Interpreting the LD Score Regression Intercept

What the Intercept Can Tell You

A genome-wide association scan can produce test statistics that are elevated for more than one reason. Many genuine genetic effects can raise the overall distribution of association statistics, while population structure, cryptic relatedness, technical artifacts, or other sources of test-statistic inflation can also raise it. The earlier material on polygenicity and on distinguishing confounding from polygenic signal motivates a diagnostic question: after allowing the mean statistic to vary with LD score, is there evidence of an additional elevation that does not follow that pattern?

LD Score Regression (LDSC) addresses this question by fitting a mean model to SNP association statistics and their LD scores. The intercept is the fitted mean statistic at an LD score of zero. Under the model’s assumptions, a value near one is consistent with no additional inflation represented by the intercept; a value above one is evidence of excess inflation not explained by the model’s LD-score-dependent component. That is a diagnostic interpretation, not a verdict about the source of the excess.

Here, \(b\) is included to describe the fitted mean model; the focus is interpreting \(c\), not giving \(b\) a biological interpretation. The distinction depends on the model being an adequate description of the mean relationship and on the assumptions that allow LD-score-dependent polygenic signal to be separated from other inflation. If these conditions fail, the fitted intercept may not isolate the intended component.

The phrase “not explained by the LD-score-dependent component” is deliberately narrower than “confounding.” In the standard diagnostic use, excess intercept is often treated as evidence consistent with inflation from confounding or other non-polygenic sources. But the number does not identify a particular mechanism, prove that a confounder is present, or measure its causal effect. Other departures from the model can also affect the estimate.

Reading an Estimate and Its Uncertainty

An LDSC analysis reports an estimated intercept, usually alongside a standard error. A point estimate above one indicates that the fitted intercept is elevated relative to the null baseline. Whether the data provide clear evidence of such an elevation also depends on uncertainty. A confidence interval that includes one does not establish that the intercept is exactly one; it indicates that the estimate is not clearly separated from one at that confidence level.

For an approximate Wald interval, suppose the intercept estimate \(\hat c\) has standard error \(SE(\hat c)\), and its sampling distribution is approximately normal. A 95% confidence interval is

To assess evidence against the reference value \(c=1\), the corresponding Wald statistic is

This comparison tests whether the fitted intercept differs from the null baseline under the approximation used for the standard error. It does not test whether any named confounding mechanism is present. A small p-value for this comparison is evidence that the intercept differs from one under the model; it is not the probability that the null is true.

Why the Intercept Is an Extrapolated Diagnostic

The definition makes the intercept sound like an observed average for SNPs with LD score zero. In practice, it is a regression estimate based on the collection of SNPs in the analysis. If the fitted mean relationship is \(E[\chi_j^2]=c+b\ell_j\), then the fitted value at a given \(\ell_j\) is \(c+b\ell_j\); setting \(\ell_j=0\) gives \(c\). Thus the intercept is generally an extrapolated model quantity, not a direct average over a group of SNPs with zero LD score.

A short algebraic justification clarifies what is being estimated. In weighted least squares, let \(w_j\) be the weight for SNP \(j\), and let \(\bar{\ell}_w\) and \(\bar{y}_w\) be the weighted means of \(\ell_j\) and the observed statistic \(y_j=\chi_j^2\). Once the regression slope coefficient is estimated as \(\hat b\), the fitted intercept is

This follows from the weighted least-squares intercept condition: the weighted residuals sum to zero, so the weighted mean of the observed statistics equals the weighted mean of the fitted values, \(\bar{y}_w=\hat c+\hat b\bar{\ell}_w\). Rearranging gives the expression above. It shows that the intercept depends on the fitted relationship across the data, including the estimated coefficient and the LD-score distribution; it is not calculated by simply counting “confounded” SNPs.

Because the intercept and the LD-score-dependent term are estimated together, their estimates can be statistically related. A narrow range of LD scores, unusual weighting, or a poor fit can make the extrapolated intercept less stable or less informative. The reported standard error conveys sampling uncertainty under the fitted procedure, but it cannot by itself diagnose every kind of model misspecification.

Diagnostics, Alternatives, and Limits

An intercept should be interpreted alongside the rest of the diagnostic evidence, rather than as a stand-alone correction or a universal measure of study quality. Useful checks include whether the regression gives a plausible description of the mean statistics across LD scores, whether the estimate is precise, and whether the data and analysis meet the assumptions needed for the LDSC interpretation. A large intercept with a wide confidence interval is not the same evidential result as a large, precisely estimated intercept. Nor does a precise estimate guarantee that the model is correctly specified.

Two common complementary diagnostics answer different questions. The genomic inflation factor \(\lambda\) summarizes the overall shift in association test statistics relative to a null reference. It does not itself separate polygenic signal from other causes of inflation. A QQ plot displays the distribution of observed statistics against the expected null distribution; its shape can help reveal departures, but it does not assign those departures to a cause. LDSC adds an LD-score-based mean model and yields an intercept for a particular component of that fitted model. None of these summaries, on its own, proves the source of an observed pattern.

Interpretation is especially limited when the mean model or its assumptions are inadequate. For example, inflation that is related to LD score in a way the model does not represent may not be cleanly assigned to the intended components. Mismatch between the analysis data and relevant LD information can also undermine the diagnostic interpretation. Details of those assumptions and their consequences require careful attention; an intercept is not a substitute for checking them.

The intercept is also not a general-purpose adjustment to apply automatically to all reported association statistics. It is an estimated diagnostic quantity. A value above one does not mean that every SNP statistic is inflated by exactly the same amount, and subtracting the excess from every statistic is not justified by the definition alone. Conclusions should state the estimate, its uncertainty, the model-based interpretation, and the limits on attributing a cause.

Common Mistake