In this paper we present results from the first application of robust geostatistical modeling techniques to radio environment and coverage mapping of wireless networks. We perform our analysis of these methods with a case study mapping the coverage of a 2.5 GHz WiMax network at the University of Colorado, Boulder. Drawing from our experiences, we propose several new methods and extensions to basic geostatistical theory that are necessary for use in a radio mapping application. We also derive a set of best practices and discuss potential areas of future work. We find that this approach to radio environment mapping is feasible and produces maps that are more accurate and informative than both explicitly tuned path loss models and basic data fitting approaches.