Xuanjun Fang

Author    Correspondence author
Bioscience Methods, 2026, Vol. 17, No.   
Received: 01 Jan., 1970    Accepted: 01 Jan., 1970    Published: 15 May, 2026

SNP-based heritability is widely interpreted as a fundamental property of complex traits, yet estimates vary substantially across methods. Here we show that this variation arises because different approaches do not estimate the same quantity: SNP-based heritability is a model-defined estimand rather than a single biological parameter. Using UK Biobank height data as a representative case, we systematically compare estimates from individual-level methods (GCTA–GREML and related estimators) and summary-statistics-based approaches (LD Score Regression and SumHer). We find that GREML-based methods consistently yield higher estimates (~0.60–0.69), LDSC produces systematically lower values (~0.56), and SumHer yields intermediate or higher estimates (~0.63). These differences persist under matched samples and SNP sets, indicating that they cannot be attributed to sampling variation alone. We demonstrate that the discrepancies arise from differences in data representation, model assumptions, and the treatment of linkage disequilibrium (LD) and allele frequency. Accordingly, each method targets a distinct estimand: GREML captures variance explained through genomic relationships, LDSC estimates LD-weighted marginal effects, and SumHer models MAF- and LD-dependent architectures. This framework resolves apparent inconsistencies in SNP heritability estimates and clarifies that cross-method comparisons are generally not statistically valid without alignment of underlying assumptions. More broadly, our results redefine SNP-based heritability as a model-dependent functional determined by SNP coverage, LD structure, and estimation framework. These findings provide a principled basis for interpreting heritability estimates and have implications for genetic studies ranging from biobank-scale analyses to genomic prediction.