ABSTRACTTidal marsh sediments from south‐central Chile provide evidence for multiple great earthquakes. Diatom transfer functions, statistical models of the relationship between species preserved in the sediment and elevation, provide quantitative estimates of coseismic vertical land‐level change associated with individual earthquakes. However, in south‐central Chile, our ability to quantify land‐level change is currently limited by a lack of understanding of the environmental variables controlling the distribution of diatoms, an essential prerequisite for converting variations in fossil diatom assemblages into quantitative estimates of past elevation changes. We present a new modern diatom dataset for the region and explore the implications of the scale of the dataset used in transfer function models on the reconstructions of land‐level change. Modern training sets containing samples from a regional scale are superior to sub‐regional and local‐scale training sets, providing closer estimates for known deformation during the great 1960 Chilean earthquake, a higher proportion of good modern analogues and uncertainty terms up to 42% smaller than previously published reconstructions. Copyright © 2017 The Authors. Journal of Quaternary Science Published by John Wiley & Sons Ltd.