A1253
Title: An investigation of the choice of time scale for studies with time-to-dementia-related outcomes
Authors: Kyle Conniff - University of Wisconsin Madison (United States) [presenting]
Daniel Gillen - University of California, Irvine (United States)
Abstract: Time to event analyses using Cox proportional hazards models require specifying a time scale, which determines the risk sets used to estimate covariate effects. Although age and time on study are the two most common choices, prior simulation studies have evaluated them only under correct specification of the underlying time process, producing mixed guidance for applied researchers. This literature is extended by examining the performance of the Cox model when the analysts chosen time scale is misspecified relative to the true event generating mechanism. Scenarios are constructed in which the hazard function changes at an unobserved time $T_0$ that is unrelated to study entry. Under an age based time scale, the hazard is low before age $T_0$ and increases afterward; under a time on study time scale, the hazard is zero until the individual reaches $T_0$, regardless of entry age. Both time scales are evaluated across settings where the predictor of interest is either correlated or uncorrelated with age at entry. When the predictor and entry age are uncorrelated, both time scales yield similar performance even under misspecification. However, when they are correlated and the true time process aligns more closely with time on study, the age time scale model has much more bias than the time-on-study model. Overall, the results indicate that time on study is more robust to misspecification of the true underlying time scale.
