Cox Regression Using Different Time-scales

Typically in cohort studies, the time-scale used in Cox regression models is time-on-study, adjusting for age as a covariate. However, age can also be used as the time-scale, where subjects enter the analysis at their baseline age (left-truncation) and exit at their event/censoring age. Using SAS PROC PHREG, we compared five methods using time-on-study and age as the time-scales. We used a subset of data (n=85,882) from the large California Teachers Study cohort, with five years of follow-up, invasive breast cancer as the outcome (n=1,428) and alcohol consumption (> 20 grams/day) as the risk factor of interest. Using a time-on-study time-scale adjusting for age continuous in years produced a slight overestimate of the alcohol effect compared with the more accurate but computationally more intensive age time-scale methods. However, allowing for different age effects for younger and older ages gave virtually identical results as using an age time-scale.