In this paper we derive hedonic regression models for single family homes that account for nonlinearity in price functions as well as spatial heterogeneity. House prices belong to four hierarchical levels of spatial units: Census tracts (level 1), municipalities (level 2), districts (level 3) and federal states (level 4). Additionally to individual house attributes, locational covariates are available on three of these resolutions. We apply a multilevel version of structured additive regression (STAR) model on this data, which allows for nonlinear covariate effects and time trends as well as spatial effects to capture unexplained spatial heterogeneity on every level of the hierarchy. The model proposed is particularly useful for automated valuation purposes, as the decomposition of spatial effects results in improved predictive quality even in the case of unobserved spatial units. Furthermore, external price information can be integrated on various levels of the model. The presented results include nonlinear covariate effects of house and locational attributes as well as the distribution of spatial heterogeneity over Austria on maps. We also show how the inclusion of external house price information improves predictions.