Given our design, we ran stratified analyses with our Australian sample, i.e., testing for effects of ND after controlling for AD (AD+ND+ versus AD+ND− and AD−ND+ versus AD−ND−) and testing for the effects of AD (AD+ND+ versus AD−ND+ and AD+ND− versus AD−ND−), in addition to testing for comorbid effects (AD+ND+ versus AD−ND−). For each pair of pools (e.g., AD+/ND+ versus AD+/ND− QIMR arrays, AD+/ND+ versus AD+/ND− deCODE arrays, AD−/ND+ versus AD−/ND− QIMR arrays, etc.) run separately, a linear model based approach was used to test for allelic association in which the response variable is the set of pooling allele frequency (PAF) estimates for each SNP. The predictor variable is case/control status. The linear model was used to estimate the pooling error across all SNPs. A test statistic which corrects for the pooling error was used to rank SNPs, with p-values based on a Chi-Square distribution. The test statistic is: T_2_X=T_simple∗(V/(V+2∗vare_pool_2+varSNPspecific))
