LAUSR

Abstract This article explores questions on perceptions of the location of the ‘Midwest’, a contested vernacular region of the United States. We created a custom online survey with R’s web framework Shiny, in which participants were presented with a blank web map and asked to ‘draw’ their definition of the Midwest. Instead of simply describing the aggregated results, we employ machine learning algorithms – Naive Bayes, Random Forest and Categorical Boosting – in an attempt to classify users into groups, with a focus on the features that most effectively separate responses. We also demonstrate a way to engineer features from a single spatial response question and provide an implementation through a small R package. Furthermore, we discuss misclassified observations and suggest some driving factors in the construction of regional perception. This research is important not only for its contribution to perceptual regions but also for the approach, which could be applied to place-based survey analysis more broadly.