Subsequently, three multivariate Cox proportional hazards regression models were constructed to identify the strongest predictors after accounting for all other variables. The starting model adjusted for eight demographic variables and included 11 risk factors and eight R/S variables. Since an early onset drinker generally progresses more slowly through these stages than later onset drinkers, this difference was controlled by an indicator of drinking before age 15. Each model was reduced by manual backward deletion to identify a final model of significant predictors. ODD, CD, MDD, and Trauma were modeled as time-dependent covariates; this eliminated cases where risk occurred after (rather than before) a given alcohol transition (causality can only be inferred if a predictor precedes transition). The approach was to construct “Person year data” using SAS, version 9.1 (SAS Institute Inc., 2006) consisting of one line of data for each person for each year of life. For cases positive for ODD, CD, MDD, or Trauma, a time-dependent variable was computed as 0 (absent) for each year up to the age of onset of the risk factor, and then 1 (present)
