This paper assesses variations in quantitative reconstructions of late Holocene relative sea-level (RSL) change arising from using modern diatom datasets from different spatial scales, applied to case studies from Alaska. We investigate the implications of model choice in transfer functions using local-, sub-regional- and regional-scale modern training sets, and produce recommendations on the creation and selection of modern datasets for reconstructing RSL change over Holocene timescales in tidal marsh environments comparable with those in Alaska. We show that regional modern training sets perform best in terms of providing fossil samples with good modern analogues, and in producing reconstructions that most closely match observations, where these are available. Local training sets are frequently insufficient to provide fossil samples with good modern analogues and may over-estimate the precision of RSL reconstructions. This is particularly apparent when reconstructing RSL change for periods beyond the last century. For reconstructing RSL change we recommend using regional modern training sets enhanced by local samples.