In UKB3, because of the large sample size, we faced computational constraints. We therefore used the REML implementation in BOLT v2.334 (with the --remlNoRefine option). Moreover, we estimated standard errors only for three phenotypes: friend satisfaction, educational attainment, and height. We chose these three phenotypes so as to have one each corresponding to a single-trait PGI with low (friend satisfaction), medium (educational attainment) and high predictive power (height).
