Abstract Model updating, structural optimisation and non-linear predictions are applications using families of models with varying properties. As full order model evaluations are too costly to be performed repeatedly, the possibility of using a constant basis of Ritz vectors to create parametric families of reduced models is studied. Standard model reduction methods developed for single configuration models take into account inputs and frequency range of interest but not type and range of design parameters that characterise a family of models. Nominal model reduction is shown to be insufficient in many respects and alternative methods for constructing models, which can be seen as parametrised super-elements, are studied. A plate element model of a cantilevered box beam with varying rib stiffnesses is used to demonstrate efficiency and highlight practical difficulties of the proposed approaches for predictions of static responses, modal frequencies, modeshapes, and sensitivities of those quantities to design parameters.