The "portability problem" has been well established in the literature. Namely, the predictive accuracy of polygenic indices is poor when the target population is of a different continental ancestry (Martin et al. 2017). Moreover, recent work has shown that even within ancestry groups, the PGI predictive accuracy decays as a function of genetic distance (Ding et al. 2023). Meanwhile, other work has shown that even within an ancestral group stratified by non-genetic variables—such as gender or education—there are asymmetries in predictive accuracy decline across groups (Mostafavi et al. 2020). A question remains, however, regarding the nature of the decay. While it is often assumed that genetic background, LD structure and allele frequency differences account for the lion's share of the portability problem (and indeed there is some evidence to this effect from admixed populations (Bitarello and Mathieson 2020), assortative mating, gene-environment interactions, and genetic nurture may all also play a role. To disentangle these drivers of accuracy decay, the present study compares the predictive accuracy across groups of PGIs based on traditional GWAS weights and those based on within-family GWAS weights, which are putatively purged of some of the environmental confounds that may plague comparisons of traditional GWAS-based PGIs across groups.
Dalton Conley, Princeton University