MSc. Defence: Parameter Estimation Using Random Forests For Two-Stage Individual Level Models of Infectious Disease Spread

CANDIDATE: Matthew Baxter

ABSTRACT: This thesis aims to develop various random forest models in order to estimate parameters for disease model transmission using the standard Individual-Level Model (ILM). Data is simulated using an ILM in a susceptible-infectious-removed (SIR) framework. The standard ILM is then extended to a Two-Level Individual-Level Model (TILM) to account for between-city and within-city disease spread. An observation model is then applied to this TILM to represent epidemics observed in the real-world. From there, random forest models are fit in three scenarios: two considering random forest regression, and one considering random forest classification. Within this, three sub-models are fit in each scenario, using 5, 10 and 25 epidemic replications per parameter combination. This results in 9 total models being considered in this study. One of the regression models, along with the classification model, follow a factorial design. The second regression model follows a random design. In order to test how these 9 different models perform, we use epidemic prediction intervals to test both the ability of the models to predict epidemic spread, as well as estimate the parameters for the TILM. This is done by generating true epidemics with known parameter estimates, substituting them into the respective models, getting predicted parameter values, and simulating multiple epidemic curves. We were able to determine that the random design regression approach was found to predict epidemic spread the best of all models, as well as provide the most accurate parameter estimates of all models considered.