Does LDSC require individual-level GWAS records?

Standard applications can use GWAS summary statistics plus suitable LD information.

Is the intercept simply the fraction of false associations?

No. Its relationship to inflation is model-based and differs from a false-discovery rate.

Does genetic correlation imply one trait causes another?

No. Shared genetic architecture can arise through several mechanisms and does not settle causal direction.

AI Glossary / Statistics and probability / Advanced

Linkage disequilibrium score regression

Definition

Linkage disequilibrium score regression, usually abbreviated LD Score regression or LDSC, is a method that relates genome-wide association statistics to measures of how much genetic variation each variant tags through correlation with other variants. It can help distinguish widespread polygenic signal from certain sources of inflated test statistics and estimate SNP heritability under stated assumptions. Extensions estimate genetic correlations or partition heritability. It analyzes population-level statistical patterns, not an individual's genetic destiny or a direct causal effect of a particular variant.

What it means

Linkage disequilibrium means that alleles at different genomic locations are statistically associated. A variant can therefore carry information about other variants. Its LD score summarizes squared correlations with variants in a specified reference set or window, with estimation details defined by the method.

If many variants contribute small effects to a trait, variants tagging more relevant variation tend to have larger association statistics on average. LDSC uses that relationship. Its interpretation depends on the genetic architecture, reference population, quality of summary statistics, and how confounding relates to LD.

How it works

In a basic single-trait formulation, the expected association chi-square statistic at variant j is approximately 1 + N a + (N h2/M) lj. Here N is sample size, M is the relevant variant count, lj is the LD score, h2 is the specified SNP-heritability quantity, and a represents confounding in the model.

The fitted slope can be rescaled to estimate h2. The intercept above one estimates a contribution to inflation under the model's assumptions. Real implementations use weighting, variant filtering, and block-based uncertainty estimation because variants are correlated and test-statistic variability is not constant.

Analysts harmonize summary statistics with suitable reference scores, check sample-size and allele information, and inspect uncertainty. Constraining the intercept or converting binary-trait estimates to a liability scale requires additional justification and inputs.

Three practical examples

A research team examines whether elevated association statistics are consistent with many small genetic contributions rather than treating all inflation as confounding.
Researchers estimate SNP heritability from suitable GWAS summary statistics when individual-level records are not available for that analysis.
A cross-trait study uses an LDSC extension to estimate shared genetic architecture, explicitly distinguishing genetic correlation from a claim that one trait causes another.

Think of it this way

Think of microphones that each pick up different amounts of a surrounding sound field. A microphone covering more sources may record more signal even without being noisier. Comparing recorded intensity with coverage can help separate components under a model. In LDSC, however, the coverage is a genetic correlation structure, and an incorrect coverage map can distort the separation.

Worked example

Construct a simplified teaching relationship using N = 100,000, M = 1,000,000, SNP heritability h2 = 0.30, and intercept 1.10. The slope implied by N h2/M is 0.03.

At LD scores 10, 20, and 40, the expected chi-square values are 1.10 + 0.03 x 10 = 1.40, 1.70, and 2.30. Fitting a line to these three constructed expectations recovers slope 0.03 and intercept 1.10.

Rescaling the slope gives h2 = 0.03 x 1,000,000/100,000 = 0.30. The intercept excess is 0.10. Across these three equally weighted teaching points, mean chi-square is 1.80, so the excess attributed to the intercept is 0.10/(1.80 - 1) = 0.125, or 12.5% of that mean excess under this construction.

This is not a valid GWAS analysis using three variants. It only checks the model's arithmetic. Real SNP statistics fluctuate, sample sizes can vary, weights matter, and uncertainty must account for dependence. The value 0.30 is a population variance proportion under the specified model, not 30% of a particular person's trait.

Illustrative case study

Fictional case: Arcadia Genomics reviews a published summary-statistics dataset before comparing it with an AI-derived phenotype. An analyst initially selects a convenient LD reference panel without checking population compatibility.

The methods reviewer asks for reference-panel justification, variant harmonization checks, and clear definitions of the phenotype and heritability scale. The team revises the analysis plan and documents any population mismatch that cannot be resolved.

It reports population-level estimates with uncertainty and avoids converting them into individual recommendations. A shared genetic signal is treated as evidence about statistical architecture, with further work needed to understand mechanisms or intervention effects.

Common misconceptions and mistakes

Do not equate SNP heritability with all inherited influences or assume it is fixed across environments and populations. Do not interpret genetic correlation as causation or as phenotypic correlation.

An intercept near one does not certify the absence of every bias. Reference mismatch, incorrect summary-statistic processing, and violations connecting LD to confounding can affect the result. A large estimated heritability with substantial uncertainty should not be treated as precise. Different variant sets and observed versus liability scales must not be silently compared as identical quantities.

Business use

For organizations evaluating genomic analytics, LDSC can help assess research claims and compatibility of summary datasets. It supports methodological scrutiny rather than direct product personalization by itself.

Ask for the analysis population, reference panel, variant filters, scale, uncertainty, and software settings. An AI-generated literature summary should preserve these qualifications. A population-level estimate cannot alone establish clinical usefulness, individual predictive accuracy, or the benefit of a proposed intervention; those are separate evaluation questions.

Check your understanding

With N = 100,000, M = 1,000,000, and slope 0.03, what does the basic model imply for h2?

Answer: It implies 0.30 on the specified scale. That arithmetic does not validate reference matching, model assumptions, or the precision of an estimate from actual data.

Related terms

linkage disequilibrium
GWAS
SNP heritability
polygenicity
population stratification
genetic correlation
block jackknife

Learn next

Next, study GWAS summary-statistic harmonization and reference-panel selection. Then compare single-trait, cross-trait, and stratified LDSC, keeping their estimands and assumptions separate.

Sources

Sources: Bulik-Sullivan and colleagues, LD Score Regression Distinguishes Confounding from Polygenicity in Genome-Wide Association Studies (2015), primary paper: https://researchmgt.monash.edu/ws/portalfiles/portal/296284768/41408712_oa.pdf . Authors' LDSC documentation: https://github.com/bulik/ldsc/wiki/Heritability-and-Genetic-Correlation . Primary formula, reference-population discussion and implementation guidance were consulted. The three-point illustration is original and independently calculated, not an empirical genetic result.

Previous: Linkage constraints Next: Linked data