| When mortality rates are collected over time and space, statistical modeling could be a challenge because it is likely that observations from the same time unit are spatially correlated, and it is also likely that observations from the same geographic area (e.g. a state or a county) are temporally correlated. In addition to the two types of correlations, space-time interaction may occur. Few studies in the literature have addressed the interpretation of space-time interaction. This dissertation will provide an exhaustive account of the interaction. Moreover, this dissertation proposes a general linear mixed model, which adjusts for heterogeneous variability, temporal correlation and spatial correlation simultaneously, and also takes into account the space-time interaction to some extent.;Based on the developed model, six approaches are proposed and applied to county-specific 30-year-period infant mortality of Alabama. In the application, besides taking care of temporal correlation, spatial correlation, spatial heterogeneity and space-time interaction, issues such as checking temporal and spatial correlations, choosing appropriate temporal and spatial covariance models, comparing models, and assessing model performance have been discussed as well.;A simulation study has been conducted to explore the performance of three general approaches in various temporal-spatial settings. The approach which summarizes different temporal trends bivariately is consistently better than the other two approaches which summarize temporal trends univariately. Comparing the two univariate approaches, VAR Approach is better than the AR Approach, especially when temporal correlation across areas changes in a small range. Since the three general approaches employ the same method of handling spatial correlation, the spatial effect did not interact with the approach effect. |
