(Joint work with John F. Stanton. Can't get lingbuzz to make him an account.)
Wilson & Obdeyn (2009) claim that the use of the Observed/Expected measure (O/E) is mathematically unsound. They argue that O/E fails to reconstruct the underlying structure of a dataset, and that it should be abandoned in favor of the loglinear model. In this paper we defend O/E by arguing that the underlying grammatical structure it diagnoses, as well as the structure diagnosed by a loglinear model, are just two possible solutions to a problem that has infinitely many. Based on purely mathematical criteria, the grammar found by O/E is no more or less plausible than that found by a loglinear model.