This paper presents a methodology for reducing the uncertainty related to the structural behaviour of an existing building in view of a vulnerability assessment regarding future earthquake actions. Estimating the lateral load resistance is a step towards evaluating the capacity of a structure to resist earthquake actions. The prediction of structural behaviour by numerical models is inevitably biased, and the importance and correlation of those modelling errors is unknown. In addition, due to the lack of knowledge on the uncertainties related to existing structures, more than one model may explain the observed behaviour. Consequently, a data-interpretation methodology that is robust in the presence of errors of unknown spatial correlation is applied to separate candidate models from unlikely ones. The first natural frequency derived from ambient vibration measurements is proposed to falsify those model instances that are inconsistent with the monitored behaviour. The identified candidate models are used to predict the lateral load resistance of a mixed concrete and masonry building studied for illustration. The number of model instances and the parameter uncertainties could be substantially reduced by interpreting ambient vibration data, showing the potential of the proposed methodology. Nevertheless, no significant reduction in the prediction range could be obtained and future work is needed to allow a meaningful assessment of the structural vulnerability.