Boston University
Change point estimation and mediation analysis using longitudinal data
Abstract
dc:description.abstractAnalyzing longitudinal data helps deepen our understanding of disease development mechanisms. Longitudinal clinical biomarkers are widely studied to investigate disease progression, however, meaningful inference from their biomarkers is complicated by non-linear trajectories. Change point models capture these non-linear patterns by allowing different slopes over time, thus identifying shifts in disease dynamics. Multiple methods are proposed to estimate change points, but few studies systematically compare these methods within a linear mixed model. Longitudinal studies go beyond simple associations by leveraging the sequence of measurements over time to estimate effect magnitudes and inform causal direction. Mediation analysis evaluates exposure-outcome relationships by quantifying the pathways through intermediate variables. Current longitudinal mediation analyses estimate either the average indirect effect over time or the time-lagged indirect effect of prior exposures or mediators. However, established methods do not estimate the cumulative effect of exposures across all time points on the current outcome. Moreover, missing data are common in longitudinal studies, and methodological challenges remain for longitudinal mediation analysis, particularly when missing data are under missing not at random (MNAR). To address these knowledge gaps, this dissertation develops and applies methods for change point estimation and mediation analysis using longitudinal data. This work is organized into three distinct projects. In Chapter 2, we compare three approaches for estimating complex change point structures within a linear mixed model. We consider subject-level change points, covariates in the change point, and different slopes after change point. In Chapter 3, we propose an unbiased estimator for cumulative indirect effect (CIE) in longitudinal mediation analysis. We estimate the variance of CIE with three approaches: exact variance method, delta method, and bootstrap procedure. We compare these variance estimators and recommend the bootstrap procedure for empirical studies. Moreover, we demonstrate that CIE estimation is subject to bias in the presence of unmeasured confounders. In Chapter 4, we extend an existing method by implementing a weighted Expectation-Maximization (EM) method to address MNAR in the mediator and outcome variables. Our method yields minimal bias in estimating CIE and reduces computational burden by handling MNAR missingness sequentially at each time point. We evaluate our approaches through extensive simulation studies and compare their performance with that of existing methods across a range of scenarios. We then apply these methods to real-world data to characterize long-term change patterns and estimate indirect effects linking exposures, mediators, and outcomes. In summary, this dissertation advances our understanding of complex changes and mediation pathways underlying disease mechanisms.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Biostatistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Li, Yi
- Advisors dc:contributor.advisor
-
- Liu, Chunyu
- Lee, Jungwun
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53449