Testing the performance of S-curves for language change

Abstract This article investigates the effectiveness of sigmoid trajectories (S-curves) for statistically modelling language change. In diachronic linguistics, it is customary to fit a sigmoid regression line through data, on the assumption that such a curve fits the ideal language change. A coefficient with an associated p-value significantly different from zero is taken as evidence for the presence of a change. Here, we take the inverse perspective: given a known change, how well can the S-curve predict the actual data? We look at 15 well-known changes in Late Modern Dutch, in a genre-balanced corpus. For each change we built four different models: one model based on all the data and three partly blinded models, where either the beginning, middle or end of the S-curve is omitted. We check how well the S-curve can predict or reconstruct the unknown data, i.e. the blinded parts of the S-curve. We investigate in which cases (the type of change, the part of the data that is omitted, etc.) it is easier or harder to reconstruct the missing data.