The density plot of age at interview for DSM-IV AD and non-dependent subjects shows that a large proportion of non-dependent subjects are younger than the AD subjects at their last interview (Figure 1). We included these high-risk non-AD subjects in the analysis after censoring them at age at latest interview and performed the age at onset analysis of AD using survival models. We used a genome-wide Cox proportional hazards (coxph) analysis, a survival analysis package in R (www.r-project.org), to estimate the hazard ratio for AD occurrence based on genotype, controlling for sex, birth cohort and first principal component (PC1) from EIGENSTRAT (Price et al, 2006). This analysis incorporates a clustered sandwich estimator to account for the familial correlation among observations. Violation of the proportional hazards assumption was tested with non-zero slope of Schoenfeld residuals versus time, using the survival analysis package in R. The analyses of AD for the strongest signals were conducted using the GWAF, an R package for genome-wide association analyses with family data (Chen and Yang, 2010). A logistic regression model was employed with gender, age and
