Main content

Back-calculation, Classification, and Emulation-based Inference for Spatial Infectious Disease Models

Show full item record

Title: Back-calculation, Classification, and Emulation-based Inference for Spatial Infectious Disease Models
Author: Pokharel, Gyanendra
Department: Department of Mathematics and Statistics
Program: Mathematics and Statistics
Advisor: Deardon, Rob
Abstract: Individual-level models (ILMs) are a class of complex probabilistic models which can be used to model infectious disease data. They can incorporate the effects of time and space, the key risk factors of disease transmission, and inference for them is carried out easily within a Bayesian MCMC framework. They are thus useful for modelling individual-level spatial epidemic data. However, fitting these models to what are typically incomplete data can result in poor parameter estimation and can miss important characteristics of the disease systems. Additionally, the complex nature of such ILMs can cause significant computational expense when fitting them to large disease systems. Here, we propose methods and models which address both the incomplete history of epidemic data as well as the computational problem of inference for ILMs while modelling complex disease systems. First, we consider the use of back-calculation of infection times in the context of spatial infectious disease models. Together with prior knowledge about the distribution of the time from infection to disease reporting, we extend the method to incorporate spatial information in the back-calculation mechanism itself. Secondly, the epidemic classification approach of Nsoesie et al.(2011) is extended to the case where the disease generating models are spatial ILMs. This method involves simulating epidemics from various spatial ILMs, and then using a classifier built from the epidemic curve data to predict which model was most likely to have generated an observed epidemic curve. Finally, we propose a method of inference for spatial ILMs based on so-called emulation techniques. The method is set in a Bayesian MCMC context, but avoids calculation of the computationally expensive likelihood function by replacing it with a Gaussian process approximation of the likelihood function of the ILM built from simulated data. All models are fitted to simulated as well as real data, specifically data from an experiment on tomato spotted wilt virus (TSWV).
URI: http://hdl.handle.net/10214/8756
Date: 2015-03
Rights: Attribution-NonCommercial 2.5 Canada


Files in this item

Files Size Format View
Pokharel_Gyanendra_201503_PhD.pdf 4.017Mb PDF View/Open

This item appears in the following Collection(s)

Show full item record

Attribution-NonCommercial 2.5 Canada Except where otherwise noted, this item's license is described as Attribution-NonCommercial 2.5 Canada