Degree Type

Creative Component

Semester of Graduation

Spring 2019

Department

Statistics

First Major Professor

Emily Berg

Degree(s)

Master of Science (MS)

Major(s)

Statistics

Abstract

The class of Markov random field models known as auto-models provides a flexible and highly-interpretable model structure for analysis of lattice data, which may be of particular interest for research in agriculture, forestry, image construction, or genetics. This paper proposes a Markov random field model with a bivariate Binary-Gaussian response, which provides a foundation for reliable statistical inference with multivariate response distributions of mixed types in the presence of spatial dependence. Pseudo-likelihood-based estimation and inference are investigated via simulation study. The model adequately captures both covariate and spatial dependence relationships under a not-too-strong spatial dependence regime, however the definition of ``strong spatial dependence" is a difficult question when distributions are mixed. The model is implemented in 2 cases: data from wheat field trials and data on observations of prairie chickens. The full bivariate model outperforms univariate and non-spatial models in capturing spatial dependence and covariate relationships together.

Copyright Owner

Harms, Steven

File Format

PDF

Share

COinS