- Why can a prediction interval be wider than a confidence interval for \(\mu\)?
- What assumption is violated if two ancestry-specific estimates come from heavily overlapping participants and their covariance is ignored?
Solutions. (1) \(\mu\) is the mean of the modeled distribution of true ancestry-specific effects; \(\tau^2\) is their between-group variance, and \(\tau\) has the same units as an effect estimate. (2) The fixed-effect mean is \(0.30\), so every deviation is zero and \(Q=0\). Since \(Q<k-1=2\), \(\hat{\tau}^2_{\mathrm{DL}}=\max(0,(0-2)/C)=0\), for \(C>0\). (3) The prediction interval includes estimated between-group variation as well as uncertainty in the mean, through \(\sqrt{\hat{\tau}^2+SE(\hat{\mu})^2}\); it also often uses a larger \(t\) multiplier when \(k\) is small. (4) The independence assumption for sampling errors is violated. Ignoring the covariance can misstate the uncertainty of the combined estimate.