Lost in the woods: shifting habitats can lead phylogeography astray

Continuous phylogeographic inference is a popular method to reconstruct the spatial distribution of ancestral populations and estimate parameters of the dispersal process. While the underlying probabilistic models can be complex and their parameters are often computationally demanding to infer, these models typically ignore that replication and population growth are tightly coupled to spatial location: populations expand into fertile uninhabited areas and contract in regions with limited resources. Here, I first investigate the sampling consistency of popular summary statistics of dispersal and show that estimators of "lineage velocities" are ill-defined. I then use simulations to investigate how local density regulation or shifting habitats perturb phylogeographic inference and show that these can result in biased and overconfident estimates of ancestral locations and dispersal parameters. These, sometimes dramatic, distortions depend in complicated ways on the past dynamics of habitats and underlying population dynamics and dispersal processes. Consequently, the validity of phylogeographic inferences, in particular when involving poorly sampled locations or extrapolations far into the past, is hard to assess and confidence can be much lower than suggested by the inferred posterior distributions.