We conducted a genome-wide Cox proportional hazards regression72 assuming an additive effect from SNP dosage. The Cox proportional hazard regression was implemented in the R survival analysis package. We incorporated sex, site and the first three principal components from EIGENSTRAT30 in all our regression models to control for their effects. For EADI, sex and four principal components were included in the model. For the Cox model, the time scale is defined as age in years, where age is age at onset for cases and age at last assessment for controls. The formula applied is as followed: h(t|X)=h0(t)exp(∑i=1pβiXi)where X = (X1, X2, …, Xp) are the observed values of covariates for subject i. The Cox model has previously been shown to be applicable to case-control datasets without an elevated type 1 error rate nor overestimation in effect sizes73,74. The model assumes log-linearity and proportional hazards. The assumption of log-linearity is common in the additive logistic regression used in a typical GWAS. We validated the assumption of proportional hazards assumed by the Cox model by conducting the Schoenfeld test in the 22
