In a comparison of generalised linear mixed-effects models, generalised linear mixed-effects model trees and random forests, the author applies the three methodologies to a binary variable from the field of interactional pragmatics, the choice between filled and unfilled pauses across varieties of English represented by components of the International Corpus of English. Based on a large number of examples annotated for linguistic and extralinguistic factors the steps and decisions involved in the analyses are demonstrated. Though different in essence, the three resulting models share central trends. A more fine-grained evaluation of results and interpretations shows, however, that the three approaches differ in their systematicity of handling multiple observations from the same source, in that only the mixed-effects models explicitly account for and systematically partial out the relatedness of data points contributed by the same speaker. As to the way the approaches balance researcher involvement and control of the outcome, the approaches also differ substantially. A modelling choice can thus lead to notably different perspectives on an identical set of data and variables.