A Comparison of two Significance Testing Methodologies or the Knox Test

  • Theophilides C. N. Center for Advanced Research of Spatial Information, Hunter College of the City University of New York
  • Binkowski E. S. Department of Mathematics and Statistics, Hunter College of The City University of New York, USA
  • Ahearn S. C. Center for Advanced Research of Spatial Information, Hunter College of the City University of New York
  • Paul W. S. Epidemiology and Disease Control Programs, Chicago Department of Health, USA

要旨

The Knox method for detecting space-time interactions of point data has been widely used in infectious Epidemiology, and crime mapping. However, the parametric methods (chi-square, Poisson) used for testing the significance of space-time interaction suffer from a major flaw, the violation of the assumption of independence of pairs. A Monte Carlo method that consists of random switching of the time labels of points is also flawed because in the case of heavy clustering of either dimension (space or time) the power of the test is reduced due to the switching of already close labels. Recently, a new Monte Carlo method for testing significance has been proposed which consists of completely random sampling in space and time, i.e., an unconditional extension of the Knox test. Comparative evaluation of these statistical tests and empirical methodologies has never been conducted. Here we present the first comparative examination between the chi-square test and the random Monte Carlo unconditional extension of the Knox test. We use a space-time version of the kappa statistic for evaluation and show that the results of the unconditional Monte Carlo methodology are significantly different and superior to those of the chi-square test.
出版済
2008-09-01
セクション
Article