Why the LD Reference Panel Matters
LD score regression relates association-test statistics to LD scores: variants tagging more nearby genetic variation are expected, under the model, to have larger mean test statistics from polygenic signal. In practice, LD scores are often calculated from a reference panel rather than the GWAS participants. That makes the analysis feasible when only summary statistics are available, but it introduces a crucial requirement: the panel’s LD scores should represent the LD structure of the GWAS population.
A reference panel can differ in allele frequencies, LD patterns, sample composition, or the variants available for calculation. The consequence is not merely that individual LD scores are imperfect. Because the regression uses those scores as its predictor, systematic mismatch can change the slope and intercept that the fitted line assigns to the GWAS statistics. The introductory material on LD score regression establishes the regression framework; here the focus is on how replacing study-population scores with reference-panel scores changes that fit.
The formula \(y_j=a+bz_j+u_j\) describes the conditional mean, not every observed statistic. Observed values also vary around that mean. The analysis below isolates the effect of using \(x_j\) in place of \(z_j\); random sampling variation in the association statistics is an additional source of uncertainty. The regression weights \(w_j\) are taken as fixed and positive for the derivations. Define weighted means by \(\bar{x}_w=\sum_j w_jx_j/\sum_jw_j\), and define weighted covariance and variance using the same normalization:
A fitted slope requires \(\operatorname{Var}_w(x)>0\). If the reference scores have no weighted variation, the regression cannot identify a slope. The same definitions apply to \(u\), such as \(\operatorname{Cov}_w(x,u)\).
What the Regression Estimates Under Mismatch
For a weighted least-squares regression with an intercept, the fitted slope is the weighted covariance of predictor and response divided by the weighted predictor variance. Applying this identity to the mean response \(a+bz_j+u_j\) gives the slope of the weighted linear projection:
This expression is the expected fitted slope when the observed regression errors have conditional mean zero and the scores and weights are treated as fixed. It also gives the exact slope of the regression of the conditional means. The first term is the effect of substituting reference scores for study scores. The second term records systematic departures \(u_j\) that happen to covary with the reference scores. Thus, even an excellent correlation between \(x_j\) and \(z_j\) does not by itself guarantee an unchanged slope: their weighted covariance-to-variance ratio matters.
In the special case \(u_j=0\), if \(b\ne0\), the fitted slope equals \(b\) exactly when \(\operatorname{Cov}_w(x,z)=\operatorname{Var}_w(x)\). If \(b=0\), the fitted slope is zero for any reference–study covariance, as long as \(\operatorname{Var}_w(x)>0\). A mismatch can therefore distort a nonzero slope, but it does not create a slope from a zero mean relationship unless systematic deviations \(u_j\) or other model departures contribute.
The intercept changes as well. For the fitted regression, the intercept is the weighted mean response minus the fitted slope times the weighted mean predictor. Under \(u_j=0\), the exact identity for the fitted intercept is:
when the response consists of the stated mean model without additional errors. If the response includes random errors with weighted mean \(\bar e_w\), add \(\bar e_w\) to the right-hand side. Taking expectations when those errors have mean zero gives \(E[\hat a_{\mathrm{fit}}]=a+b\bar z_w-E[\hat b_{\mathrm{fit}}]\bar x_w\). In particular, the expected-intercept expression contains the expected fitted slope, not the random fitted slope. A changed fitted intercept can therefore reflect the predictor’s mean and the altered slope, even if the underlying \(a\) is unchanged. It is not, by itself, evidence that the underlying confounding process changed.
Worked Example: A Mismatch Changes the Slope
Consider four hypothetical variants with equal weights, no systematic deviations (\(u_j=0\)), and study mean model \(y_j=1+0.2z_j\). Suppose their study and reference scores are \(z=(1,3,7,9)\) and \(x=(2,4,6,8)\), respectively. The reference scores have positive variance, so the regression slope is defined.
The weighted means are \(\bar{x}_w=5\) and \(\bar{z}_w=5\). The weighted variance of the reference scores and covariance with the study scores are
Therefore the slope when regressing the mean response on the reference scores is \(b_{\mathrm{fit}}=0.2(7/5)=0.28\), rather than \(0.2\). The study-model responses are \(y=(1.2,1.6,2.4,2.8)\), with mean \(2\). The fitted intercept is \(2-0.28(5)=0.60\). Equivalently, the intercept identity gives \(1+0.2(5)-0.28(5)=0.60\). The calculations show how the changed slope and predictor scale shift the intercept even though the study-model intercept is \(1\).
Worked Example: Systematic Deviations Can Add Slope Bias
Suppose four hypothetical variants have equal weights, \(x=z=(2,4,6,8)\), and the study mean model is \(y_j=1+0.2z_j+u_j\), with \(u=(-0.2,-0.1,0.1,0.2)\). Here the reference scores match the study scores, but deviations are not unrelated to the predictor. Both score means are \(5\), and \(\operatorname{Var}_w(x)=5\).
The weighted mean deviation is zero. Its weighted covariance with \(x\) is \(\operatorname{Cov}_w(x,u)=[(-3)(-0.2)+(-1)(-0.1)+(1)(0.1)+(3)(0.2)]/4=1.4/4=0.35\). Thus the projected slope is \(0.2+0.35/5=0.27\), not \(0.2\). The mean response is \(1+0.2(5)=2\), so the intercept of the projected line is \(2-0.27(5)=0.65\). This example is not a pure panel mismatch: it shows why a score match alone cannot correct a model departure that covaries with the score.
How Score Correlation Can Mislead
A reference panel may preserve the ordering of variants reasonably well and still produce a different fitted slope. The regression uses the scale and covariance of the scores, not just their rank correlation. If the panel compresses the range of LD scores, changes their scale, or assigns unusually high scores to variants whose study scores are not high, the projection ratio can differ substantially from one. Restricted predictor variation also makes the fitted slope more sensitive to errors and outlying variants.
Worked Example: Similar Score Ranges, Different Projection
Take four hypothetical variants with equal weights, study scores \(z=(1,2,3,4)\), reference scores \(x=(1,3,2,4)\), and mean responses \(y=1+0.5z=(1.5,2,2.5,3)\). Both score sets have mean \(2.5\), and the reference scores have variance \(1.25\).
The weighted covariance is \(\operatorname{Cov}_w(x,z)=[(-1.5)(-1.5)+(0.5)(-0.5)+(-0.5)(0.5)+(1.5)(1.5)]/4=4/4=1\). The fitted slope is therefore \(0.5(1/1.25)=0.4\), much smaller than the study-score slope \(0.5\). The mean response is \(2.25\), so the fitted intercept is \(2.25-0.4(2.5)=1.25\), compared with the study-model intercept \(1\). This small example has only a modest rearrangement in scores, yet the regression projection changes markedly.
Assumptions, Diagnostics, and Practical Choices
The projection calculation is algebraic, but interpreting it as a diagnosis of reference-panel mismatch requires care. The correspondence between variants must be correct: scores and summary statistics need to refer to the same variants, with consistent allele handling and compatible LD-score definitions, windows, and variant filters. The study-population mean relationship must also be a reasonable approximation over the analyzed variants. If these conditions fail, an apparent panel effect may instead reflect data harmonization errors or a misspecified regression.
The panel should be sufficiently representative of the GWAS population’s LD structure for the intended analysis. In finite panels, estimated LD scores also have sampling error. That error can be substantial for small panels or scores dominated by a few estimated correlations. Consequently, a low correlation between study and reference scores may signal a serious mismatch, while a high correlation is reassuring but not conclusive: slope calibration depends on the weighted covariance-to-variance ratio and on deviations from the assumed mean model.
Confirm that summary statistics and reference scores align by variant identity and that the score construction and filtering rules are compatible.
Compare score distributions and, where study-genotype data are available, calculate the weighted correlation and covariance-to-variance ratio between \(x_j\) and \(z_j\).
Check whether fitted slopes and intercepts are sensitive to influential variants, score ranges, weights, or reasonable changes in variant filters.
Compare results across plausible panels using the same summary statistics and regression specification; report the sensitivity rather than selecting a panel solely because its result is preferred.
Useful warning signs include systematic disagreement in score distributions, a fitted slope that changes substantially across reasonable reference panels, influential points concentrated at one end of the score range, and evident curvature or structured residuals. A straight-line fit can conceal these problems if only its slope and intercept are reported. Diagnostics cannot prove that a panel is representative, and matching the mean and range of scores alone is insufficient.
The main alternative, when individual-level GWAS genotypes are available and the analysis permits it, is to calculate LD scores directly in the study sample using a consistent definition. That avoids transporting LD structure from a separate panel, although finite-sample estimation error and other model assumptions remain. When only summary statistics are available, a carefully selected external panel is often necessary; using multiple plausible panels and evaluating sensitivity is preferable to treating one panel as certainly correct. These choices address score construction, not all possible causes of deviations from the mean model.
Common Mistakes and Limits
- Equating high correlation with no bias. Correlation measures linear association after centering and scaling; the slope formula depends on \(\operatorname{Cov}_w(x,z)/\operatorname{Var}_w(x)\). Compare that ratio and the fitted results, not correlation alone.
- Calling every slope change a panel effect. A term \(\operatorname{Cov}_w(x,u)\) also changes the projection. Model misspecification or other structured departures can mimic or compound a reference mismatch.
- Reading the fitted intercept as a direct biological or confounding measure. The intercept depends on the weighted score means and fitted slope as well as the underlying mean model. A shift in the fitted intercept does not identify its cause.
- Ignoring uncertainty in the reference scores. The formulas above describe projections for fixed scores. Estimated scores from finite panels add uncertainty that should be considered when assessing stability.
Reference-panel mismatch is a limitation of the predictor used in LD score regression, not a correction that can be summarized by one universal adjustment factor. The projection formulas explain the direction and scale of slope changes only when the relevant weighted moments and mean-model departures are known. In real analyses those quantities may be incompletely observed, so diagnostics and sensitivity analyses are essential and should be reported with the regression specification.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- If \(u_j=0\), \(b=0.3\), and \(\operatorname{Cov}_w(x,z)/\operatorname{Var}_w(x)=0.8\), what is the projected slope?
- For \(u_j=0\), \(b=0\), and positive reference-score variance, can mismatch alone produce a nonzero projected mean slope? Explain.
- With \(a=1\), \(b=0.2\), \(\bar z_w=5\), \(\bar x_w=4\), and \(E[\hat b_{\mathrm{fit}}]=0.25\), calculate the expected fitted intercept, assuming zero-mean regression errors.
- Why is a high correlation between reference and study LD scores not sufficient to establish that the fitted slope is unchanged?
- Name two checks that can help diagnose an unsuitable reference panel.
Solutions. (1) \(0.3(0.8)=0.24\). (2) No. The mismatch term is \(b\operatorname{Cov}_w(x,z)/\operatorname{Var}_w(x)=0\), provided the reference-score variance is positive. (3) \(1+0.2(5)-0.25(4)=1\). (4) The slope depends on the covariance-to-variance ratio, not correlation alone; score scaling and weighted predictor variation matter. (5) Examples include comparing score distributions and the covariance-to-variance ratio when study scores are available, checking variant correspondence, and comparing fitted results across plausible panels.