Summary: One of the major concerns when analysing incomplete longitudinal data is the fact that models necessarily rest on strong assumptions, unverifiable from the data. In response to these concerns, there is growing awareness of the usefulness of sensitivity analysis. We focus on repeated ordinal data. Specifically, we implement a formal approach to such a sensitivity assessment, based on local influence, in the presence of multivariate categorical data. We explore the influence of perturbing a MAR dropout model in the direction of non-random dropout, and apply the proposed method to data from a longitudinal multicentre psychiatric study.