Interpreting the LD Score Regression Slope

Why the Slope Needs Context

A positive association between a SNP's LD score and its association test statistic is evidence of a systematic pattern: SNPs tagging more genetic variation tend, on average, to have larger test statistics. LD Score Regression summarizes that pattern with a slope. But a slope is not a stand-alone measure of how much heritability a trait has. Its expected size depends on both the trait's SNP heritability and the study's sample size.

That dependence matters when interpreting a result or comparing studies. A larger study can have a steeper slope than a smaller study even when both have the same underlying SNP heritability. Conversely, comparing raw slopes across studies without accounting for sample size can suggest a difference in polygenic signal where there is none.

The Mean Model and the Meaning of the Slope

For SNP \(j\), let \(\chi_j^2\) be its association test statistic and \(l_j\) its LD score, the sum of squared correlations between that SNP and the variants in the reference set. A usual LD Score Regression mean model is

Here \(a\) is the regression intercept, \(b\) is the slope, \(N\) is the study's sample size (or the appropriate effective sample size), \(h_g^2\) is SNP heritability on the scale used for the analysis, and \(M\) is the number of variants defining the genetic-effect model. In the standard interpretation, the slope reflects the component of expected test-statistic elevation that increases with LD score. The intercept represents the model's expected value at LD score zero; it is not the slope's substitute.

The units help make the interpretation concrete: the response is a chi-square statistic and the predictor is LD score, so the slope is measured in chi-square units per LD-score unit. To express the slope on a heritability scale, rearrange the model:

This conversion is meaningful only when the \(M\), \(N\), phenotype scale, variant set, and model conventions are appropriate to the analysis. It does not turn every estimated slope into a reliable heritability estimate: uncertainty, model fit, and data quality still matter.

Why the Slope Scales With Heritability and Sample Size

A brief derivation shows the source of the scaling. Suppose genotypes and the phenotype are standardized, and the genetic effects across \(M\) variants are modeled as mutually uncorrelated, mean-zero effects, each with variance \(h_g^2/M\). If \(r_{jk}\) is the correlation between SNPs \(j\) and \(k\), the marginal effect at SNP \(j\) is approximately the sum of the effects at correlated variants, weighted by those correlations. Because the effects are mutually uncorrelated, the cross-covariance terms in this squared sum have expectation zero. Averaging its squared value over the random effects gives

For a sample of size \(N\), the association statistic's noncentrality parameter is approximately \(N\beta_j^2\) under this standardization. A one-degree-of-freedom noncentral chi-square variable has expectation \(1+\lambda_j\), where \(\lambda_j\) is its noncentrality parameter. Thus the genetic contribution to the expected chi-square is approximately

Adding the intercept term gives the stated mean model and identifies the slope as approximately \(N h_g^2/M\). This is an expectation under a model, not an identity for every SNP. In practice, LD Score Regression fits a weighted regression to noisy summary statistics, and the estimate is affected by sampling variability and how well the model describes the data.

The derivation relies on a consistent variant set and LD-score definition, an appropriate sample-size input, and the assumed relationship between effect variance and LD tagging. It also presumes that the test statistics and phenotype scale correspond to the model being used. These are conditions for interpreting the slope, not guarantees that a particular fitted regression is reliable.

Comparing Slopes Across Sample Sizes

If two studies have the same underlying \(h_g^2\) and use the same \(M\), the model predicts slopes proportional to sample size. Thus raw slopes should not be compared as if they were sample-size-free measures. Scaling the slope to a common sample size is one way to make the model's prediction visible; converting each slope to heritability is another.

Reading the Fitted Line in Context

The slope describes how the fitted mean chi-square changes with LD score, not the statistic for a particular SNP. For example, a slope of \(0.040\) means that the fitted mean chi-square increases by \(0.040\) for each one-unit increase in LD score. It does not mean that every SNP with a one-unit higher LD score has a statistic exactly \(0.040\) larger.

Uncertainty, Diagnostics, and Failure Modes

A slope estimate should be read with its standard error or confidence interval. A positive point estimate may be imprecise, and a small or negative estimate can occur through sampling variation even when the underlying genetic component is nonnegative. A negative fitted slope is not evidence for negative SNP heritability; it is a warning that the data do not support a stable positive slope under this fitted model.

Useful checks include examining the distribution of LD scores and whether the fitted relationship is driven by a small number of high-LD-score observations; reviewing standard errors and the number of SNPs contributing to the regression; and confirming that sample size, variant count, and LD-score conventions match the intended interpretation. Large scatter around the fitted line, influential points, or a slope that changes substantially under reasonable analysis choices can make a point estimate difficult to interpret.

The slope is also not the same quantity as the genomic inflation factor \(\lambda\), which summarizes a distribution of test statistics, nor is it the intercept. In the LD Score Regression mean model, the slope captures the part of expected chi-square that varies with LD score. The intercept is a separate model component. Neither a slope nor an intercept alone diagnoses every possible source of bias.

Comparison and Limits of Interpretation

The slope is useful precisely because it uses the relationship between association statistics and LD scores rather than only the overall level of test statistics. However, its interpretation remains model-dependent. Differences in variant sets, LD-score reference data, phenotype scaling, or the sample-size quantity used can undermine a numerical comparison. In case-control or meta-analysis settings, use the effective sample-size convention appropriate to the analysis rather than assuming that a simple participant count is interchangeable.

A common mistake is to say, “Study A has the larger slope, so it must have greater SNP heritability.” That conclusion ignores \(N\): since \(b\approx Nh_g^2/M\), a larger sample can produce a larger slope at the same heritability. The correction is to compare slopes on a common sample-size scale or convert them using \(M/N\), while checking that the variant and model conventions match. Another mistake is to interpret a fitted slope as a causal effect per SNP; it is an aggregate regression summary of expected test-statistic patterns, not an estimate of an individual variant's causal effect.