Location estimation has been a backbone for location-aware services as wireless networks and mobile devices are more pervasively available. By operating on the signal strength space, nearest neighbor methods like RADAR have proved to be simple yet effective for location estimation. It has been common to take locations as classes, and then to infer location classes based on signal strength measurements. Under such a location classification setting, this paper investigates in detail the k-nearest neighbor approach in RADAR, and demonstrates that considering more neighboring signal strength measurements usually cannot help. Instead the orientations in which the signal strength is taken should be more carefully treated. This paper also develops a refinement step for RADAR, by building nearest neighbor classifiers to further clarify several top location estimates by RADAR. At a very economic cost, our refinement step can significantly boost the accuracy.