Recomputing LoA(LD+MAF) with fGWAS PGIs gives interesting picture: downward shift for most phenotypes in AFR, mixed in SAS, & upward for many in EAS. fGWAS-based PGIs can alter the relative contribution of LD and MAF to prediction loss, with direction and magnitude varying by phenotype and ancestry.

Observed portability is broadly similar for standard vs family-based GWAS (fGWAS) PGIs across traits/ancestries, with notable gains for BMI in AFR—consistent with population-specific confounds affecting select traits rather than a uniform shift.

Because h2 was imprecisely estimated in SAS/EAS, we evaluate LoA(LD+MAF+ h2) in AFR. Adding h2 closes most remaining gap for height and increases explained loss for BMI.

We estimate the share of observed loss in PGI relative accuracy (LoA) explained by cross ancestry MAF and LD differences. LoA(LD+MAF) is highest in AFR, lower in EAS/SAS, and varies by category: blood biomarkers highest; fertility/sexual development & substance use lower.

Biologically proximal traits (blood lipids, diabetes, height) are more portable than behavioral/social traits (e.g., educational attainment, neuroticism, risk tolerance), likely reflecting stronger environmental influences interacting with genetics for the latter.

We dissect why EUR-trained PGIs lose accuracy in non-EURs across 54 traits. We build on model by Wang et al (2020), but use genome-wide, LD-adjusted PGIs, model ancestry-specific h2, and compare standard vs family-based GWAS. Average portability is lowest in AFR (~24%), then EAS (~37%), SAS (~51%).

PGI predictive accuracy is substantially smaller when weights obtained from one ancestry are applied to another. The first step in addressing this portability problem is to understand how different factors influence the loss of PGI predictive accuracy across ancestries.