Date of Award

Spring 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Public Health

First Advisor

Li, Fan

Abstract

Causal mediation analysis is a popular statistical tool for evaluating the extent to which an observed exposure-outcome relationship is mediated by an intermediate variable, also referred to as the mediator. Through mediation analysis, the total effect (TE) from the exposure on the outcome can be decomposed into a natural indirect effect (NIE) through the mediator and a natural direct effect (NDE) whose impact is due solely to the exposure and possibly other mediators. Causal mediation analysis has been applied in a variety of empirical research fields, including but not limited to epidemiology, implementation science, economics, and psychology. In implementation science, for example, the causal mediation analysis is widely-used to understand the extent to which an implementation strategy causing a health outcome is mediated by an intermediate implementation outcome, such as acceptability, adoption, and fidelity of the implementation strategy. This dissertation develops advanced statistical methodologies for causal mediation analysis. Objectives of this dissertation are two-folded. First, we develop measurement error correction methods for mediation analysis when a continuous exposure is mismeasured for independent data. Measurement error is a common problem in mediation analysis in observational studies, and standard mediation analysis approaches will be biased if the measurement error is not appropriately addressed. In existing literature, no methods are available to address measurement error in a continuous exposure, although exposure measurement error in continuous variables is a major source of bias in epidemiologic studies. Based on this observation, we develop easily-implemented strategies for correcting the exposure measurement error-induced bias in mediation analysis. Second, we develop a semiparametric framework for conducting causal mediation analysis in cluster-randomized trials, which feature correlated data observations. The current literature on mediation analysis primarily focuses on independent data, whereas statistical methods are limited for assessing mediation with cluster data, as occurs, for example, in cluster-randomized trials. Existing statistical methodologies for assessing mediation with cluster data either lack of an ability to address interactions among individuals in the same cluster or lack methods for explorations of the spillover mediation effect. Several applied studies of mediation analysis in clustered data settings also neglect the correlation structure of data and therefore the resulting inference and conclusion may be biased. In this dissertation, we propose appropriate assumptions to identify the NIE, NDE, and spillover mediation effect in cluster-randomized settings and then develop a doubly robust approach to estimate these mediation effects that is robust against some degree of model misspecifiation of the working models.

Download

Share

COinS