LD Score Regression Assumptions and Their Consequences

Why Assumptions Matter

LD Score Regression (LDSC) uses variation in SNP LD scores to model variation in association chi-square statistics. Its slope is informative about polygenic signal only when the assumed relationship is a reasonable description of the data. If the mean relationship is misspecified, the fitted slope can absorb patterns caused by confounding or other departures. If the SNP statistics are dependent in ways ignored by the uncertainty calculation, a numerically plausible slope can still have an unreliable standard error.

The key distinction is between assumptions about the conditional mean of each statistic and assumptions about the residual variation around that mean. Mean-model failures can distort the intercept or slope; residual dependence and incorrect variance estimates can make confidence intervals and tests too optimistic or too conservative.

The Mean Model and Its Assumptions

Let \(\chi_j^2\) be the one-degree-of-freedom association chi-square statistic for SNP \(j\), let \(\ell_j\) be its LD score, let \(N\) be the association sample size, \(M\) the number of SNPs in the genetic model, and \(h_g^2\) the SNP heritability on the scale used in that model. A common LDSC mean model is

E(\chi_j^2\mid \ell_j)\approx 1+a+\frac{N h_g^2}{M}\ell_j.

Here \(a\) denotes excess mean inflation not explained by the modeled polygenic component. Thus the model intercept is \(1+a\), and the slope is \(N h_g^2/M\). The equation is an approximation, not a universal identity: it describes expected association statistics under specified genetic and sampling conditions.

Definition: In the LDSC mean model, the intercept is the expected value at LD score zero, \(1+a\), and the slope is the change in expected chi-square statistic per unit increase in LD score, \(N h_g^2/M\). This interpretation requires the conditional mean to be approximately linear in the LD score and the SNP statistics to be adequately calibrated.

The usual slope interpretation relies on a set of linked assumptions. Genetic effects are treated as having mean zero and approximately equal variance across the \(M\) modeled SNPs, rather than being systematically concentrated among SNPs with high or low LD scores. The LD scores used in the regression must represent the relevant linkage disequilibrium pattern. The association statistics must also be on a comparable scale, with sample size and analysis conditions sufficiently consistent that differences in their means are not driven by unmodeled changes in sample size or calibration.

For the simple expectation above, a further requirement is that the average non-polygenic contribution to \(\chi_j^2\) does not vary systematically with \(\ell_j\). If it is roughly constant, it can contribute to the intercept. If it increases or decreases with LD score, it can be mistaken for part of the slope. Finally, standard errors require a suitable account of heteroscedasticity and dependence among SNP statistics; SNPs in LD do not provide independent observations.

Why the Slope Is Linear Under the Genetic Model

A short derivation explains the central genetic assumption. For standardized genotypes, the marginal genetic effect at SNP \(j\) can be written approximately as a sum of effects at causal SNPs, weighted by correlations:

\beta_j^{\mathrm{marg}}\approx\sum_{k=1}^{M} r_{jk}\beta_k.

Suppose the effects \(\beta_k\) have mean zero, are mutually uncorrelated, and each has variance \(h_g^2/M\). Then cross-product terms have expectation zero, so

E\left[(\beta_j^{\mathrm{marg}})^2\right] \approx\sum_{k=1}^{M}r_{jk}^2\frac{h_g^2}{M} =\frac{h_g^2}{M}\ell_j, \qquad \ell_j=\sum_{k=1}^{M}r_{jk}^2.

For a well-calibrated one-degree-of-freedom association statistic, its expected value is approximately one under a null effect, plus the noncentrality contribution. With sample size \(N\), that contribution is approximately \(N E[(\beta_j^{\mathrm{marg}})^2]\). Adding a constant excess component \(a\) gives the stated mean model. The derivation shows why equal effect variance across SNPs matters: if that variance depends on genomic architecture, the weighted sum need not be proportional to the ordinary LD score.

Worked Examples: Reading a Well-Specified Mean Pattern

Worked Example: Recovering the Slope and Intercept

Suppose a hypothetical analysis has \(N=10{,}000\), \(M=10{,}000\), and \(h_g^2=0.020\), with \(a=0\). Under the mean model, the expected chi-square values at LD scores 5, 10, and 20 are

1+\frac{10{,}000(0.020)}{10{,}000}\ell_j =1+0.020\ell_j,

which gives \(1.10\), \(1.20\), and \(1.40\), respectively. The slope between the first two points is \((1.20-1.10)/(10-5)=0.020\); between the last two it is \((1.40-1.20)/(20-10)=0.020\). The intercept is 1.00. In this hypothetical setting the slope equals \(N h_g^2/M=0.020\), and the points are consistent with the assumed linear mean. This agreement does not by itself establish that the assumptions hold in real data.

How Violations Distort the Fitted Relationship

A constant excess component across LD scores shifts the mean upward and is represented by the intercept. The more consequential case for slope interpretation is excess inflation that is correlated with LD score. That pattern can arise if an unmodeled source of association-statistic inflation varies with genomic features that also track LD score. The regression cannot identify its cause from the line alone: it estimates the best-fitting mean trend in the observed statistics.

Worked Example: LD-Correlated Inflation Biases the Slope

Consider hypothetical expected statistics at LD scores 5, 10, and 20. Suppose the true polygenic contribution is \(0.020\ell_j\), but an additional, unmodeled contribution is \(0.010\ell_j\). The resulting means are

1+0.020\ell_j+0.010\ell_j =1+0.030\ell_j,

so they are \(1.15\), \(1.30\), and \(1.60\). The fitted slope is \((1.30-1.15)/(10-5)=0.030\), and the endpoint slope is \((1.60-1.15)/(20-5)=0.030\). The slope exceeds the true polygenic component \(0.020\) by \(0.010\) per LD-score unit. The line still looks perfectly linear and has intercept 1.00: linearity alone cannot rule out this violation. Interpreting the full slope as polygenic signal would therefore be incorrect.

Effect-size architecture can also violate the assumed proportionality. If effect variance differs across SNPs in a way related to LD score, the expected marginal association signal may bend away from a straight line. A straight-line fit then averages over that pattern. The direction and size of slope distortion depend on the distribution of LD scores and on how the departures covary with them; curvature does not have one universal effect.

Worked Example: Curvature and the Fitted Slope

For a hypothetical illustration, suppose the expected means follow \(1+0.020\ell_j+0.0005(\ell_j-12)^2\), rather than a straight line. At LD scores 2, 12, and 32, the means are

\begin{aligned} \ell=2 &: \quad 1+0.020(2)+0.0005(2-12)^2=1+0.04+0.05=1.09,\\ \ell=12 &: \quad 1+0.020(12)+0.0005(12-12)^2=1.24,\\ \ell=32 &: \quad 1+0.020(32)+0.0005(32-12)^2=1+0.64+0.20=1.84. \end{aligned}

The endpoint slope is \((1.84-1.09)/(32-2)=0.75/30=0.025\), not the linear component \(0.020\). For ordinary least squares on all three equally weighted points, the mean LD score is \(46/3\) and the mean response is \(1.39\). The slope is

\frac{\sum_j(\ell_j-\bar{\ell})(y_j-\bar{y})} {\sum_j(\ell_j-\bar{\ell})^2} = \frac{12}{466.667} \approx 0.02571.

Thus the curvature changes the fitted slope in this asymmetric set of LD scores. The discrepancy between the endpoint and all-point slopes also illustrates that the fitted value depends on the weighting and distribution of the SNPs used.

Uncertainty, Diagnostics, and Alternatives

Even if the mean line is reasonable, uncertainty estimates can fail when residuals are treated as independent or have the wrong variance. Chi-square statistics have nonconstant variability, and nearby SNPs share genotype information through LD. A regression standard error that treats every SNP as an independent replicate can consequently understate uncertainty. Block jackknife methods address this dependence by repeatedly omitting genomic blocks and measuring how much the estimate changes across omissions; their validity still depends on having enough suitably chosen blocks and on the genomic blocks capturing the relevant dependence.

Worked Example: Dependence Changes the Uncertainty

Suppose a hypothetical fit gives a slope estimate of \(0.020\), and a regression calculation that treats SNP rows as independent reports \(SE=0.0015\). A block jackknife based on four leave-one-block-out slope estimates gives \(0.018\), \(0.022\), \(0.020\), and \(0.030\). Their mean is \(0.0225\). The jackknife standard error is

SE_{\mathrm{jack}} = \sqrt{\frac{B-1}{B} \sum_{b=1}^{B}(\hat{\theta}_{(b)}-\bar{\theta})^2} = \sqrt{\frac{3}{4} \left[(-0.0045)^2+(-0.0005)^2+(-0.0025)^2+(0.0075)^2\right]} \approx 0.00789.

The jackknife uncertainty is much larger than \(0.0015\), indicating that the independent-row calculation is not a defensible summary of uncertainty in this example. Four blocks are too few for a stable practical estimate, so the numerical comparison is illustrative rather than a recommended analysis. It demonstrates why the standard error depends on the dependence-aware procedure, not just on the fitted line.

Diagnostics should examine both the mean pattern and the uncertainty procedure. Plotting binned mean chi-square statistics against LD score can reveal broad curvature or a changing trend; residual patterns can reveal lack of fit not obvious from a single slope. Large changes in estimates under block deletion indicate sensitivity to particular genomic regions. The plot is a diagnostic, not proof: a confounding pattern proportional to LD score can preserve a straight line, as the preceding example showed.

LDSC is not interchangeable with a single genomic inflation factor or a QQ plot. Those summaries describe overall test-statistic behavior, whereas LDSC uses LD-score variation to separate a linear polygenic component from a constant component under its assumptions. Conversely, LDSC's regression interpretation depends on those assumptions and on appropriate uncertainty estimates. No one summary diagnoses every failure mode; the substantive interpretation should be checked against model fit, data processing, and the study design.

Common Pitfalls and Practical Limits

Calling every slope increase polygenicity. A trend in non-polygenic inflation with LD score can raise the slope. The correction is to treat the slope interpretation as conditional on the mean-model assumptions.
Assuming a straight-looking plot proves the model. A correlated inflation component can itself be linear in LD score. Examine study design and calibration, not only visual fit.
Using a small regression standard error as evidence of precision. SNP statistics are not independent replicates. Use an uncertainty procedure that accounts for genomic dependence and assess sensitivity to blocks.
Ignoring curvature or influential regions. A single slope can conceal changing trends. Inspect binned means and residuals and consider whether a few regions dominate the fit.

These checks cannot establish all assumptions from summary statistics alone. A fit may be stable and nearly linear while its slope still combines multiple sources of signal. LDSC is therefore an informative diagnostic and model-based summary, not an automatic correction for every source of bias.

Check Your Understanding

Answer each question, then compare with the brief solutions.

If the model is \(E(\chi_j^2\mid\ell_j)=1.04+0.015\ell_j\), what are the intercept and slope?
Why can confounding correlated with LD score bias the slope even when the regression line is straight?
For the hypothetical means 1.09 at LD score 2 and 1.49 at LD score 22, calculate the endpoint slope. Does this alone show that a curved pattern has altered the slope?
Why may treating every SNP as an independent regression observation produce an unreliable standard error?

Solutions. (1) The intercept is 1.04 and the slope is 0.015 per LD-score unit. (2) The regression assigns any systematic mean increase with LD score to its slope; it cannot distinguish such a component from polygenic signal using linearity alone. (3) The endpoint slope is \((1.49-1.09)/(22-2)=0.40/20=0.020\). Those endpoints alone do not establish whether the relationship is curved or whether curvature changed the fitted slope; an interior point and the regression weighting matter. (4) SNP statistics are correlated through LD, so the number of SNP rows overstates the effective independent information and can make the standard error too small.