Evaluating Statistical Approaches to Model Multi-Pollutant Mixtures on Common Bottlenose Dolphin (Tursiops Truncatus) Health Along the Eastern Coast of Florida and South Carolina Open Access

Kennerley, Victoria (Spring 2020)

Permanent URL: https://etd.library.emory.edu/concern/etds/r207tq45p?locale=en


Background: Bottlenose dolphins (Tursiops truncatus) are the most common cetacean species found in coastal and estuarine ecosystems along the southeastern coast of the United States. Their widespread distribution and role as apex predators make them an ideal sentinel species for monitoring pollutants. Previous studies have revealed associations between individual chemical pollutants and pathophysiological endpoints in dolphins. However, the reality is that dolphins are exposed to a large number of pollutants simultaneously and single-pollutant models do not capture the mixture and potential interplay of combined exposures. In recent years, an increased number of studies have implemented more sophisticated statistical methods to assess the relationship between multi-pollutant mixtures and health outcomes in humans. These methods have not been previously applied to marine mammal research.

Methods: In this study, we focus on combining the application of principal component analysis and Bayesian kernel machine regression to evaluate the association between environmental exposure mixtures and absolute counts of MHCII+ cells in Atlantic bottlenose dolphins while simultaneously examining the impact of missing values using random forest imputation and multiple imputation.

Results: Multiple imputation resulted in the highest average pollutant concentrations. A statistically significant association was found between absolute counts of MCHII+ and the first and second principal components primarily made up of 1) ∑PFCs, ∑PFCAs, and ∑PFSAs and 2) ∑Pesticides, ∑PCBs, and ∑PBDEs across all methods for handling missing data. Bayesian Kernel Machine Regression with hierarchical variable selection identified PFCAs as most influential. Principle components 1 and 2 were still found to be significant in Bayesian Kernel Machine Regression analyses with principle components as predictors.

Conclusions: Of the methods presented, Bayesian Kernel Machine Regression with hierarchical variable selection yielded the most straightforward results by identifying a single predictor with the most influence. Analyses conducted with different methods for handling missing data yielded similar results across all three methods.

Table of Contents

Introduction (Pages 1 - 3)

Methods (Pages 3 - 8)

-Study Population (Pages 3 - 4)

-Exposure Assessment and Handling of Missing Values (Pages 4 - 6)

-Statistical Analysis (Pages 6 - 8)

--Principal Component Analysis and Regression (Page 6)

--Bayesian Kernel Machine Regression with Variable Selection (Pages 6 - 8)

--Bayesian Kernel Machine Regression with Principal Components (Page 8)

Results (Pages 8 - 10)

-Principal Component Analysis (Page 9 )

-Bayesian Kernel Machine Regression (Pages 9 - 10)

Discussion (Pages 10 - 11)

Tables and Figures (Pages 12 - 17)

Supplementary Tables (Page 18)

References (Pages 19 - 21)

About this Master's Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
Subfield / Discipline
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Last modified

Primary PDF

Supplemental Files