Linkage disequilibrium score imputation (LDimp) of genome wide association study summary statistics for multivariate analyses

Introduction: Multivariate genome wide association study (GWAS) analyses, such as genetic correlations and genetic structural equation modelling (e.g. Genomic SEM), are restricted by the number of genetic variants that are in common between the summary statistics included. This means that downstream multivariate analyses examine dissimilar subsets of genetic variants per pair of traits for genetic correlations. We sought to develop an easy to use and quick method to harmonise summary statistics based on imputing missing variants using an LD reference calculated from a whole genome sequenced population panel.

Methods: We validated our simple GWAS summary statistics imputation method Linkage Disequilibrium score IMPutation of GWAS summary statistics (LDimp). Additionally, we developed and applied an enhanced implementation of LD score regression (LDSC++) which accounts for 1.) the number of variants included per trait combination, 2.) association test power, 3.) correlation/attenuation bias, 4.) GWAS summary statistics imputation quality, and 5.) effect direction for bivariate trait covariance.

Results: Augmentation by LDimp retained variants in multivariable and multivariate analyses, and our results suggested that imputed effects are unbiased and yield modest predictive performance. We also found evidence of improvements in LD score regression estimated genetic covariance precision using GWAS summary statistics imputation in LDSC++.

Conclusion: LDimp may benefit multivariable and multivariate analyses, such as Genomic SEM. Improved accuracy and validity of genetic covariance estimates are however conditioned on improvements to the LD score regression method in LDSC++, and any method’s capability to account for noise incurred by GWAS summary statistics imputation.