Editing Openai/69177fbb-14c4-800f-bc57-ccb5a1d88c95 (section)

=== - Population stratification and confounding. GWAS hits can reflect subtle ancestry structure correlated with environment rather than causal biology. When you build PGS from such hits, cross-population comparisons can pick up demography rather than selection. Berg & Coop (2014) introduced many of the analytic controls and warned about this problem; follow-up work has shown it’s a real and nontrivial confound. PMC<ref>{{cite web|title=PMC|url=https://pmc.ncbi.nlm.nih.gov/articles/PMC4125079/|publisher=pmc.ncbi.nlm.nih.gov|access-date=2025-11-17}}</ref> ===
* LD and ascertainment biases. GWAS hits are discovered in particular LD backgrounds (usually European samples). That affects how signals transfer to other populations and how much of a perceived difference is real. Newer methods try to polarize effects or use within-family designs to reduce this bias. OUP Academic<ref>{{cite web|title=OUP Academic|url=https://academic.oup.com/evlett/article/3/1/69/6697303|publisher=OUP Academic|access-date=2025-11-17}}</ref>
* Phenotype proxy limitations. “Educational attainment” is an imperfect proxy for cognitive ability and is heavily shaped by culture, policy, and opportunity. Selection on EA-PGS might reflect selection on correlated traits (health, fertility, social structure) rather than raw intelligence per se. PMC<ref>{{cite web|title=PMC|url=https://pmc.ncbi.nlm.nih.gov/articles/PMC9582364/|publisher=pmc.ncbi.nlm.nih.gov|access-date=2025-11-17}}</ref>