2024 Theses Doctoral
Modeling and causal inference methods for analyzing and transporting an environmental mixture effect
An environmental mixture is composed of multiple environmental exposures. Quantifying this joint effect on health outcomes mirrors what occurs in nature, offering significant benefits to environmental epidemiological research. However, analyzing the impact of an environmental mixture poses numerous statistical and inferential challenges. Motivated by the Strong Heart Study (SHS), a prospective cohort study of cardiovascular disease (CVD) outcomes among three American Indian communities where urine samples were collected at three visits and analyzed for concentrations of various metal exposures, this dissertation aims to improve our ability to analyze the effect of multiple, continuous, and correlated exposures with complex relationships on a health outcome using observational study data, such as the metal mixture exposure in the SHS. The contributions of this dissertation address two challenges inherent in environmental mixture analyses: modeling methods and the transportability of estimated effects.
In the first project presented in this dissertation, our goal was to evaluate the performance of available modeling methods for estimating the impact of an environmental mixture on survival outcomes. While survival time outcomes (also known as time-to-event outcomes) are very common in epidemiological studies, little attention has been given to examining the performance of existing modeling methods when estimating the effect of an environmental mixture on a survival outcome. In this chapter, we identified applicable and readily available modeling methods, assessed their performance through simulations replicating various real-world scenarios, and applied the selected methods to estimate the effect of a metal mixture on CVD incidence in the SHS. We examined proportional hazards (PH) based models as well as more flexible, machine learning-style models. Our simulations found that, when the PH assumption held, the effect estimates via flexible models had higher bias and variance compared to PH methods. However, when the PH assumption was violated, this discrepancy between the methods decreased and the more flexible methods achieved higher coverage. These simulation findings underscore the importance of demonstrating the robustness of findings across various modeling approaches in environmental epidemiology. In the SHS analysis, all methods found a significant, harmful effect of the metal mixture on incident CVD. However, the more flexible approaches found larger point estimates with wider confidence bands.
The second and third projects of this dissertation focus on constructing a framework for transporting an environmental mixture effect across populations. Numerous methods exist for analyzing environmental mixture effects within a population where sample data is available for. However, being able to adjust these effects based on the exposure/covariate distribution of a different target population would enable more precise estimation of the mixture effect in that population. This, in turn, allows for more accurate estimation of effects for populations distinct from those sampled. This broadens available data sources and provides significant advantages to researchers and policymakers interested in specific populations.
The second project leverages causal inference concepts to formally extend the transportability literature to the environmental mixtures context. We defined a relevant intervention with favorable properties concerning the exposure concentration positivity assumption and explicitly outline the assumptions needed to transport its effect across two observational studies. To assess whether the target population is well represented in the study population sample, which is required for the positivity of population membership assumption, a matching algorithm is proposed. Subject's environmental mixture exposure profiles are incorporated into subject matching on exposures and covariates between the two populations. Simulation results demonstrate that the matching algorithm effectively detects non-overlap across populations, with well-overlapped populations yielding minimally biased transported effect estimates, while those with insufficient overlap exhibit greater bias. Applying this framework, we estimated the effects of a metal mixture on coronary artery calcification (CAC) in the SHS cohort by transporting the effects observed in the Multi-Ethnic Study of Atherosclerosis (MESA). Although CAC was not directly measured in the SHS, its importance as a subclinical indicator of advanced atherosclerosis and its link to elevated cardiovascular risk underscore the significance of exploring its relationship with metal exposures in the SHS. Despite larger effects observed in the MESA population, significant effects persisted within the SHS, providing insights for innovative strategies in preventing and treating atherosclerosis progression among American Indian populations.
In the third project, we turned our focus to violations of internal and external validity exchangeability assumptions. We proposed the use of a negative control exposure analysis modeled with Bayesian Kernel Machine Regression with hierarchical variable selection to identify unmeasured confounding in the context of multiple, continuous, and correlated exposures when exposures are measured at various time points. Additionally, we developed a novel method for detecting violations of transportability assumptions by assessing the transportability of an effect within a study population. If the internal validity assumptions are plausible, then the inability to transport an effect within a study population suggests the presence of unmeasured effect modification in the study sample. Through simulations, we demonstrated the efficacy of these methods in detecting unmeasured confounding and effect modification. We applied these methods to assess the robustness of the estimated effect of a metal mixture on fasting blood glucose levels in the SHS to violations of transportability assumptions and found evidence of both unmeasured confounding and effect modification. However, the internal effect estimate remained significant and robust to unmeasured confounding.
Geographic Areas
Subjects
Files
- Mayer_columbia_0054D_18635.pdf application/pdf 3.52 MB Download File
More About This Work
- Academic Units
- Biostatistics
- Thesis Advisors
- Valeri, Linda
- Degree
- Ph.D., Columbia University
- Published Here
- August 14, 2024