Within-Family GWAS does not Ameliorate the Decline in Prediction Accuracy across Populations

As polygenic prediction extends beyond the research domain to involve clinical applications, the urgency of solving the “portability problem” becomes amplified—that is, the fact that polygenic indices (PGI) constructed based on discovery analysis in one population (typically of exclusively continental European descent) predict poorly in other populations. In the present paper we test whether population differences in genetic nurture, assortative mating, or population stratification contribute to the fact that polygenic indices constructed based on GWAS results from European-descent samples predict more poorly in admixed populations with Native American and African ancestry. We do this by comparing the rates of decline in prediction accuracy of classical-GWAS-based PGIs versus within-family-based PGIs, each estimated in a population of European descent, as they are deployed in two samples of Latino Americans and African Americans. Within-family GWAS putatively eliminates the effects of parental genetic nurture, assortative mating, and population stratification; thus, we can determine whether without those confounding factors in the PGI construction, the relative prediction accuracy in the out-groups is ameliorated. Results show that relative prediction accuracy is not improved, suggesting that the differences across groups can be almost entirely explained by variation in genetic architecture (i.e. allele frequencies and short-range LD) rather than the aforementioned factors. Additional analysis of the impact of genetic architecture on the decline in prediction accuracy supports this conclusion. Future researchers should test within-family analysis at the prediction rather than the discovery stage.