A cellular association trace consists of timestamped events recording user activity in labeled cells in a cellular network. From such data one can infer that if a user appears in two different cells within a short span of time, that a handover took place, and that the coverage areas of the two cells overlap. That is, one can infer geographic information from handover behavior. One would like to expand this kind of inference to a larger scale, perhaps reconstructing a proximity graph of the cellular sites, or creating an approximate 2-dimensional embedding of the cells. We have analyzed a large-scale cellular association trace of several months of activity for several million users on a 3G network, and have found that handover behavior is actually incredibly diverse and complicated, making it very dicult to make any sort of global inferences, even in small sections of a network. In this paper we present some stable elements of handover behavior, and present several methods one can use to extract proximity information from such a trace.