Statistical modelling of seismic vulnerability of RC, timber and masonry buildings from complete empirical loss data

•Detailed and complete building-to-building loss data used in the study.•In total 4800 residential buildings exposed to two Mw6.5 earthquakes.•Both near fault and far field data included.•No buildings collapsed and there was no loss of life.•Novel seismic vulnerability model presented based on zero-inflated beta regression. In June 2000 two shallow, strike slip, Mw6.5 earthquakes occurred in the middle of the largest agricultural region in Iceland. The epicentres were close to small towns and villages and almost 5000 residential buildings were affected. A great deal of damage occurred but no residential buildings collapsed and there was no loss of life. Insurance against natural disasters is compulsory for all buildings in Iceland and they are all registered in a comprehensive official property database. Therefore, to fulfil insurance claims, a field survey was carried out after the two earthquakes where repair cost was estimated for every damaged building. By combing the loss data with the property database it was possible to establish a complete loss database, where all residential buildings in the affected area were included, both buildings with loss as well as buildings with no-loss. The main aim of the study was to fit a statistical vulnerability model to the data. Due to the high proportion of no-loss buildings in the database (~84%) a new and novel vulnerability model was used based on a zero-inflated beta regression model. The model was fitted to the three main building typologies in the affected region, i.e. low-rise structural wall RC, timber, and masonry buildings. The proposed model can be used to predict the mean and desired prediction limits of the losses for a given intensity level as well as to create fragility functions. All the typologies showed outstanding performance in the two destructive earthquakes, which is important to report, model and learn from.