LD Score Regression as an Inflation Diagnostic

Why Regress Association Statistics on LD Scores?

A genome-wide association scan can show an excess of large test statistics for more than one reason. A widespread polygenic effect can raise statistics across many variants, while confounding or technical problems can also inflate them. The earlier material on distinguishing confounding from polygenic signal motivates the central diagnostic question: does the inflation track the amount of linkage disequilibrium surrounding a variant, or does it appear more like a common upward shift?

LD Score Regression (LDSC) addresses this question by relating each variant’s association chi-square statistic to its LD score. The basic diagnostic is a fitted line across variants: its slope summarizes the increase in association statistics with LD score, and its intercept summarizes the fitted level at LD score zero. These are diagnostic summaries, not a definitive separation of every possible source of inflation.

In an analysis, the fit is generally weighted rather than an unweighted line through every point. The weights affect how much each variant contributes to the estimated line; uncertainty is also important when judging whether a fitted slope or intercept differs meaningfully from a reference value. Small examples below use ordinary least squares so that the arithmetic is visible. The regression’s diagnostic interpretation requires more than calculating a line.

What the Slope and Intercept Summarize

Under the usual diagnostic interpretation, polygenic effects tend to contribute more to the expected association statistic at variants with larger LD scores: such variants tag, on average, more genetic variation through correlation with other variants. This creates a positive trend of chi-square statistics with LD score. Thus, a positive slope is consistent with polygenic signal. Its numerical value is not, by itself, a portable measure of genetic effect: it depends on such features as sample size, trait architecture, and the scale and construction of the LD scores.

The intercept describes the fitted chi-square level at \(\ell_j=0\), rather than the average statistic among the variants actually observed. In a well-calibrated setting with no excess inflation, the null reference level is approximately 1. An intercept above 1 can indicate inflation that does not follow the LD-score trend, such as effects of population structure, technical artifacts, or residual relatedness. But an intercept is not a direct measurement of all confounding, and its interpretation depends on the model and data quality.

Fitting the Line: A Short Justification

For an unweighted fit to \(m\) variant-level pairs \((\ell_j,\chi_j^2)\), ordinary least squares chooses \(a\) and \(b\) to minimize the sum of squared residuals. Write \(\bar{\ell}\) for the mean LD score and \(\overline{\chi^2}\) for the mean chi-square statistic in this set of pairs. Differentiating the sum of squares with respect to \(a\) and setting the result to zero gives \(a=\overline{\chi^2}-b\bar{\ell}\). Substituting this expression into the derivative with respect to \(b\) gives the slope formula:

The numerator measures how much the two quantities vary together; the denominator measures variation in LD score. If the LD scores barely vary, the denominator is small and the slope is unstable. In practice, a weighted fit changes the estimating equations, but the diagnostic idea remains the same: estimate the trend with LD score and examine the fitted level at zero.

Assumptions, Diagnostics, and Alternatives

The regression is most informative when its central working relationship is a reasonable summary of the data: expected chi-square statistics change approximately linearly with LD score over the analyzed variants. The association statistics should be on a comparable, adequately calibrated scale, and the LD scores should represent the study variants and population well enough for that comparison. The variants included and the weighting scheme should not allow a small number of extreme or high-leverage points to dictate the result. These conditions are not guaranteed by drawing a line.

Inspect the data and fit for warning signs. A scatterplot can reveal curvature, clusters, or individual points far from the trend. A small range of LD scores weakens slope estimation. A negative slope or a large discrepancy between the fitted line and the bulk of the points deserves investigation, not an automatic biological explanation. Compare the intercept and slope with their uncertainty: a point estimate alone does not show how precisely either quantity has been estimated. Also ask whether the reference panel and the association analysis are sufficiently comparable for LD-score differences to be meaningful.

LD Score Regression complements, rather than replaces, other inflation diagnostics. The genomic inflation factor \(\lambda\) summarizes a genome-wide shift in test statistics, while a QQ plot displays the distribution of observed statistics against a null reference. Neither directly models inflation as a function of LD score. LDSC adds that relationship, which can help distinguish an LD-associated trend from a broadly elevated baseline. The diagnostics can disagree because they summarize different features and rely on different assumptions.

A common mistake is to read an intercept above 1 as the exact amount of confounding, or a positive slope as proof that all inflation is polygenic. The correction is to treat both as conditional diagnostics: the slope describes association between chi-square statistics and LD score under the fitted model, while the intercept describes the model’s extrapolated level at zero LD score. Reference mismatch, model misspecification, residual confounding, or other complications can make those summaries misleading. Report them with uncertainty and context, and do not turn association patterns into causal claims.