Accounting for spatial processes is an important aspect of modelling data which are associated with a geographical location. Failure to do so can compromise a model’s performance. These processes can operate in different ways depending on the underlying mechanisms at play. It is reasonable, for example, to expect that the characteristics of a single agricultural field should be closely related to those of other fields within the same farm, tended to by the same farmer. Similarly, various such farms in a single area of governance may be subject to one set of regulations leading them to more closely resemble each other than farms in another jurisdiction. By capturing this spatial hierarchy of field within farm within jurisdiction in a model, we would expect the model to perform better. In another sense, it also seems reasonable to expect that fields which are geographically close to each other should be more similar than distant fields. They may be subject to similar climate, soil composition and local traditions of land use, so even if they are operated by different farmers on either side of a fence, their characteristics are not independent. Here we would expect that a spatially autoregressive model which accounts for this should perform better. This thesis develops models which combine both of these types of spatial processes. Rather than looking at fields and farms, we instead focus on voter behaviour in individual constituencies across the UK, all of which are nested within counties and regions. We begin by applying such a modelling structure, accounting for both of these types of spatial effect, to the 2019 UK General Election in England and Wales. Using this methodology, we can examine the proportion of variation in behaviour which is attributable to different levels of grouping, and estimate spatially varying coefficients. A key component of such modelling is the construction of neighbourhood matrices which encode whether or not spatial units are to be considered as neighbours and thus more likely to share similarities than other units. We present an R package, sfislands, which reduces the workload in creating such matrices when complications occur due to the presence of islands or other geographic sources of discontiguity. We conclude by applying a novel methodology to the 2024 UK General Election, which seeks to capture both of the above spatial effects in a different way. The proposed model allows the degree to which neighbouring constituencies are expected to be similar to vary according to hierarchical position. By comparing the plausibility of this framework to other candidate combinations of spatial structure, we find that this model represents the more plausible explanation of the underlying spatial processes of party support in this election.