Why Apparent Resolution Can Be Misleading
Trans-ancestry fine-mapping can use differences in linkage disequilibrium (LD) to distinguish variants that are difficult to separate in one ancestry. But a small credible set is not automatically a reliable result. Incorrect allele alignment, an unsuitable LD matrix, unmodeled effect differences, incomplete variant coverage, or unequal information across ancestry groups can make the analysis look more decisive than it is.
The practical question is not only, “Which variant has the largest posterior probability?” It is also, “Would this ranking remain credible if the inputs and assumptions were checked?” The earlier material on ancestry-specific effects, posterior probabilities, LD references, variant coverage, and sample-size imbalance supplies tools for answering that question. Here the focus is recognizing combinations of failures and choosing checks that reveal them.
Failure Mode 1: Allele Flips Masquerading as Effect Heterogeneity
A reported genetic effect is tied to an effect allele: the allele whose dosage is counted in the association analysis. Reversing which allele is counted reverses the sign of the effect estimate. Thus, opposite signs across ancestry-specific results do not by themselves establish that the biological effects differ. First check whether the estimates refer to the same effect allele.
This issue matters for fine-mapping because trans-ancestry methods compare evidence across groups. A sign error can make a shared signal look heterogeneous, or can make incorrectly oriented evidence contribute in an unintended direction. Inspect effect and other alleles, allele frequencies, and effect directions in the association results before interpreting differences. A strand or coding ambiguity should be resolved or handled explicitly, not silently treated as agreement.
Worked Example: An Apparent Sign Reversal
Suppose a hypothetical variant has alleles A and G. In ancestry group 1, dosage is counted for G, and the estimated effect is \(0.12\) outcome units per G allele, with standard error \(0.03\). In ancestry group 2, dosage is counted for A, and the estimate is \(-0.11\) outcome units per A allele, also with standard error \(0.03\). The estimates are not yet oriented to the same allele.
For a biallelic variant, the A dosage is the complement of the G dosage: \(G=2-A\). Changing the counted allele from A to G therefore changes the sign of the slope. Orienting group 2 to G gives
The two estimates oriented to G are \(0.12\) and \(0.11\) outcome units per G allele. Their signs agree, and their difference is \(0.01\), not the \(0.23\) difference suggested by comparing \(0.12\) directly with \(-0.11\). For illustration, an equal-weight average of the unaligned estimates is \((0.12-0.11)/2=0.005\), whereas the equal-weight average after alignment is \((0.12+0.11)/2=0.115\). These averages are only a demonstration of the sign problem, not a substitute for a model that accounts for the study designs and their uncertainties.
The diagnostic is to verify the allele being counted and orient both estimates consistently before assessing effect heterogeneity or combining evidence. The aligned estimates are compatible in direction in this example, but this alone does not prove that the effects are identical.
Failure Mode 2: Ancestry-Mismatched LD Can Favor the Wrong Candidate
Association statistics at nearby variants are correlated because the genotypes are correlated. For standardized genotypes and a linear model with a vector of genetic effects \(\boldsymbol{\beta}\), the expected vector of marginal association coefficients is proportional to \(R\boldsymbol{\beta}\), where \(R\) is the genotype correlation matrix. If just one variant has a nonzero effect, the expected marginal pattern is proportional to the corresponding column of \(R\). This is why a fine-mapping analysis needs LD that represents the study ancestry and variants.
The following justification states the relationship in a simple setting. Let \(X_j\) be a standardized genotype at variant \(j\), let the outcome satisfy \(Y=\sum_k X_k\beta_k+\varepsilon\), and assume the residual has mean zero and is uncorrelated with the genotypes. The marginal slope for variant \(j\) is its covariance with the outcome. Therefore,
If only variant \(c\) has a nonzero effect, this becomes \(\operatorname{Cov}(X_j,Y)=R_{jc}\beta_c\): the marginal pattern follows column \(c\) of the relevant LD matrix. With comparable sample sizes and standardized association statistics, expected \(Z\)-score patterns have the same LD-based structure up to a scale factor. The approximation is a diagnostic intuition; real fine-mapping likelihoods also account for sampling uncertainty and their specified data model.
Worked Example: A Valid LD Mismatch Favors the Wrong Candidate
Consider three variants and a hypothetical ancestry in which variant 1 is causal. Suppose the true genotype correlation matrix for that ancestry is
This is a valid correlation matrix: its diagonal entries are 1, its pairwise correlations are within \([-1,1]\), and its determinant is \(1+2(0.95)(0.20)(0.20)-(0.95^2+0.20^2+0.20^2)=0.0935>0\). In a simplified association-score scale, let the causal signal have size 5. The expected association pattern is five times column 1:
Now suppose the analysis uses a mismatched reference matrix
This is also a valid correlation matrix: its determinant is \(1+2(0.95)(0.45)(0.20)-(0.95^2+0.45^2+0.20^2)=0.026>0\), and its two-variable principal minors are positive. Compare the observed pattern to five times each candidate column, using the sum of squared discrepancies as a deliberately simple diagnostic score. Candidate 1 predicts \((5,4.75,2.25)\), so
Candidate 2 predicts \((4.75,5,1)\), giving
Under this simplified comparison, the mismatched reference fits candidate 2 better, even though candidate 1 generated the pattern. With the true matrix \(R_A\), candidate 1 reproduces the pattern exactly. This calculation is not a formal fine-mapping posterior calculation; it isolates how an incorrect LD pattern can mislead a candidate comparison. A practical check is to repeat the analysis with ancestry-matched LD where possible and examine whether the ranking or credible set changes under plausible alternative references.
Failure Mode 3: Shared Causality Does Not Guarantee a Shared Effect
A variant can be causal in more than one ancestry while its estimated effect differs between ancestries. This distinction—causal sharing versus effect portability—is important: a model that imposes one common effect can give misleading combined evidence when effects are heterogeneous. Conversely, a difference in estimates can reflect sampling variation, allele misalignment, or model problems rather than genuine biological heterogeneity. Interpret a heterogeneity diagnostic only after checking orientation and the analysis inputs.
Worked Example: A Common-Effect Summary Hides a Difference
Suppose, after aligning the effect allele, two hypothetical ancestry-specific studies estimate effects of \(0.20\) and \(-0.10\) outcome units per effect allele, with standard errors \(0.05\) and \(0.04\). Assume independent estimates and consider the fixed-effect inverse-variance summary as a diagnostic of what a common-effect assumption would imply. The weights are \(w_1=1/0.05^2=400\) and \(w_2=1/0.04^2=625\). The weighted mean is
The summary is close to zero, but that does not mean both ancestry-specific effects are near zero. Their disagreement can be assessed with Cochran’s \(Q=\sum_i w_i(\hat{\beta}_i-\hat{\beta}_{\mathrm{FE}})^2\). Here,
For two independent estimates, the heterogeneity test has \(df=2-1=1\); a \(Q\) this large is a strong warning against treating these data as straightforward evidence for one common effect. The fixed-effect number is therefore not a sufficient description of the locus. Check allele harmonization, ancestry-specific estimates and uncertainty, and the assumptions of the trans-ancestry model. The calculation flags a modeling concern; it does not establish why the estimates differ or rule out a shared causal variant.
Failure Mode 4: Unequal Variant Coverage Changes the Candidate Set
Fine-mapping ranks variants that are represented in the analysis. If a candidate is absent, poorly imputed, or filtered in one ancestry, that group cannot provide ordinary association evidence for or against that candidate. A posterior probability is conditional on the variants and model included; a posterior renormalized over a reduced candidate set can look concentrated without having learned that the missing variant is unlikely.
Worked Example: A Confident Ranking After a Candidate Is Missing
Suppose a hypothetical analysis in ancestry group 1 includes variants \(v_7\), \(v_8\), and \(v_9\), with relative evidence weights 100, 10, and 1. In this simplified illustration, normalizing those weights gives the probability for \(v_7\) as
Now suppose ancestry group 2 has no usable genotype for \(v_7\). If one normalizes only the weights for the remaining variants \(v_8\) and \(v_9\), the resulting share for \(v_8\) is \(10/(10+1)=0.9091\). That number describes the relative evidence between the two available variants in this toy calculation. It is not evidence that \(v_7\) has been tested and rejected, nor can it be compared as though the two groups had identical candidate sets.
Check which variants were tested, their imputation or genotype quality, and whether the same candidate set enters each ancestry-specific and joint analysis. Treat absent or low-quality variants as a coverage limitation, not as negative association evidence. If coverage differs materially, report the limitation and assess conclusions using a defensible shared set of well-measured variants.
Failure Mode 5: Sample Imbalance Lets One Ancestry Dominate
When sample sizes differ substantially, the ancestry with more information can dominate a combined result. This is not automatically an error: a larger study can legitimately contribute more evidence. The failure occurs when a pooled result is interpreted as strong, balanced cross-ancestry support even though the smaller group contributes little precision or has an unexamined conflicting pattern.
Worked Example: Different Precision for the Same Effect Size
Assume two hypothetical studies have the same allele frequency, outcome variance, and true effect, so that the standard error is approximately proportional to \(1/\sqrt{n}\). If one ancestry has \(n_1=10{,}000\) and the other \(n_2=1{,}000\), the expected standard-error ratio is
For instance, if \(SE_1=0.020\) outcome units per allele, the corresponding approximate standard error in the smaller study is \(SE_2=0.020(3.162)=0.0632\). With an estimate of \(0.10\) in each study, the \(Z\)-statistics are \(0.10/0.020=5.00\) and \(0.10/0.0632=1.58\). The smaller study is less precise; its weaker \(Z\)-statistic is not, by itself, evidence of a different effect.
Inspect ancestry-specific estimates and standard errors alongside the joint result. Consider whether the claimed resolution persists in analyses that make the contribution of each group visible, such as leave-one-ancestry-out comparisons or appropriately designed sensitivity analyses. Do not interpret a small joint credible set as equal support from every ancestry when one group supplies most of the information.
A Practical Diagnostic Sequence
The failure modes can overlap. For example, a poorly covered variant may also have an LD estimate based on a different set of individuals, while an allele-orientation error may create apparent effect heterogeneity. A useful review follows the data path from variant representation to combined inference rather than treating each diagnostic in isolation.
Confirm that effect estimates refer to the same counted allele before interpreting signs or comparing ancestry-specific effects. Investigate ambiguous or inconsistent allele records rather than assuming the signs are comparable.
List which variants were tested in each ancestry and review genotype or imputation quality. A missing candidate is not evidence against that candidate.
Verify that the LD reference is appropriate to the ancestry and variants being analyzed. Compare conclusions across defensible ancestry-matched references or other justified LD sensitivity analyses.
Review effect estimates, standard errors, and heterogeneity diagnostics together. Distinguish a weak estimate caused by limited precision from a precise estimate pointing in a different direction.
Compare variant rankings and credible sets under relevant sensitivity analyses, including leave-one-ancestry-out checks when appropriate. Report instability as uncertainty, not as a single definitive variant ranking.
Common Mistakes and Key Takeaway
- Reading opposite signs as biological heterogeneity before aligning alleles. Reversing the counted allele reverses the effect sign; orient estimates consistently first.
- Assuming that more ancestries automatically improve resolution. Added data help only to the extent that the variants, LD inputs, and model are appropriate and informative.
- Treating an absent variant as evidence against it. A variant that was not measured or analyzed cannot receive ordinary association evidence from that ancestry.
- Calling a common-effect summary portable because it is precise. Precision under a common-effect model does not validate the common-effect assumption; review ancestry-specific estimates and heterogeneity.
- Trusting a small credible set without sensitivity checks. A ranking that changes with plausible LD references or analysis inputs is less robust than its nominal concentration may suggest.
Check Your Understanding
For each question, explain what you would check and why.
- Two ancestry-specific estimates have opposite signs, but the studies report different effect alleles. What should be done before interpreting heterogeneity?
- In a three-variant example, why must each candidate’s predicted marginal-association pattern come from a valid LD-matrix column, with a self-correlation of 1?
- One ancestry lacks a candidate variant, while the posterior over the variants it does contain is highly concentrated. What can and cannot be concluded?
- Why can a smaller \(Z\)-statistic in a smaller ancestry group be consistent with the same underlying effect?
- Name two checks that could reveal whether a fine-mapping result depends strongly on an LD-reference choice.