Population size history from short genomic scaffolds: how short is too short?

The Pairwise Sequentially Markov Coalescent (PSMC), and its extension PSMC9, model past population sizes from a single diploid genome. Both models have been widely applied, even to organisms with scaffold-level genome reference assemblies of limited contiguity. However it is unclear how PSMC and PSMC9 perform on short scaffolds. We evaluated psmc and msmc, implementations of the PSMC and PSMC9 models respectively, on simulated genomes with low contiguity, and compared results to those from fully contiguous data. Simulations with scaffolds from 100 Mb to 10 kb revealed that psmc maintains high accuracy down to lengths of 100 kb, while msmc is accurate down to 1 Mb. The discrepancy is not due to differing models, but stems from an implementation detail of msmc---homozygous tracts at the ends of scaffolds are discarded, making msmc unreliable for low contiguity genomes. We recommend excluding data that are aligned to shorter scaffolds when undertaking demographic inference.