Do you always need to clump polygenic risk scores?

Question

[kl]

Hi,

I have always clumped my polygenic risk score but I recently I came across a paper that did not perform clumping on polygenic risk scores created at quite liberal thresholds. Does this not worsen the predictive accuracy and bias the effect estimates by over representing regions in high LD?

Answer 1

[Sam]

Depends on the algorithm used. If you follow the traditional p-value thresholding approach, then clumping is necessary to avoid bias by LD as suggested by you. However, if they are using more sophisticated models that already accounts for LD e.g. LDpred, lassosum, PRS-CS etc, then they don't need to perform clumping as these methods already accounted for the whole LD structures.