The proportional hazards assumption (PHA) underlying Cox proportional hazards analyses requires that risk remain constant over time. The PHA was tested by the Grambsch and Therneau test of the Schoenfeld residuals (Grambsch & Threneau, 1994). Violations indicated uneven change across time (such as a large effect in early years and little effect in later years). In the first model (age to first full drink), 12 variables violated the PHA, and the model was adjusted by splitting the age distribution into ages 1–11 years, 12–14 years, 15–17 years, 18–20 years, and 21 years and up. Interaction terms then represented the original variables as two (or more) age-specific sub-variables that covered the entire period and met the PHA (e.g., CD required two interaction terms, one modeling ages 0 to 14 and the other modeling ages 15 and up). The same procedure was used with the second model except that the time measure was the duration in years from first drink and onset of at-risk drinking. In this case, the distribution of years of duration was split into four divisions: 0–2 years, 3–5
