Simulation results of rates of false positives (a = 0.05) under different trait & predictor residualization approaches, compared to a full model, under a null in which there is no gene-gene interaction effect.

Compared models include the approach used in the main analyses, ‘Residualize Expression’, in which the imputed expression and trait values were residualized on all covariates prior to running the test (yresid = T1resid + T2resid + T1resid*T2resid), and the “Residualize Expression and GxG Term”, in which the trait, imputed expression and imputed expression interaction terms were all residualized prior to running the test (yresid = T1resid + T2resid + resid(T1*T2)). “Full Model” includes the observed trait values, the imputed expression and covariate main effects, the T1*T2 and all expression*covariate terms. Simulations used a sample size of 5,000 and 2,000 replicates. Residualization of the trait and the imputed expression never leads to systematically higher rates of false positives than the full model, but residualizing the T1*T2 term separately leads to high rates of false positives when covariates and imputed gene expression values are correlated.
(XLSX)