Author ORCID Identifier
Date of Graduation
Doctor of Philosophy (PhD)
Mendelian Randomization (MR) is an epidemiological framework using genetic variants as instrumental variables (IVs) to examine the causal effect of an exposure on an outcome. It is widely used to detect causal factors of diseases and provide insight into the biological pathway of diseases. Current methods under the MR framework are built to estimate the unidirectional causal effects of exposures on outcomes and neglect the potential bidirectional causal effects. However, a bidirectional causal effect creates a feedback loop that biases the casual inference in MR studies. Furthermore, current MR methods estimate the causal effect as a single value using cross-sectional data and ignore the possibility of time-varying exposures, outcomes, and causal effects. This assumption can lead to biased estimation, even invalid inference. In this dissertation, we propose novel statistical methods based on the MR framework for a more accurate causal effect under two scenarios: 1) there is a bidirectional causal effect between two phenotypes 2) the causal effect is time-varying.
To overcome the limitations of estimating bidirectional causality with the standard MR method, we proposed two novel approaches to estimate bidirectional causal effects using MR: BiRatio and BiLIML, which are extensions of the standard ratio, and limited information maximum likelihood (LIML) methods, respectively. We compared the performance of the two proposed methods with the naïve application of MR methods through extensive simulations of several scenarios involving varying numbers of strong and weak IVs. Our simulation results showed improved performance of our proposed methods than naïve applications of standard methods. We also applied the proposed methods to investigate the potential bidirectional relationship between obesity and diabetes using the Multi-Ethnic Study of Atherosclerosis cohort data. Our results from the BiLIML method revealed the bidirectional causal relationship between obesity and diabetes across all racial subpopulations.
In addition, since many cohort studies collected longitudinal data, we proposed a longitudinal MR (LMR) model based on a mixed effects model to estimate the time-varying causal effects with available longitudinal data. We performed an extensive simulation to compare the accuracy of the estimation of time-varying causal effect estimations by applying the standard MR and LMR methods on longitudinal data collected across 18 years. Our simulation shows that our LMR model provided a more accurate estimation than the standard MR model and higher power when the causal effect is time-constant or time-varying.
The methodologies and frameworks developed in this dissertation generate novel insights into MR studies for estimating more accurate causal effects using MR methods. Our proposed methods will contribute to detecting risk factors for disease with higher accuracy and understanding the disease progression, eventually contributing to disease prevention and intervention through prioritizing treatment targets.
Causal inference, Mendelian Randomization, Bidirectional causal inference, Longitudinal data, Observational study
Available for download on Wednesday, July 02, 2025