Statistical methods in analyzing complex longitudinal data

Eunice Kennedy Shriver National Institute Of Child Health & Human Development

Investigators

Linked publications & trials

Abstract

In studying repeated pregnancies, the goal is to model the trajectory of pregnancy outcomes, and identify individual risk factors both at the baseline and longitudinally. However, directly application of longitudinal data methods, e.g., generalized estimating equations, may lead to biased inference due to several reasons. First, women with more pregnancy complications are likely to have fewer pregnancies, so the longitudinal observations may be subject to informative cluster size. Second, the time gap between pregnancies may be indicative of a womans underlying fertility, and may also be correlated with the pregnancy outcomes, resulting in informative gap time. We will develop new statistical models to account for the unique data structure with application to the pregnancy study. Specifically, we will employ parametric models for the cluster size and gap time, and then develop a propensity score approach to reweight the estimating equations. In addition, we will investigate how informative cluster size and gap time affect the inference in transition model frameworks, where the interest is in estimating the recurrence rate. In many longitudinal studies, both the outcome and covariates are time-varying, the challenge is to quantify the association between two correlated processes. A shared random effects model provides a natural and easy-to-interpret way to model the association, but it does not directly estimate association measures such as odds ratios or relative risks. We will explore the transition model as an alternative framework, which is compatible to the shared random effects model, and provides concurrent and prospective association measures.