Abstract For genetic association analysis based on multiple SNP regression of genotypes obtained by dense DNA sequencing or array data imputation, multi-collinearity can be a severe issue causing failure to fit the regression model. In this study, we proposed a method of Dimension Reduction using Local Principal Components (DRLPC) which aims to resolve multi-collinearity by removing SNPs under the assumption that the remaining SNPs can capture the effect of a removed SNP due to high linear dependency. This approach to dimension reduction is expected to improve the power of regression-based statistical tests. We apply DRLPC to chromosome 22 SNPs of two data sets, the 1000 Genomes Project (phase 3) and Canadian Longitudinal Study on Aging (CLSA), and calculated Variance Inflation Factors (VIF) in various SNP-sets before and after implementing DRLPC as a metric of collinearity. Notably, DRLPC addresses multi-collinearity by excluding variables with a VIF exceeding a predetermined threshold (VIF=20), thereby improving applicability for subsequent regression analyses. The number of variables in a final set for regression analysis is reduced to around 20% on average for larger-sized genes, whereas for smaller ones, the proportion is around 48%; suggesting that DRLPC is more effective for larger genes. We also compare the power of several multi-SNP statistics constructed for gene-specific analysis to evaluate power gains achieved by DRLPC. In simulation studies based on 100 genes with ≤500 SNPs per gene, DRLPC effectively increased the power of the multiple regression Wald test from 60% to around 80%.

Dimension Reduction using Local Principal Components for Regression-based Multi-SNP Analysis in 1000 Genomes and the Canadian Longitudinal Study on Aging (CLSA)