Why the LD Reference Panel Matters
Trans-ancestry fine-mapping uses differences in linkage disequilibrium (LD) across ancestry groups to help distinguish variants that are correlated in one group but less correlated in another. The association evidence is interpreted together with ancestry-specific LD. In practice, LD is often estimated from a reference panel rather than from the participants contributing the association summary statistics. The panel is therefore not just a data-management choice: its estimated correlations help determine how the model explains the observed association pattern.
A reference panel can be unsuitable because its ancestry composition differs from the target group, or because it is too small to estimate LD precisely. These are distinct problems. A large panel from a poorly matched population may estimate the wrong LD structure precisely; a well-matched but small panel may estimate the relevant structure noisily. Sensitivity analysis asks whether the fine-mapping conclusions persist across credible alternatives, while keeping other analysis choices fixed.
In a joint summary-statistic model, the LD matrix helps describe how association statistics at correlated variants relate to one another. A panel-specific \(\hat{R}\) is therefore an estimated input to the likelihood, not direct evidence that any particular variant is causal. Fine-mapping results also depend on the association data, the assumed causal architecture, the prior specification, and other model choices. To isolate panel sensitivity, hold those choices fixed and change only the LD reference input being assessed.
Three Sources of Panel Sensitivity
| Source | What changes | Why it matters |
|---|---|---|
| Ancestry mismatch | The panel’s population composition differs from the target group | Allele frequencies and LD patterns can differ, so the panel may represent the wrong correlation structure |
| Panel sample size | The number of usable, approximately independent individuals changes | Smaller samples generally produce noisier estimates of pairwise LD |
| LD estimation error | The estimated entries of \(\hat{R}\) differ from the underlying population correlations | Fine-mapping can respond to changes in correlations among candidate variants, especially when evidence is hard to distinguish |
“Ancestry-matched” should not be treated as a guarantee of perfect correspondence. A label may cover substantial population diversity, and a reference sample may differ from the association participants in relevant ways. Conversely, a panel that is not labelled identically is not automatically unusable: its suitability must be judged using the study context and available evidence. The practical issue is whether the panel provides a defensible estimate of the LD relevant to the association group.
Sample size affects precision, but does not by itself correct mismatch. Increasing the number of individuals from a population with different LD can narrow uncertainty around an estimate that remains inappropriate for the target group. Related individuals, nonrepresentative sampling, genotyping differences, and inconsistent allele coding can also compromise the estimate; a nominal panel count is not necessarily the effective number of independent, usable individuals.
Quantifying Uncertainty in a Pairwise LD Estimate
For a rough assessment of sampling uncertainty in one pairwise correlation, Fisher’s transformation is useful. If \(r\) is the sample correlation from \(n\) approximately independent observations, then \(z=\operatorname{atanh}(r)\) is approximately normal with standard error \(1/\sqrt{n-3}\), under the usual large-sample approximation for correlation. This is an approximate diagnostic for a pair, not a complete uncertainty model for an entire LD matrix.
The approximation is most defensible when observations are independent and the correlation estimate is not being distorted by strong departures from the assumptions behind the correlation approximation. In genetic reference panels, relatedness and population structure can make the nominal \(n\) overstate the effective independent sample size. Pairwise intervals also do not capture all dependence constraints among entries of an LD matrix. They are helpful for understanding scale and sampling uncertainty, but should not be mistaken for a full probabilistic treatment of \(\hat{R}\) in fine-mapping.
Worked Example: A Noisier Correlation Estimate in a Smaller Panel
Suppose a hypothetical ancestry-matched panel has \(n=100\) approximately independent individuals and estimates the correlation between two variants as \(r=0.20\). Calculate an approximate 95% interval, and compare its width with what the same estimate would have in a much larger panel.
First transform the estimate: \(z=\operatorname{atanh}(0.20)\approx0.202733\). The transformed-scale standard error is \(SE_z=1/\sqrt{100-3}=1/\sqrt{97}\approx0.101535\). Thus the transformed interval is \(0.202733\pm1.96(0.101535)\), or approximately \(0.0037\) to \(0.4017\).
Transforming the endpoints back gives \(\tanh(0.0037)\approx0.004\) and \(\tanh(0.4017)\approx0.3814\). The approximate correlation interval is therefore 0.004 to 0.381. This wide range illustrates how uncertain a modest estimated correlation can be in a small panel; it does not say that the true correlation has a 95% probability of lying in this particular interval.
The calculation concerns one pair of variants under an approximation. It does not establish that the panel is ancestry-matched or that its entire LD matrix is accurate. Those require separate checks.
Worked Example: How a Larger Panel Narrows the Approximate Interval
Now suppose a hypothetical panel estimates the same \(r=0.20\), but has \(n=2498\) approximately independent individuals. The point estimate is unchanged; only the amount of sampling information differs.
The transformed estimate remains \(z=\operatorname{atanh}(0.20)\approx0.202733\). The standard error is \(SE_z=1/\sqrt{2498-3}=1/\sqrt{2495}\approx0.020020\). The margin of error is \(1.96(0.020020)\approx0.039239\), so the transformed endpoints are approximately \(0.202733-0.039239=0.163494\) and \(0.202733+0.039239=0.241972\). Rounded to four decimal places, this is about \(0.1635\) to \(0.2420\).
Applying \(\tanh\) gives an approximate correlation interval of \(0.1621\) to \(0.2374\). It is narrower than the interval from the 100-person panel because the standard error is smaller. This comparison isolates the effect of sample size only under the hypothetical assumption that both panels estimate the same underlying population correlation and have equally suitable sampling.
Comparing Fine-Mapping Results Across Panels
A useful sensitivity comparison begins with a fixed analysis specification: the same harmonized variants and alleles, association summary statistics, priors, causal model, and computational settings. Then fit the analysis using each defensible ancestry-matched reference panel in turn. Record the posterior inclusion probabilities (PIPs), credible sets, and any locus-level interpretation that depends on them. The PIP for a variant is its posterior probability of inclusion in the causal configuration under the specified model and inputs; it is not the probability that the variant is truly causal without qualification.
Compare both the overall pattern and the changes in individual candidates. A leading variant that stays near the top across plausible panels is more robust to the tested LD inputs than one whose ranking changes sharply. A credible set that expands, contracts, or replaces members across panels is panel-sensitive. Such movement is a reason to report uncertainty about localization, not a reason to select whichever panel produces the narrowest set.
Worked Example: Detecting Panel-Sensitive Posterior Evidence
Consider a hypothetical locus with four candidate variants. All fine-mapping choices other than the LD input are held fixed. Two plausible reference panels produce the following illustrative PIPs:
| Variant | Panel A PIP | Panel B PIP |
|---|---|---|
| Variant 1 | 0.62 | 0.31 |
| Variant 2 | 0.25 | 0.51 |
| Variant 3 | 0.08 | 0.12 |
| Variant 4 | 0.05 | 0.06 |
For each panel, the four PIPs sum to 1.00. Under Panel A, variants 1, 2, and 3 accumulate \(0.62+0.25+0.08=0.95\), forming a cumulative-PIP set that reaches 0.95. Under Panel B, the first three in descending order—variants 2, 1, and 3—sum to \(0.51+0.31+0.12=0.94\), so variant 4 is also needed to reach at least 0.95; all four are included in the cumulative-PIP set. These sets have 95% posterior coverage if exactly one variant is causal; with multiple causal variants, cumulative PIPs do not determine the probability that at least one variant in a set is causal.
The leading variant changes from variant 1 to variant 2, and the cumulative-PIP set expands from three variants to four. These sets have 95% posterior coverage under an exactly-one-causal-variant model; under a model allowing multiple causal variants, the probability that at least one variant in a set is causal must be calculated from the joint posterior over causal configurations. Precise localization is sensitive to the LD reference. These hypothetical PIPs illustrate how to read a comparison; they are not calculated from the table alone.
Why LD Uncertainty Can Change Fine-Mapping
A brief model-based argument explains the sensitivity. Let \(Z\) denote a vector of standardized association statistics across variants for an ancestry group, and let \(R\) be the population LD correlation matrix for that group. Under a joint summary-statistic model, the covariance structure of \(Z\) depends on \(R\). Replacing \(R\) by an estimate \(\hat{R}\) changes the likelihood assigned to candidate causal configurations. Since posterior probabilities combine likelihood and prior information, changing \(\hat{R}\) can change PIPs even when the observed association statistics and prior are fixed.
The implication follows directly from Bayes’ rule. For candidate configuration \(C\), posterior support is proportional to its likelihood times its prior probability:
If a different panel gives \(\hat{R}_{A}\) rather than \(\hat{R}_{B}\), the likelihood term can change from \(P(Z\mid C,\hat{R}_{A})\) to \(P(Z\mid C,\hat{R}_{B})\). Unless those likelihood changes are the same across all configurations, their normalized posterior probabilities can change. This is why a variant’s PIP may shift even though its own association statistic is unchanged: the model evaluates the pattern of evidence jointly, including correlations with other variants.
The direction of the change is not predetermined. A revised LD estimate may make an association pattern more compatible with one candidate configuration and less compatible with another. Neither an increase nor a decrease in a PIP, by itself, identifies which panel is correct. Panel suitability must be assessed using ancestry match, sample composition, genotype quality, and other relevant information about the reference data.
A Practical Sensitivity Workflow
Use the same variant set, allele orientation, summary statistics, priors, model, and analysis settings. Verify that LD matrices use matching variant order and allele coding.
Document ancestry relevance, usable sample size, relatedness, and any filtering. Do not treat a larger panel as automatically more appropriate.
Compare PIPs, leading variants, credible-set membership, and set size. Identify which findings are stable and which depend materially on the LD input.
State the panel choices and the conclusions that change. If a result is panel-sensitive, describe localization as uncertain under the tested inputs rather than presenting one result as definitive.
Common Pitfall and Interpretation
A common mistake is to reason: “Panel A gives the smallest credible set, so Panel A must be the best panel.” That conclusion does not follow. A smaller set can reflect a more informative and suitable LD estimate, but it can also reflect an inaccurate correlation pattern that makes the model overconfident. Instead, assess panel relevance independently and report how the findings vary across plausible panels.
Sensitivity comparisons are conditional on the alternatives tested. Stability across two panels does not prove that the LD input is correct, and instability does not prove which panel is preferable. The value of the comparison is diagnostic: it shows whether the fine-mapping interpretation relies strongly on a particular reference-panel estimate. This is especially important when candidate variants have similar association evidence or are correlated, because modest changes in the estimated LD structure can redistribute posterior support.
Check Your Understanding
Answer each question using the distinctions and calculations developed above.
- Why can a very large LD reference panel still be unsuitable for a target ancestry group?
- For an estimated pairwise correlation of 0.30 from \(n=200\) independent individuals, write the formula for the approximate Fisher-transformed standard error. What happens to this standard error as \(n\) increases?
- Two panels produce different leading variants but use the same association statistics and priors. Give a model-based reason their PIPs can differ.
- Why is it not valid to treat a 95% credible set from one panel as proof that its member variants are causal?
- When two credible sets differ across plausible panels, what should be reported about the fine-mapping conclusion?