Parsimony and complexity in epidemiological models for decision support in animal health.

摘要

Chapter 1 provides an introduction to the subject of epidemiological models, parsimony and complexity in modeling, and a review of the different models used in animal health policy, focusing on the levels of complexity and parsimony approaches used in the published literature. For this, I make a distinction between complexity in the model structure (structural complexity) and in the corresponding parameters used in the epidemiological models (parameter complexity), and I use examples such as the 2001 Foot-and-mouth disease (FMD) outbreak in the UK to depict how complexity and parsimony can directly affect health policy.;In Chapter 2, I explore the effect that structural complexity and parsimony in the model (specifically, population, contact, and spatial heterogeneity) can have in the results commonly used for policy. For this, I developed a flexible herd-level model that simulates the spatial and temporal spread of infectious diseases between animal populations, and used it to evaluate sixteen scenarios, involving combinations of multiple production-types (PT) with heterogeneous contact structure versus single PT with homogeneous contact structure; random versus actual spatial distribution of population units (based on an existing dataset from the state of Minnesota); high versus low disease infectivity; and no vaccination versus preemptive ring vaccination. The results from the scenarios revealed that for fast spreading epidemics, the actual locations of population units (e.g. herds) may not be as relevant to predict outbreak size and duration as information on population and contact heterogeneity. Nonetheless, both population and spatial heterogeneity might be important to model slower spreading epidemic diseases. This information is relevant to inform data collection and model building efforts for epidemiological models used to inform health policy.;In chapter 3, I used an epidemiological modeling framework to estimate the potential losses from a new emerging disease (ED) in channel catfish ponds in Mississippi, with the purpose of estimating animal inventory losses for agricultural insurance purposes. Given the uncertain epidemiology of a new ED, the predictions naturally have a high level of uncertainty, which motivated the design of a structurally complex model to try to evaluate the potential spread of the disease from the "bottom-up". For this, I used two coupled stochastic models that simulate the spread of an ED between and within ponds under high, medium, and low disease impact scenarios, which were parameterized based on a meeting with fish disease experts. The models provided a systematic method to organize the current knowledge on the emerging disease perils and, ultimately, use this information to help develop actuarially sound agricultural insurance policies and premiums. The conclusions from this chapter was that a structurally complex model was necessary to make inferences about a hypothetical ED for which no empirical data is available, but the estimates obtained included a large amount of uncertainty driven by the stochastic nature of disease outbreaks, by the uncertainty in the frequency of future ED occurrences, and by the often sparse data available from past outbreaks.;Chapter 4 evaluates the impact that parsimony and complexity in model parameters can have in model predictions. For this, I developed a Bayesian model that estimates the confidence on individual infection progression using longitudinal screening test results, and use the results to estimate infectious disease model parameters using a Monte Carlo simulation model. The disease trajectories were used to estimate the joint uncertainty distributions of the transition probabilities of the stochastic Markov Chain model, and were then used to project the yearly progression of disease in 20 years. The joint uncertainties in both, the test characteristics and the disease parameters exhibited a significant level of correlation, and sensitivity analysis showed that ignoring parameter correlation considerably underestimated the variance of the model predictions. The main conclusion from this chapter is that the correlation between disease parameters can have an important impact in the variance of relevant disease model outputs and therefore, this correlation should be taken into account when parameterizing stochastic epidemic models. (Abstract shortened by UMI.)