Nonrigid registration (NRR) is routinely used in the study of neuroanatomy and function and is a standard component of analysis packages such as SPM. There remain many unresolved correspondence problems that arise from attempts to associate functional areas with specific neuroanatomy and to compare both function and anatomy across patient groups. Problems can result from ignorance of the underlying neurology which is then compounded by unjustified inferences drawn from the results of NRR. Usually the magnitude, distribution, and significance of errors in NRR are unknown so the errors in correspondences determined by NRR are also unknown and their effect on experimental results cannot easily be quantified. In this paper we review the principles by which the presumed correspondence and homology of structures is used to drive registration and identify the conceptual and algorithmic areas where current techniques are lacking. We suggest that for applications using NRR to be robust and achieve their potential, context-specific definitions of correspondence must be developed which properly characterise error. Prior knowledge of image content must be utilised to monitor and guide registration and gauge the degree of success. The use of NRR in voxel-based morphometry is examined from this context and found wanting. We conclude that a move away from increasingly sophisticated but context-free registration technology is required and that the veracity of studies that rely on NRR should be keenly questioned when the error distribution is unknown and the results are unsupported by other contextual information.