Estimating Genetic Correlation from Summary Statistics

54 concepts50 papers1 proposal

Most complex human traits and diseases are polygenic, meaning they are driven not by a single "risk gene," but by the collective action of thousands of tiny genetic variants scattered across the genome. As reviewed by Robert Maier et al. in Embracing polygenicity: a review of methods and tools for psychiatric genetics research, understanding the shared genetic architecture (or pleiotropy) between these traits can reveal deep biological connections between seemingly unrelated conditions.

Historically, calculating the genetic correlation between two traits required individual-level genotype data from large cohorts, using frameworks like Genome-Wide Complex Trait Analysis (GCTA), as seen in work by Nicole Harlaar et al. (Word Reading Fluency: Role of Genome-Wide Single-Nucleotide Polymorphisms in Developmental Stability and Correlations With Print Exposure). However, sharing raw, individual-level genomic data across research groups faces severe privacy restrictions and massive computational hurdles.

Linkage Disequilibrium Score Regression (LDSC) solves this problem. It allows researchers to accurately estimate heritability and cross-trait genetic correlation using only summary statistics—the publicly available, aggregated results from Genome-Wide Association Studies (GWAS).

The Intuition: Linkage Disequilibrium as a Net

To understand why LDSC works, we have to look at Linkage Disequilibrium (LD), which is the non-random statistical association between neighboring genetic variants (SNPs) due to shared inheritance.

Imagine the genome as a landscape where some SNPs are highly connected to their neighbors (high LD) and others are relatively isolated (low LD). If we calculate an "LD Score" for a specific SNP, we are simply counting how many neighboring SNPs it is correlated with. A high LD score means the SNP sits in a crowded neighborhood; a low LD score means it stands alone.

Now, suppose a trait is highly polygenic, meaning true causal variants are randomly distributed across the genome. If you pick a SNP with a high LD score, it acts like a wide net: because it is correlated with many neighbors, it has a much higher chance of being statistically associated with a true causal variant purely by proximity. Conversely, an isolated SNP with a low LD score will only show an association if it happens to be the causal variant itself.

Consequently, under a polygenic model, the strength of a SNP’s association statistic ($\chi^2$) scales linearly with its LD Score. The more genomic neighbors a SNP tags, the higher its expected association statistic will be.

Separating Signal from Confounding

This linear relationship provides an elegant way to solve a classic problem in genetics: distinguishing true biological signals from confounding factors like population stratification (subtle ancestral differences between cases and controls) or cryptic relatedness.

Confounding factors affect the entire genome uniformly. If a study suffers from population stratification, every SNP's association statistic will be artificially inflated by a flat amount, regardless of whether it sits in a high-LD or low-LD region.

By running a simple linear regression where the independent variable is the pre-computed LD score and the dependent variable is the GWAS $\chi^2$ statistic, we can separate these forces:

Bivariate LDSC: Cross-Trait Correlation

The authors later extended this logic from a single trait to a bivariate framework to quantify the genetic overlap between two distinct phenotypes (Chloe X. Yap et al., Dissection of genetic variation and evidence for pleiotropy in male pattern baldness).

Instead of regressing the $\chi^2$ statistics of one trait, bivariate LDSC multiplies the signed association statistics ($Z$-scores) of Trait 1 and Trait 2 for each SNP. If the two traits share a genetic basis, variants that tag dense genomic regions will capture more shared causal variants, causing the product of their $Z$-scores to scale linearly with their LD scores. The slope of this regression line directly quantifies the genetic covariance between the two traits.

Single-Trait LDSC:  Expected [ χ² ]     = (Sample Size * Heritability * LD Score) + Confounding Intercept
Bivariate LDSC:     Expected [ Z₁ * Z₂ ] = (Sample Size Factor * Covariance * LD Score) + Overlap Intercept

This bivariate approach elegantly neutralizes another major real-world issue: sample overlap. If the GWAS for Trait 1 and the GWAS for Trait 2 share some of the same participants, the environmental noise between the two studies will be correlated. In traditional analyses, this overlap creates severe mathematical bias. In bivariate LDSC, however, the correlation due to sample overlap is completely absorbed by the regression intercept, as noted by Jiyuan An et al. (Gastroesophageal reflux GWAS identifies risk loci that also associate with subsequent severe esophageal diseases). The slope remains an unbiased measure of true genetic correlation.

Methodological Trade-offs

While LDSC is highly robust, it operates within a broader ecosystem of summary-statistic methods. As evaluated by Yiliang Zhang et al. in Comparison of methods for estimating genetic correlation between complex traits using GWAS summary statistics, newer likelihood-based frameworks like High-Definition Likelihood (HDL) can achieve greater statistical precision by utilizing full linkage disequilibrium matrices rather than summary scores.

However, these advanced methods are highly sensitive to mismatches between the study population and the reference panels. Because LDSC relies on a straightforward linear regression, it remains a widely used standard in genomic research—providing a mathematically clean, computationally fast, and exceptionally resilient strategy for mapping the shared blueprint of human complex traits.

Key Foundational Work

Treating phenotype as given: a simple resampling method for genome-wide association studies (2011) — contributed 1 concept elements to this direction.
From phenotype to genotype: an association study of longitudinal phenotypic markers to Alzheimer's disease relevant SNPs (2012) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_linkage_disequilibrium_score_regression_289f1c87,
  title = {linkage disequilibrium score regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linkage-disequilibrium-score-regression/rdr_289f1c87fe2b4504b68d752cbb195cba},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_linkage_disequilibrium_score_regression_289f1c87_v1,
  title = {linkage disequilibrium score regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linkage-disequilibrium-score-regression/rdr_289f1c87fe2b4504b68d752cbb195cba/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake