Date of Award
Doctor of Philosophy
This thesis focuses on constructing appropriate statistical models to monitor the dynamics of disease transmission in animal disease surveillance system. One big challenge in analyzing such disease surveillance data is that the diagnostic tests are usually known to have imperfect sensitivity and specificity, thus the observations are usually misclassified, which introduces uncertainty in determination and modeling of the true disease status among animals. The thesis consists of three projects focusing on three different models and statistical inferences for different disease surveillance datasets. In the first project (Chapter 2), we propose a latent spatial piecewise exponential model for the misclassified disease surveillance data and apply the model to a data from the porcine reproductive and respiratory syndrome virus (PRRSV) disease. The misclassification of test outcomes are accounted for by using a two-level survival model. Spatial distance and time-varying covariates are incorporated to account for disease transmission. We show that our model is efficient in capturing the data features and easy to implement. In the second project (Chapter 3), we are motivated by parameter estimations in hidden Markov models (HMM) and mixed HMM (MHMM). The HMM can be applied to the animal disease surveillance data where the outcomes are with misclassification, and with a group level random effect added, the MHMM can model the correlation structure. However, the parameters estimation in these models are challenging because of the latent variables and random effect. We propose a pairwise fractional imputation using the idea of parametric fractional imputation as well as the Markov property. The proposed estimation method is shown to provide efficient parameter estimates and achieves computational efficiency. In the third project (Chapter 4),
we further investigate into the piecewise exponential model and consider estimation of the hazard functions where a monotone restriction is put on the hazard. When observations are with misclassification, the estimation involves EM-algorithm and the principle of isotonic regression is used for constraint optimization of the model parameters. Details of the estimation algorithm is developed in this chapter and the bootstrap confidence interval is constructed for measuring the variability of the estimates. The proposed model is then applied to another PRRSV surveillance study in the swine population.
Sun, Yaxuan, "Statistical methods in modeling disease surveillance data with misclassification" (2017). Graduate Theses and Dissertations. 16223.