To test association, initiation was regressed on the best guess genotype and covariates. The covariates were sex, age at the last survey, the birth cohort (i.e., two birth cohorts containing individuals born between 1951 and 1970 and 1971–1999, respectively, and the 1915–1950 birth cohort as the reference category), 3 principal components to correct for Dutch population substructure (Abdellaoui et al. 2013), and sample specific covariates to account for batch and for chip effects. A GEE (Carey et al. 2012) logistic model was employed. To model the familial relatedness, we used an exchangeable working correlation matrix. This accounts for the familial correlations by means of a single correlation among the family members. The effect of possible misspecification of the familial covariances on the standard errors was corrected by means of a sandwich correction (Minica et al. 2014; Dobson 2002). The sandwich-corrected GEE approach was implemented by using the R-package gee (Carey et al. 2012), accessed from Plink (Purcell et al. 2007) which communicates with R (Team 2013) via the Rserve package (Urbanek 2013).
