Abstract
Spatial cluster detection is a classical question in epidemiology: Are cases located near other cases? In order to classify a study area into zones of different risks and determine their boundaries, we have developed a spatial partitioning method based on oblique decision trees, which is called spatial oblique decision tree (SpODT). This non-parametric method is based on the classification and regression tree (CART) approach introduced by Leo Breiman. Applied to epidemiological spatial data, the algorithm recursively searches among the coordinates for a threshold or a boundary between zones, so that the risks estimated in these zones are as different as possible. While the CART algorithm leads to rectangular zones, providing perpendicular splits of longitudes and latitudes, the SpODT algorithm provides oblique splitting of the study area, which is more appropriate and accurate for spatial epidemiology. Oblique decision trees can be considered as non-parametric regression models. Beyond the basic function, we have developed a set of functions that enable extended analyses of spatial data, providing: inference, graphical representations, spatio-temporal analysis, adjustments on covariates, spatial weighted partition, and the gathering of similar adjacent final classes. In this paper, we propose a new R package, SPODT, which provides an extensible set of functions for partitioning spatial and spatio-temporal data. The implementation and extensions of the algorithm are described. Function usage examples are proposed, looking for clustering malaria episodes in Bandiagara, Mali, and samples showing three different cluster shapes.