Date of Award
Doctor of Philosophy
State space models are important tools for the analysis of biological data, and although relatively unexplored in the realm of agricultural data, can be used to great effect there as well. We consider cases where data on the underlying system is observed with some form of error; ranging from combining underlying states to misclassification to continuous error-prone measurements of a state-specific process. These cases may be fit into a single common structure based on a simple matrix relationship between the underlying true state and the observed data. When applied to biological lifecycle data, this takes the form of latent unobserved stages which are related to the observed data through a sum-to-the-mean constraint. This allows for estimation of vital rates and other parameters of biological interest from partially observed complex lifecycles, as demonstrated in an application to the fungus \cfullns. We also consider applications to prediction, based on data taken from the National Resources Inventory survey of land use, with applications to Iowa and Pennsylvania. We calculate a distribution for predicting states in unobserved timepoints, based on incorporating categorical auxiliary information from the Cropland Data Layer, and consider both prediction for individual locations and area totals by incorporating survey weights and design variability. Finally, we use simulated data to develop methods for continuous time series auxiliary information, where the mean of the observed value depends on both time and the underlying state. Individual variability is allowed for by use of a random effect, and multiple approximations to the full conditional distribution are considered. We analyze the performance of each predictor using the area under the receiver operator characteristic surface, and discuss methods of estimating the mean curves.
Demuth, Gabriel, "State space models for partially observed biological and agricultural data" (2018). Graduate Theses and Dissertations. 16568.