by LD-pruning—either without regard to P-values75, or via informed LD-pruning76 (clumping) that preferentially retains markers with more significant P-values. More recent work has shown that explicitly modeling LD using an LD reference panel and estimating posterior mean causal effect sizes can improve prediction accuracy from summary statistics77. An alternative to summary statistic based methods is to fit effect sizes of all markers simultaneously using Best Linear Unbiased Prediction (BLUP) methods and their extensions78–80, which require individual-level training data. Fitting all markers simultaneously is theoretically more appropriate and can produce more accurate predictions, although the relative advantage is small when overall prediction accuracies are modest (Box 3). In their simplest form, polygenic risk scores and BLUP methods assume infinitesimal (Gaussian) architectures in which all markers are causal, but these methods have been extended to increase prediction accuracy in the case of non-infinitesimal architectures; this has been accomplished for polygenic risk scores via restricting to markers below a P-value threshold75 or estimating posterior mean causal effect sizes under a point-normal prior77, and for BLUP methods by estimating (joint-fit) posterior mean causal effect sizes under a normal mixture prior81,82. Although polygenic risk scores must await even larger training sample sizes to attain clinical
