5.4 Area data: Local measures of spatial autocorrelation
Health data within a GIS are often organized as disease rates for geographical areas such as wards, output areas, or census tracts. These rates can be converted into probability maps or smoothed using empirical Bayes estimation before further analysis takes place. These techniques reduce ‘small number’ problems, where rates in areas with small populations are unstable. Further analysis often involves identifying areas that show an unexpectedly high concentration of disease or clusters.
One of the more common techniques for identifying clusters from disease data for areas is to use Local Indicators of Spatial Autocorrelation, otherwise known as LISA. Spatial autocorrelation is the propensity for neighbouring objects to share similar properties. There are two types of measures of spatial autocorrelation: global and local. Global measures assess all locations in a study area and then produce a single summary statistic to indicate the degree of similarity between neighbours across the entire area. Local measures of autocorrelation can be calculated for each location within the study area – it is these measures that are used to detect disease clusters. Thus, LISA statistics measure the degree of similarity between the disease rate for a particular place and disease rates in neighbouring areas. High values of LISA statistics indicate the presence of geographical clusters of disease. Examples of such LISA techniques include the local Moran’s I statistic and Getis and Ord’s G* statistic
Activity
View the presentation below which explains the calculation of Moran’s I statistic and look at the use of LISA in the recommended reading.
This example is based around the classification of disease rates for one area relative to the neighbouring polygons. Consider the annual number of new cases of trachoma (an eye disease) per 1,000 schoolchildren. Our example study is concerned with a whole country with an average of 25 trachoma cases per 1,000. The presentation below covers just a small part of this country – we will focus on the central polygon.
Practical Activity – Moran’s I, spatial weights, and global and local cluster detection
Download the practical instructions here in both ArcGIS Desktop and Pro and undertake the exercise, which involves computing global and local Moran’s I for some life expectancy data for the East Midlands in the UK.
If you wish to take things further (optionally – only if you are particularly interested in this topic!), visit the web site for the course textbook by Cromley and McLafferty at http://www.guilford.com/p/cromley/ and follow the link to the companion web page. Next, download ‘exercise 6: cluster analysis’ – you will find the instructions for the practical exercise are in the ‘pdfs’ subfolder and the file ending ‘v10’ contains instructions for ArcGIS version 10. The map layers for the exercise are in the ‘data’ subfolder.
References (Essential reading for this learning object indicated by *)
* Jacquez, G. M., and Greiling, D. A. (2003) Local clustering in breast, lung and colorectal cancer in Long Island, New York International Journal of Health Geographics 2, 3 Available at http://www.ij-healthgeographics.com/content/2/1/3
* See also the documentation on the Clusterseer software web site, which is designed for spatial cluster detection. The Clusterseer adviser (under ‘products’) describes the local Moran’s I statistic and Getis and Ord’s G* statistic: http://www.biomedware.com/files/documentation/clusterseer/default.htm