Using a serial marker to predict a repeated measures outcome in a cohort study

Abstract

Consider the cohort design and suppose that the outcome of primary interest is a continuous random variable observed repeatedly over time. Suppose that the value of a "clinical marker," which is thought to be predictive of the primary outcome, is also recorded. We would like to determine whether there is an association between the two variables as they evolve over time. We might also want to predict the pattern for the primary outcome conditionally on a specific profile of clinical interest for the serial marker. A model is developed to address these issues. One regression model is created for the primary outcome while a second regression model is developed for the clinical marker. The vector autoregressive model, i.e., VAR(1), is used to characterize the covariance structure between the two sets of repeated observations. Hypotheses of interest are described and procedures for testing them are elaborated. The Diabetes Control and Complications Trial is used to illustrate the procedures.