Data as a Lever: A Neighbouring Datasets Perspective on Predictive Multiplicity