Why the LD Reference Panel Matters
A polygenic score combines allele dosages across variants, but the weights assigned to those variants can depend on how the variants are correlated. Those correlations are described by linkage disequilibrium (LD). If a score-construction method uses LD to convert marginal association estimates into joint weights, the LD reference panel is part of the statistical model, not merely a source of convenient data.
The central question is not simply whether a panel has the same broad ancestry label as the score’s target population. The panel should represent the population whose LD structure the particular construction method needs. For example, a method that models how discovery GWAS marginal associations arose from discovery-population LD needs an appropriate discovery-LD reference. A method explicitly constructing target-aligned weights may require target-population LD. Substituting one for the other without regard to the method’s assumptions can change the weights in ways that do not have the intended interpretation.
A Two-Variant Model for LD-Aware Weights
Consider a linear model for a quantitative outcome, \(Y=G_1\beta_1+G_2\beta_2+\varepsilon\), where each genotype \(G_j\) has been centered and scaled to variance 1. Assume the residual \(\varepsilon\) has mean zero and is uncorrelated with both genotypes. Let \(R\) be the genotype correlation matrix, and let \(b_j\) denote the marginal regression coefficient of \(Y\) on \(G_j\), measured in outcome units per one genotype standard deviation. Then
The second equality follows because \(b_j=\operatorname{Cov}(G_j,Y)\). Substituting the outcome model gives \(b_j=\sum_k\operatorname{Cov}(G_j,G_k)\beta_k\), and the genotype variances are 1, so these covariances are the entries of \(R\). Thus, when \(R\) is known and nonsingular, joint weights can be recovered from marginal coefficients by solving \(R\beta=b\). In practice, GWAS coefficients and the LD matrix are estimated, and real score methods may add shrinkage or other modelling steps.
This identity also clarifies an important cross-ancestry limitation. It describes marginal coefficients and LD from the same population under the stated model. If marginal GWAS estimates come from a discovery population, the LD needed to interpret those marginal estimates is generally discovery-population LD. A target-ancestry panel is not automatically a better replacement. It is appropriate when the method’s stated construction targets that population’s LD, or when its model explicitly accounts for the difference. A panel choice should follow the method’s estimand and assumptions, not an assumption that target-matched LD is always preferable.
Worked Comparisons of Reference Panels
Worked Example: Matched and Mismatched LD Give Different Weights
Suppose a hypothetical scoring analysis has two marginal coefficients, \(b_1=0.100\) and \(b_2=0.080\), in outcome units per genotype standard deviation. For this illustration, assume the relevant model is the two-variant system above and compare a panel matching the LD structure required by the procedure, \(r=0.20\), with a mismatched panel, \(r=0.80\). The example isolates the effect of the LD estimate; it does not establish which set of weights will predict better in a real target sample.
With the matched panel, the determinant is \(1-r^2=1-0.20^2=0.96\). The weights are
With the mismatched panel, \(1-r^2=1-0.80^2=0.36\), giving
The same marginal coefficients have produced different joint weights: the second variant’s weight is zero under the mismatched panel. That is a consequence of the assumed correlation pattern, not evidence that the variant has no effect. Whether either construction is appropriate depends on which population’s LD the score method is intended to model.
Worked Example: More Individuals Reduce Sampling Error, Not Mismatch
Suppose a hypothetical reference panel contains unrelated individuals and the correlation between two standardized genotypes is estimated as \(r=0.25\). Under a bivariate-normal approximation, the standard error of a sample correlation is approximately \((1-r^2)/\sqrt{m}\), where \(m\) is the number of individuals. This approximation is illustrative; ancestry structure, relatedness, and other departures from the model can affect actual precision.
For \(m=100\), the approximate standard error is
For \(m=2{,}500\), it is
A larger panel can therefore estimate this correlation more precisely if its sampling assumptions are reasonable. But increasing the size of a panel from a different LD population does not make its population correlation become the target’s correlation. More observations can reduce random error around the wrong population quantity while leaving systematic mismatch intact.
To see how the estimated correlation affects weights, use \(b_1=0.100\) and \(b_2=0.080\) again. If the supplied correlation is \(0.25\), the weights are \((0.100-0.25(0.080))/(1-0.25^2)=0.08533\) and \((0.080-0.25(0.100))/(1-0.25^2)=0.05867\). These differ from the weights at \(r=0.20\), even though the change in correlation is modest.
Why High LD Can Make Weights Unstable
The two-variant solution divides by \(1-r^2\). When \(|r|\) is close to 1, the variants carry nearly redundant genotype information, the denominator is small, and small errors in \(r\) can produce large changes in the weights. More generally, a matrix with nearly dependent columns has a small eigenvalue; inversion can amplify errors in its entries. The problem is not that LD is intrinsically undesirable, but that the data may contain little information for separating the contributions of highly correlated variants.
The following calculation makes this sensitivity concrete. It assumes the same marginal coefficients are used for both candidate LD estimates, so that only the panel-based correlation changes.
Worked Example: Small LD Error Changes Weights in a High-LD Region
Take hypothetical marginal coefficients \(b_1=0.100\) and \(b_2=0.080\). A panel estimates \(r=0.95\). Since \(1-r^2=0.0975\), the weights are
Now suppose a different panel or sampling fluctuation gives \(r=0.96\). Then \(1-r^2=0.0784\), and
A change of only \(0.01\) in the supplied correlation has changed both weights appreciably. The negative weight is not automatically an error: conditional weights can differ in sign from marginal associations when predictors are strongly correlated. But in this setting the size and sign of weights are sensitive to the LD input, so they should not be interpreted as precisely determined biological effects without further evidence.
Representativeness, Panel Choice, and Stability Checks
A reference panel is useful only to the extent that it represents the LD structure required by the analysis. Broad ancestry labels are imperfect summaries: individuals within a label can differ in ancestry, and a target group can have a different composition from the panel. In admixed or heterogeneous samples, a single pooled correlation matrix may not describe every subgroup. Variant selection and allele coding must also agree across the GWAS summary data and reference genotypes; otherwise, apparent LD discrepancies can arise from incompatible inputs rather than population differences.
Panel size and representativeness address different sources of error. A larger sample generally improves the precision of estimated correlations when individuals are sampled from the relevant population and dependence is handled appropriately. It does not correct population mismatch, inconsistent variant sets, or systematic selection. Conversely, a well-matched but small panel may have noisy estimates, especially for rare variants or strongly correlated blocks.
When estimated LD is noisy or nearly singular, some procedures use regularization or shrinkage. For example, replacing an unstable inverse with a solution based on \(R+\lambda I\), where \(\lambda>0\) and \(I\) is the identity matrix, can reduce sensitivity to small eigenvalues. This changes the estimator and generally introduces bias in exchange for improved stability; it does not repair a mismatched panel. Regularization choices should be reported as part of the weight-construction procedure.
Useful diagnostics include comparing the relevant LD entries across plausible panels, checking whether the constructed weights change substantially when the reference panel changes, and identifying blocks where the LD matrix is poorly conditioned. Such checks diagnose dependence on the reference data; they do not by themselves identify the best-performing score. Predictive performance is an empirical question and requires appropriate evaluation in the intended target population.
Determine whether the construction models discovery-population LD, target-population LD, or another explicitly defined LD quantity.
Check population representativeness, sample size, variant overlap, allele coding, and whether relatedness or composition could distort the estimates.
Compare plausible panels or LD estimates, paying particular attention to highly correlated variants and near-singular matrix blocks.
Recognize that weights depend on the GWAS estimates, the LD input, and any regularization; they are not automatically direct measures of variant-specific causal effects.
Common Pitfall
Wrong reasoning: “The score will be applied to this target ancestry, so any target-ancestry LD panel must be the correct panel for every step of weight construction.” The conclusion does not follow. If the scoring procedure uses LD to explain marginal associations from a discovery GWAS, replacing discovery LD with target LD may violate the model that relates those associations to joint effects.
Correction: First specify what population’s LD the algorithm requires at each step. Then use a panel representative of that population, or use a method that explicitly accommodates differences between discovery and target LD. A target-matched panel can be valuable, but matching is defined relative to the method’s statistical role for LD, not simply the ancestry of the people who will receive the score.
Check Your Understanding
Answer each question using the model and distinctions developed above.
- A method uses discovery GWAS marginal coefficients and models how they arose from discovery-population LD. Why is a target-ancestry panel not automatically the appropriate replacement?
- In the two-variant model, what happens to the denominator of the weight formulas as \(|r|\) approaches 1, and why can this make weights unstable?
- Explain the difference between reducing sampling error by increasing panel size and reducing systematic LD mismatch.
- If a high-LD panel produces a negative conditional weight for one variant, does that alone establish a biological protective effect? Why or why not?
- Name two practical checks that can reveal whether estimated weights depend strongly on the selected LD reference panel.