| The objective of this research is to develop a set of analytical procedures to detect local spatial clustering of spatial phenomena in a network space while enabling non-uniform distribution of risk population over the network to be explicitly taken into account and then to investigate the relationships between the detected clusters and the characteristics of the network space. The first stage of the analysis aims to detect local spatial clustering in the network space. L&barbelow;ocal i&barbelow;ndicators of n&barbelow;etwork-constrained c&barbelow;lusters&barbelow; (LINCS) are developed as tools of exploratory spatial data analysis. Discussed then is the treatment of the risk population that may not be distributed uniformly over the network. The second stage is to construct models that explain the detected clusters in relation to the characteristics of the network space. To carry out the second stage, this research utilizes inductive learning techniques, more specifically, decision tree induction algorithms and feedforward neural networks, together with traditional statistical techniques, such as regression and discrete choice models.; A spatial cluster in the network space can be defined in two different ways depending on the phenomenon of interest and the scale of data available. When the phenomenon has individual events as its basic element (e.g., vehicle crashes and disease cases) and those events are represented as points distributed over the network space, physical concentration of the points is referred to as a cluster. There is yet another type of spatial phenomenon that does not have countable events and is represented by attribute values associated with the network links (e.g., traffic volume and travel speed). In this case, high values of the attribute and their concentration are referred to as clusters since they imply a high incidence level of the phenomenon in those locations. Low values and their concentration can also be regarded as clusters of low values, implying a low incidence level. This research develops different analytical methods for the event-based and link-based clusters considering that their definitions and properties are quite different.; Performance of the proposed methodologies is illustrated via a case study on vehicle crash distribution in the Buffalo, NY area, in 1997. |
