Bayesian latent variable methods for longitudinal processes with applications to fetal growth.

摘要

We consider methods for joint models of exposure and response in epidemiologic studies. In particular, we show how latent variable methods provide a structure for obtaining inference about multistate growth processes and multiple longitudinal and cross-sectional outcomes. Each model utilizes underlying, subject-specific latent variables to account for the correlation that arises from taking multiple observations on the same sampling unit. We also consider latent variable mixture models in order to more exibly model the latent variable distributions and identify latent classes of subjects who are of particular scienti c importance. We apply our methods to applications in reproductive health, obtaining interesting new insights while developing and applying statistical methodology.; We first consider the problem of estimating a multistate growth process with unknown initiation time to determine individual early fetal growth. Using cross-sectional data, we identify fetuses that have a latent tendency to grow relatively quickly and slowly and show that slow growth early in pregnancy is associated with an increased risk of future pregnancy loss. These results are important to researchers who use early ultrasounds to date pregnancies under the assumption that there is no measurable variability in early fetal growth.; Paper two is concerned with jointly modeling the unusual, asymmetric distributions of birth weight and gestational age. Using latent variable mixture models, we identify a latent class of subjects who are more likely to deliver early and have low weight. We also allow observed covariates to be associated with latent class membership. Our approach provides researchers a new method for examining low birth weight and pre-term birth.; In paper three, we aggregate multiple ultrasound measurements on fetal size and blood restriction using latent variables that follow mixture distributions to identify a latent class of subjects who are growth restricted during pregnancy. We then consider a joint model that examines the associations between covariates, early growth restriction, and outcomes measured at birth. Our methods are able to identify a latent class of subjects who have increased blood ow restriction and below average intrauterine size during the second trimester who are more likely to be growth restricted at birth.