Thursday, June 17 2021
9:00am - 11:00am
PhD Thesis Presentation
Prediction and Cluster Detection for Spatial Data

Abstract: This dissertation is concerned with spatial statistics. Topics that we primarily discuss include cluster detection for geographically aggregated data and spatial prediction for big geostatistical data. The spatial scan method is intended to detect clusters that are not explained by baseline random processes. We propose a novel methodology that integrates likelihood-based spatial information to optimally estimate the tuning parameter controlling the size of windows used by the scan method. This estimation avoids detecting unrealistically large disease clusters. We then explore a more versatile, adjusted likelihood version of the scan method that searches arbitrarily shaped connected subsets of regions over elliptical windows whose shape, angle, and size vary. This construction allows the identification of regions contained in clusters of irregular shapes to a great extent. Lastly, we develop a semiparametric method for extending a well-known dimension-reduction kriging technique, fixed rank kriging, in which the dimension is reduced by projecting spatial process onto a low dimension space spanned by a fixed number of known basis functions. The nonparametric estimation of complex mean structures of large data sets results in more flexible and local prediction for fixed rank kriging. Methods presented in this research are assessed and compared using simulated benchmark data sets as well as real data sets.
Speaker:Mohammad Meysami
Location:see email for zoom link

Download as iCalendar