Bayesian Distributed Lag Interaction Models to Identify Perinatal Windows of Vulnerability in Children's Health

TL;DR

This paper introduces a novel Bayesian distributed lag interaction model that simultaneously estimates windows of vulnerability and effect heterogeneity in prenatal exposure studies, improving understanding of developmental timing effects on children's health.

Contribution

The method allows for flexible modeling of subgroup-specific windows and effects, addressing limitations of previous stratified approaches.

Findings

Identified specific windows of vulnerability for prenatal air pollution exposure.

Detected variation in effects by sex and maternal obesity.

Demonstrated improved bias and error properties in simulations.

Abstract

Epidemiological research supports an association between maternal exposure to air pollution during pregnancy and adverse children's health outcomes. Advances in exposure assessment and statistics allow for estimation of both critical windows of vulnerability and exposure effect heterogeneity. Simultaneous estimation of windows of vulnerability and effect heterogeneity can be accomplished by fitting a distributed lag model (DLM) stratified by subgroup. However, this can provide an incomplete picture of how effects vary across subgroups because it does not allow for subgroups to have the same window but different within-window effects or to have different windows but the same within-window effect. Because the timing of some developmental processes are common across subpopulations of infants while for others the timing differs across subgroups, both scenarios are important to consider when evaluating health risks of prenatal exposures. We propose a new approach that partitions the DLM into a constrained functional predictor that estimates windows of vulnerability and a scalar effect representing the within-window effect directly. The proposed method allows for heterogeneity in only the window, only the within-window effect, or both. In a simulation study we show that a model assuming a shared component across groups results in lower bias and mean squared error for the estimated windows and effects when that component is in fact constant across groups. We apply the proposed method to estimate windows of vulnerability in the association between prenatal exposures to fine particulate matter and each of birth weight and asthma incidence, and estimate how these associations vary by sex and maternal obesity status, in a Boston-area prospective pre-birth cohort study.