Using a serial marker to predict a repeated measures outcome in a cohort design--results of a simulation study

Consider the cohort design and suppose that the outcome of primary interest is a continuous random variable observed repeatedly over time. Suppose that there is a second variable of clinical relevance which is also observed repeatedly. We are interested in assessing whether the "serial marker" is in some sense predictive of the primary outcome. We would also like to predict the trend for the primary outcome assuming that the clinical marker follows a profile of specific clinical interest. In series of earlier papers, we have addressed these issues by applying a bivariate repeated measures model. One regression model was prescribed to relate the primary outcome to important explanatory variables, while a second regression model was prescribed for the serial marker. In this paper, we perform a series of simulation studies to investigate the empirical properties of this approach. Bivariate repeated measures data were generated at random, and basic study parameters including the sample size, the number of time points, the degree of serial correlation within the clinical marker, and type of association between the serial marker and the primary outcome were varied. The ability of the methodology to capture the underlying relationship between the two set of repeated measures was assessed. The ability to predicting the primary outcome corresponding to a known marker profile of specific interest was examined.