Don Edwards

Department of Statistics

University of South Carolina

Alternatives to the Paired T-Test for Testing Spatial Change

Testing for change in a region, e.g. loss of wetlands over a period of several years, is a frequent goal for assessment of conservation strategies. Typically, using digital image processing techniques in conjunction with Geographic Information Systems (GIS) tools, the region is subdivided into cells and percent wetlands estimated for each cell both before and after the period in question. The change in this percentage in each cell can thus be calculated and used as a basis for a formal statistical test. These cell changes are usually spatially correlated, however, which can invalidate a simple paired t-test. For example, when the region is subdivided into smaller cells, results using the paired t-test tend to artificially become more statistically significant. Using both theory and computer simulation, we examine the operating characteristics of the paired t-test and several modified versions of it which attempt to correct for spatial correlation using results from geostatistical (kriging) analyses. The simple test performs poorly under spatial correlation, whereas the alternative tests hold their level well in these instances and have comparable power to the simple t-test when there is no correlation. An example application to Murrells Inlet, SC data is provided.

Back to Colloquium Series