Efficient design and modeling strategies for follow-up studies with time-varying covariates

Abstract

Epidemiological studies can often be designed in several ways, some of which may be more optimal than others. Possible designs may differ in the required resources or the ability to provide reliable answers to the questions under study. In addition, once the data are collected, the selected modeling approach may affect how efficiently the data are utilized. The purpose of this dissertation is to investigate efficient designs and analysis meth ods in follow-up studies with longitudinal measurements. A key question is how to select optimally a subcohort for a new longitudinal covariate measurement if we cannot afford to measure the entire cohort. Another key question we consider is how to determine the reasonable number of longitudinal measurements. Different ways to utilize longitudinal covariate measurements in modeling cardiovascular disease (CVD) mortality are also studied. Follow-up data are modeled using parametric or semiparametric proportional haz ards models. Subcohort selections are carried out using optimality criteria initially developed for optimal experimental design. Measures of model discrimination are ap plied to plan the number of longitudinal measurements. The topics are studied using simulations and the East–West data, which are Finnish part of an international follow- up study in the field of cardiovascular epidemiology, the Seven Countries Study. This work demonstrates that the cost-efficiency of follow-up designs can be improved by careful planning. The proposed method for selecting optimal subcohorts is shown to outperform simple random sampling and it is demonstrated how the number of longi tudinal measurements can be determined using simulated data and data from previous similar studies. The results also indicate that individual-level changes and cumulative averages of classical risk factors are good predictors of CVD mortality.