The seismic analysis of existing unreinforced masonry buildings is a challenging task, troubled by different sources of material and modelling uncertainties. The historical heritage value of a building can further complicate the assessment and design of retrofit interventions as any kind of intervention needs to respect limits imposed from the conservational perspective. At the same time earthquakes worldwide keep causing unacceptable losses and damages, reminding us about the sensitivity of this building typology to the seismic loading. Across historical city centres of Europe, masonry buildings are often part of building aggregates, which developed when the layout of the city or village was densified. In these aggregates, adjacent buildings can share structural walls to support floors and roofs. Meanwhile, the masonry walls of the façades of adjacent buildings are often connected by dry joints since adjacent buildings were constructed at different times. Observations after, for example, the Central Italy and Croatia earthquakes showed that the joints between the building units were often the first elements to be damaged. It is hypothesised that, due to the lacking interlock, the joints opened up leading to pounding between the building units and a complicated interaction at floor and roof beam supports. The analysis of such building aggregates is very challenging and modelling guidelines are missing. Advances in the development of analysis methods have been impeded by the lack of experimental data on the seismic response of such aggregates. This leads to such buildings often modelled as separated or fully connected. The first part of this thesis is concentrated on enhancing the understanding of the seismic behaviour of such aggregates. By testing a large-scale unreinforced stone masonry aggregate, consisting of two units, the necessary experimental data was generated and confirmed the hypotheses about significant interaction at the interface between the units. A blind prediction competition accompanying the experimental campaign was organized with participants coming from both the industry and academia. This was among the only very few (to our knowledge three) blind prediction competitions on the seismic behaviour of unreinforced masonry buildings (i) the first to address a building developing in-plane and out-of-plane failure modes, (ii) the first that used bi-directional shaking as input and (iii) the first on a masonry aggregate. Modelling approaches among the blind prediction participants were well distributed with regard to the level of detail and modelling assumptions. The scatter between the predictions was very large, although all the participants were given detailed information on the material properties including the results of cyclic shear-compression tests. It can therefore be assumed that the scatter can be largely attributed to the modelling uncertainties. We participated with an equivalent frame model in the blind prediction competition. To model the interaction between units of an aggregate using the equivalent frame approach, a new material model was developed and implemented in OpenSEES, coupling normal and shear behaviour, and allowing us to model interaction in terms of separation, pounding, and shear. Recent advances in macroelement modelling using the equivalent frame approach have enabled to simultaneously model both the in-plane and out-of-plane behaviour.