Existing groupwise image registration algorithms for longitudinal data generally ignore continuous movements and signal changes that occur throughout image acquisition. We emphasize the case of functional magnetic resonance images, which present spatio-temporal distortion due to the combination of head motion during scanning and staggered slice acquisition. While there exist techniques to correct for motion and slice timing separately, a common dilemma is to determine which correction should be applied first. This paper proposes a four-dimensional realignment algorithm to perform both tasks simultaneously. Experiments conducted on simulated datasets with known movements suggest that the proposed algorithm provides more accurate image reconstruction than the classical two-step realignment procedure (temporal then spatial) as implemented, for instance, in the statistical parametric mapping software.