Data analysis was divided into three parts: First, we analyzed associations between the GRS and developmental phenotypes of smoking behavior. Second, we analyzed associations between the GRS and mature phenotypes. Third, we tested whether developmental phenotypes mediated associations between the GRS and mature phenotypes. We used different statistical models to analyze outcome data as required by the outcome's distribution. We analyzed continuously distributed outcome data (e.g. lifetime cigarette consumption in pack years) using ordinary least squares (OLS). We analyzed dichotomous outcome data (e.g. daily smoker by age 15 years) using Poisson regression models as this is a standard method to derive relative risks.41 We analyzed count outcome data (e.g. the number of assessments at which the study member met criteria for nicotine dependence) using negative binomial regression models to account for the over-dispersion of many of the count measures.42 We analyzed hazards of smoking initiation, progression to heavy smoking, becoming nicotine dependent, and relapsing from a quit attempt using Cox proportional hazard models. To account for differences in the frequency with which study members attempted cessation, we constructed panel datasets
