To correct for confounding effects on we regressed out the effect of several potential confounders from , using step-wise multiple linear regression analysis [55]. The method begins by regressing out the effect of a variable with high correlation with the gene score; it then adds the next significant variable, and evaluates whether the added variable should be kept and whether any existing variables should be eliminated from the regression model. The latter step is repeated until all variables are considered. A variable was added at p<0.05 and removed at p>0.1. The step-wise nature of this method should account for correlations between the variables. We initially tested this model using 1,000 DGI GWA permutations and six gene properties as potential confounders (predictor variables). In this case, step-wise multivariate linear regression was applied to using the full list of genes, and the coefficients α, β, δ, γ, η, and were estimated such that for every gene g one can calculate:(2)where is the residual of the association score for gene g that cannot be explained by the effects of the predictor variables considered.
