Warning, the concept of the heritability model is quite complicated; it took me about nine years to understand!
All heritability analyses begin by specifying a heritability model, which describes how much each predictor is expected to influence the phenotype. We have shown that for many analyses, the choice of heritability model is very important. For example, changing the heritability model can lead to very different estimates of SNP heritability and heritability enrichments, and result in more accurate polygenic scores.
The first heritability analyses used the Uniform Model, which assumes that all predictor are expected to explain equal heritability (we used to refer to this as the GCTA Model, because of its use in the software GCTA). Since then, many different models have been developed, that try to more accurately describe how heritability varies across the genome. For example, we proposed the LDAK Model, where the expected heritability contributed by a predictor depends on its minor allele frequency (MAF) and local levels of linkage disequilibrium (LD).
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
In our paper Evaluating and improving heritability models using summary statistics (Nature Genetics, 2020), we developed a method for measuring how well heritability models fit real data. We used this method to compare 12 different heritability models across a wide range of complex traits. When we restricted to simple heritability models, the best-performing was the LDAK-Thin Model, which is an improved version of the LDAK Model. When we considered all models, the BLD-LDAK Model performed best, a model where the expected heritability contributed by a predictor depends on its MAF, LD and functional annotations.
Recommendation:
We now generally recommend using either the Human Default Model or the Alpha Model. Both these models allow the expected contribution of a predictor to depend on its MAF (the difference is that the Human Default Model specifies the relationship between per-predictor heritabilities and MAF, whereas the Alpha Model estimates it from the data).
In particular, we recommend using the Human Default Model when performing REML, Haseman Elston Regression or PCGC, when running LDAK-GBAT and when estimating SNP Heritability and Genetic Correlations. Meanwhile, we recommend using the Alpha Model when running LDAK-KVIK and MegaPRS. Although there exist more realistic heritability models (e.g., those that take into account functional annotations), we recommend these two models because they have robust performance and are easy to use.
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
Technical Details describes how to implement different models, while Comparing Models explains how you can use LDAK to test different heritability models.
