Date of Award

2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Laurent Hébert-Dufresne

Second Advisor

Jean-Gabriel Young

Abstract

Effective public health interventions must balance an array of interconnected challenges, and decisions must be made based on scientific evidence from existing information. Building evidence requires extrapolating from limited data using models. But when data are insufficient, it is important to recognize the limitations of model predictions and diagnose how they can be improved. This dissertation shows how principles from Bayesian experimental design can be applied to surveillance and control efforts to allow researchers to get more out of their data and direct limited resources to best effect. We argue a Bayesian perspective on data gathering, where design decisions are made to maximize utility on average over a joint distribution of beliefs and outcomes, is better suited to the epidemiological setting where observational studies are the norm. We illustrate these ideas using a range of models and topics across epidemiology.

We focus first on Chagas disease, where in Guatemala an endemic vector continues to cause a high rate of domiciliary infestation in rural communities, and shortages of insecticides and resources for critical house improvements hamper control efforts. Using an adaptive sampling and geospatial modeling framework, we show that interpolating from a traditional design goal of minimizing prediction uncertainty to targeting houses of high risk can satisfy competing objectives, namely, to efficiently identify houses in need of treatment while mitigating sampling bias. We next focus on tick surveillance in the southeastern United States. By framing tick collection surveys as a design problem over time and space, we show optimal survey design can yield greater information compared to random or convenience sampling. Finally, we shift attention from experimental design to the closely related concept of practical identifiability. We propose a novel method to quantify practical identifiability which reflects the average amount of posterior shrinkage that would occur in a Bayesian analysis, without requiring computationally expensive techniques like Markov Chain Monte Carlo. With this method, we demonstrate the limits of using epidemiological models to derive standard statistics such as the basic reproductive number early in an outbreak.

Language

en

Number of Pages

136 p.

Download

Share

COinS