Summary. According to the hypothesis of configural encoding, the spatial relationships between the parts of the face function as an additional source of information in the facial perception of emotions. The paper analyses experimental data on the perception of emotion to investigate whether there is evidence for configural encoding in the processing of facial expressions. It is argued that analysis with a probabilistic feature model has several advantages that are not implied by, for example, a generalized linear modelling approach. First, the probabilistic feature model allows us to extract empirically the facial features that are relevant in processing the face, rather than focusing on the features that were manipulated in the experiment. Second, the probabilistic feature model allows a direct test of the hypothesis of configural encoding as it explicitly formalizes a mechanism for the way in which information about separate facial features is combined in processing the face. Third, the model allows us to account for a complex data structure while still yielding parameters that have a straightforward interpretation.