We propose a method for the statistical analysis of fMRI data that tests cluster units rather than voxel units for activation. The advantages of this analysis over previous ones are both conceptual and statistical. Recognizing that the fundamental units of interest are the spatially contiguous clusters of voxels that are activated together, we set out to approximate these cluster units from the data by a clustering algorithm especially tailored for fMRI data. Testing the cluster units has a two-fold statistical advantage over testing each voxel separately: the signal to noise ratio within the unit tested is higher, and the number of hypotheses tests compared is smaller. We suggest controlling FDR on clusters, i.e., the proportion of clusters rejected erroneously out of all clusters rejected and explain the meaning of controlling this error rate. We introduce the powerful adaptive procedure to control the FDR on clusters. We apply our cluster-based analysis (CBA) to both an event-related and a block design fMRI vision experiment and demonstrate its increased power over voxel-by-voxel analysis in these examples as well as in simulations.