Conditional association and imputation using summary statistics critically rely on accurate LD information from a population reference panel. Even in the best case where the reference population closely matches the GWAS population, the relatively small size of reference panels for which LD information is publicly available (typically hundreds or at most thousands of individuals) makes accurate estimation of a large number of LD parameters a challenge. This motivates regularization of the estimated LD matrix, both to maximize accuracy and to ensure robustness in the case of imputation using summary statistics, as mis-estimation of the variance of imputed statistics can lead to false-positive associations. A simple approach to regularization is to set all correlations between distal SNPs to zero, based on a fixed distance threshold7 or approximately independent LD blocks inferred from the data23. An alternative is to specify a prior distribution and compute Bayesian posteriors12; data can be combined across multiple ancestry reference panels to further boost accuracy17,18. Singular value decomposition based approaches for LD regularization have also been proposed in other contexts10. In general, the accuracy of conditional association
