Preprint warns LDSC heritability estimates can be biased by meta-analysis and imputed variant inputs

A preprint posted to bioRxiv presents a systematic investigation of bias in linkage disequilibrium score (LDSC) regression, one of the most widely used methods for estimating SNP heritability from genome-wide association study (GWAS) summary statistics. LDSC regression is popular partly because it does not require access to individual-level genotype data, making it practical for large collaborative studies. However, the authors demonstrate that the method can produce substantially biased heritability estimates under two common conditions: when summary statistics derive from meta-analyses across multiple studies, and when they include imputed rather than directly genotyped variants.

The investigation used summary statistics from several large-scale Alzheimer's disease GWAS as a test case, comparing LDSC estimates against results from methods that use individual-level data. The authors also examined the sensitivity of estimates to the choice of LD reference panel, finding that mismatches between the reference panel and the target study population can compound bias further.

The findings are relevant to the statistical genetics community because LDSC-derived heritability estimates are routinely used to benchmark traits, compute genetic correlations, and assess the contribution of specific genomic annotations — applications that downstream analyses, including partitioned heritability and cross-trait analyses, depend upon. The preprint does not invalidate LDSC regression as a tool but identifies conditions under which its outputs warrant additional scrutiny. The work has not yet been peer-reviewed.