|dc.description.abstract||Complex adaptive systems exhibit certain types of behaviour that are difficult to predict or understand using reductionist approaches, such as linearization or assuming conditions of optimality. This research focuses on the complex adaptive systems associated with public health. These are noted for being driven by many latent forces, shaped centrally by human behaviour.
Dynamic simulation techniques, including agent-based models (ABMs) and system dynamics (SD) models, have been used to study the behaviour of complex adaptive systems, including in public health. While much has been learned, such work is still hampered by important limitations. Models of complex systems themselves can be quite complex, increasing the difficulty in explaining unexpected model behaviour, whether that behaviour comes from model code errors or is due to new learning. Model complexity also leads to model designs that are hard to adapt to growing knowledge about the subject area, further reducing model-generated insights.
In the current literature of dynamic simulations of human public health behaviour, few focus on capturing explicit psychological theories of human behaviour. Given that human behaviour, especially health and risk behaviour, is so central to understanding of processes in public health, this work explores several methods to improve the utility and flexibility of dynamic models in public health. This work is undertaken in three projects.
The first uses a machine learning algorithm, the particle filter, to augment a simple ABM in the presence of continuous disease prevalence data from the modelled system. It is shown that, while using the particle filter improves the accuracy of the ABM, when compared with previous work using SD with a particle filter, the ABM has some limitations, which are discussed.
The second presents a model design pattern that focuses on scalability and modularity to improve the development time, testability, and flexibility of a dynamic simulation for tobacco smoking. This method also supports a general pattern of constructing hybrid models --- those that contain elements of multiple methods, such as agent-based or system dynamics. This method is demonstrated with a stylized example of tobacco smoking in a human population.
The final line of work implements this modular design pattern, with differing mechanisms of addiction dynamics, within a rich behavioural model of tobacco purchasing and consumption. It integrates the results from a discrete choice experiment, which is a widely used economic method for study human preferences. It compares and contrasts four independent addiction modules under different population assumptions. A number of important insights are discussed: no single module was universally more accurate across all human subpopulations, demonstrating the benefit of exploring a diversity of approaches; increasing the number of parameters does not necessarily improve a module's predictions, since the overall least accurate module had the second highest number of parameters; and slight changes in module structure can lead to drastic improvements, implying the need to be able to iteratively learn from model behaviour.||