Why the LD Reference Panel Needs an Audit
Linkage disequilibrium (LD) is correlation among alleles at different variants in a population. In polygenic risk score work, an LD reference panel is a set of genotyped individuals used to estimate these correlations. Those estimates can inform methods that use LD alongside genome-wide association study (GWAS) results. A panel is useful only to the extent that it represents the LD structure relevant to that use.
There are three distinct data sources to keep straight. The discovery cohort supplies GWAS estimates; the target cohort is the population in which a score may be applied or evaluated; the reference panel supplies genotype data for estimating LD. They can differ in ancestry, sample size, and which variants were measured or passed quality control. A large panel is not automatically a good panel, and matching a target cohort does not necessarily make a panel appropriate for modelling discovery GWAS results.
The appropriate ancestry match depends on the purpose. If LD estimates are being used to interpret or model GWAS summary statistics from a discovery cohort, the panel should represent the ancestry composition and relevant LD structure of that discovery cohort. A target cohort with a different ancestry does not, by itself, make its LD the right reference for those discovery statistics. If the method instead needs LD in a particular target population, the relevant target structure must be considered. For multi-ancestry discovery data, a single pooled panel may obscure real differences; ancestry-specific estimates or a method designed for that setting may be needed.
Ancestry labels are useful summaries, not guarantees of genetic similarity. Within-group variation, admixture, recruitment differences, and the mixture of ancestries in a cohort can all matter. Compare available ancestry information and, where possible, allele frequencies and LD patterns in the actual variants of interest. The earlier material on how LD changes marginal GWAS effects explains why the correlation structure can affect the relationship between marginal estimates and underlying effects; here the question is whether the panel estimates the relevant structure well.
Three Audit Dimensions
A practical audit asks three separate questions. First, does the panel represent the cohort whose LD is needed? Second, are there enough effective, unrelated individuals to estimate the correlations with acceptable precision? Third, does the panel cover the variants required by the analysis? These checks are complementary: success on one does not compensate automatically for failure on another.
| Dimension | What to compare | Warning signs |
|---|---|---|
| Ancestry and population structure | Discovery and target composition, panel composition, and relevant genetic similarity | Broad ancestry mismatch, substantially different mixture proportions, or discordant allele frequencies |
| Sample size | Number of usable, effectively independent panel individuals | Small effective sample, unstable correlations, or large uncertainty in key LD estimates |
| Variant coverage | Variants needed by the analysis versus variants available and retained in the panel | Low overlap, missing key variants, inconsistent variant definitions, or restricted coverage in relevant regions |
For ancestry, compare the panel to the cohort whose LD is being represented, not simply to whichever cohort is easiest to inspect. Summary ancestry percentages can be a first screen. For example, a discovery cohort composed mostly of one ancestry group and a panel composed mostly of another deserve investigation, even if both contain a small proportion of the same group. If genetic principal components or comparable information are available, they can refine this comparison. Allele-frequency differences at shared variants can provide an additional diagnostic, but do not prove that all LD patterns match.
For sample size, use the number of individuals remaining after variant and sample quality control, and account for relatedness or other dependence when judging effective information. A nominal panel of 2,000 people can contain less independent information than 2,000 unrelated individuals. Precision is also variant-specific: rare alleles and variants with substantial missingness can yield less stable estimates than common, well-measured variants.
Coverage should be calculated against the actual required variant set, after making sure the variants are defined on compatible genomic coordinates and allele coding. A raw count of shared variants can conceal an important gap: a panel may cover nearly all common variants but omit a small set that is essential to the planned analysis. Conversely, a few missing variants may be inconsequential if they are not required. Coverage is a property of the particular analysis, not a universal percentage cutoff.
Precision of an Estimated Correlation
A useful rough calculation illustrates why sample size matters. Under independent sampling and a bivariate-normal approximation, the Fisher transform \(Z=\operatorname{atanh}(r)\) of a sample correlation \(r\) has approximate standard error \(1/\sqrt{n-3}\), where \(n\) is the number of independent individuals with usable data for both variants. Genotype dosages are discrete, so this approximation is a guide rather than a guarantee; it is least reassuring for rare variants, strong structure, or small samples.
The first approximation is the large-sample variance of a sample correlation when the population correlation is \(\rho\). Applying the delta method to \(g(r)=\operatorname{atanh}(r)\), whose derivative is \(g'(r)=1/(1-r^2)\), cancels the factor involving the correlation. The conventional finite-sample approximation replaces \(n\) by \(n-3\). An approximate interval can be formed on the transformed scale and then mapped back with the hyperbolic tangent:
This interval is a diagnostic for uncertainty in an estimated pairwise correlation, not a test of whether two cohorts have identical LD. Comparing whole LD structures requires examining many correlations and appropriate uncertainty, rather than treating each pairwise interval as a definitive pass/fail rule.
Worked Example: Ancestry Composition Can Outweigh Panel Size
Suppose a hypothetical discovery GWAS is 90% Group A and 10% Group B. A hypothetical target cohort is 25% Group A and 75% Group B. Two candidate panels have compositions A: 88%, B: 12% for Panel 1, and A: 30%, B: 70% for Panel 2. Panel 1 has 5,000 individuals; Panel 2 has 1,200.
For LD intended to represent the discovery GWAS, Panel 1 is much closer by these broad composition summaries: its Group A proportion differs by \(90\%-88\%=2\) percentage points, compared with \(90\%-30\%=60\) points for Panel 2. Panel 1 is therefore the more plausible ancestry match to the discovery cohort, subject to checking finer structure and actual shared variants.
For an LD task specifically intended to represent the target cohort, the comparison reverses. Panel 2 differs from the target Group A proportion by \(30\%-25\%=5\) points, while Panel 1 differs by \(88\%-25\%=63\) points. Panel 2 is the closer broad match, but its smaller sample means its estimates may be less precise.
The conclusion is not that either panel is universally best. The reference must match the cohort whose LD is needed. A larger but mismatched panel can estimate the wrong population’s correlations very precisely; sample size does not repair that mismatch.
Worked Example: Comparing Correlation Precision
Consider two hypothetical panels with \(n=400\) and \(n=1{,}600\) independent individuals. For a pair of common variants, each gives an estimated correlation \(r=0.30\). Assume for illustration that both panels are well matched in ancestry and that the bivariate-normal approximation is reasonable.
For the smaller panel, the Fisher-scale standard error is \(1/\sqrt{400-3}=1/\sqrt{397}\approx0.0502\). Since \(\operatorname{atanh}(0.30)\approx0.3095\), an approximate 95% interval on that scale is \(0.3095\pm1.96(0.0502)\), or approximately \([0.2111,0.4079]\). Applying \(\tanh\) gives an interval for \(r\) of approximately \([0.208,0.387]\).
For the larger panel, the standard error is \(1/\sqrt{1{,}600-3}=1/\sqrt{1{,}597}\approx0.0250\). The Fisher-scale interval is \(0.3095\pm1.96(0.0250)\), or approximately \([0.2605,0.3585]\). Transforming back gives approximately \([0.255,0.344]\).
The larger panel yields a narrower approximate interval, as expected. Both estimates are 0.30, but that agreement does not establish that the panels match the discovery cohort: ancestry suitability is a separate check. These intervals also describe sampling uncertainty under the stated approximation, not uncertainty from ancestry mismatch or variant-definition errors.
Variant Coverage and an Integrated Decision
Count coverage against the set of variants that the intended analysis needs. If \(M\) variants are required and \(M_{\mathrm{shared}}\) are available in both the panel and the analysis data after quality control, the unweighted coverage fraction is \(M_{\mathrm{shared}}/M\). This summary is easy to communicate, but it treats all variants equally and does not explain why variants are absent. Report missingness by region or variant class when that affects the intended use.
Worked Example: High Overall Coverage Can Hide a Critical Gap
A hypothetical analysis requires 100 variants. Panel A contains 96 of them, so its unweighted coverage is \(96/100=0.96\), or 96%. However, four absent variants are the only required variants in a region the analysis specifically needs to represent. Panel B contains 92 of the 100 variants, so its coverage is 92%, and its eight missing variants are outside that region.
By the overall percentage alone, Panel A appears better. But if the analysis requires LD information for the specified region, Panel A fails that requirement: it has none of the four needed variants there. Panel B may be more suitable for that particular task despite lower total coverage, provided it includes the necessary regional variants and passes the ancestry and precision checks.
This example shows why a coverage percentage is a screen rather than a decision rule. The audit should list which variants are absent, whether absence is concentrated, and whether the missing variants matter to the analysis. It should not infer that the panel’s remaining variants can substitute for missing ones.
An integrated decision can be organized as a short sequence. First specify the cohort whose LD must be represented and the variants needed. Then compare candidate panels on ancestry, effective sample size, and coverage for that use. Finally, document unresolved concerns and decide whether to use a candidate, seek another panel, or limit the analysis. There is no universal sample-size or coverage threshold that makes every panel adequate: tolerable uncertainty depends on the method and the role of LD in it.
Identify whether the required structure is that of the discovery cohort, the target population, or a specified ancestry component of a multi-ancestry analysis.
Compare ancestry composition and, where available, finer genetic structure and allele frequencies at shared variants. Investigate discrepancies rather than relying on labels alone.
Use the post-quality-control effective sample size, assess the stability of relevant correlations, and enumerate missing required variants.
State which limitations are acceptable for the intended analysis. If no candidate represents the needed population or covers essential variants, do not describe it as an adequate match.
Diagnostics, Alternatives, and Limitations
Several patterns should prompt caution: estimated correlations change substantially across subsets or resamples; allele frequencies at shared variants differ sharply from the cohort to be represented; the panel’s ancestry mixture is unlike that cohort; or required variants are systematically missing. These are signals for investigation, not automatic proof of failure. Differences may reflect sampling variability, data processing, or genuine population structure. Check that variants are comparable before attributing a discrepancy to ancestry.
The main alternatives are not interchangeable. A single ancestry-matched panel is straightforward when the discovery cohort is reasonably homogeneous and the panel is large and well covered. A larger but less well-matched panel may have smaller sampling error but greater population mismatch. Separate ancestry-specific panels can be more appropriate for multi-ancestry data, but require analyses that use those groups coherently. Estimating LD in the target cohort is relevant when target LD is the goal and suitable target genotypes exist; it does not automatically provide the correct LD for discovery summary statistics. Each choice changes the estimand or assumptions, so report the reason for the choice.
An LD reference panel cannot remove limitations in the underlying GWAS, guarantee transfer of predictive performance, or make unobserved variants available. Its estimates can be distorted by small sample size, related individuals, population substructure, missingness, and inconsistent variant definitions. Broad ancestry categories also fail to capture all within-group variation. A suitability audit therefore supports a reasoned choice; it does not certify that the panel is a perfect representation.
A common mistake is to argue, “This panel is the largest available, so it must be the best reference.” That reasoning considers precision but ignores whether the estimated correlations describe the relevant cohort. The correction is to evaluate ancestry match, effective sample size, and required variant coverage separately, then make a use-specific decision. Another mistake is to choose the panel that matches the target merely because the target is where predictions will be used. For LD estimates intended to represent discovery GWAS statistics, the discovery population remains the relevant comparison.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A GWAS discovery cohort is predominantly Group A, but the target cohort is predominantly Group B. If a method uses LD to model the discovery GWAS statistics, which cohort should guide the panel’s ancestry match?
- A panel estimates a correlation from 100 independent individuals. Using the Fisher approximation, what is the approximate standard error on the transformed scale?
- A panel covers 180 of 200 required variants. What is its unweighted coverage fraction, and why is that not sufficient to decide suitability?
- Why can a very large panel still be a poor LD reference?
- Name two reasons why an observed difference in allele frequency between a panel and a cohort should be investigated before it is attributed to ancestry mismatch.
Solutions. (1) The discovery cohort, because its LD is the structure the method is using to model those GWAS statistics; target ancestry is not a substitute. (2) \(1/\sqrt{100-3}=1/\sqrt{97}\approx0.102\). (3) \(180/200=0.90\), or 90%; the missing variants may be concentrated among variants essential to the analysis, and the panel must also pass ancestry and precision checks. (4) Its ancestry or population structure may not represent the cohort whose LD is needed, so it can estimate the wrong correlations precisely. (5) Variant definitions or allele coding may not be comparable, and sampling variability, missingness, or quality-control differences may also explain the discrepancy.