Robust detection of plant species distribution shifts under biased sampling regimes

Abstract

The great diversity of terrestrial plants testifies to a wide variety of life-history strategies for dealing with the problem of surviving to reproduce in a stressful environment while competing with other biota. It is essential to develop an understanding of how global environmental changes are perturbing the distribution of species throughout the globe. However, the most abundant collections of historic data on species distributions, i.e., museums and herbaria, are affected by sample biases that strongly compromise our ability to use these data for change detection studies. Here, we present a simulation study to find robust methods for rejecting spurious shifts in the geographic range or the environmental niche occupied by species, under a variety of sampling biases. We present two methods for addressing bias. The first method is a Bayesian weighting method from machine learning theory, in which each specimen is reweighted to achieve uniform sampling intensity in time, based on sampling intensity for all specimens. An alternative method uses binomial probabilities of selecting the observed number of samples of a target relative to all specimens, and estimates the likelihood associated with the binomial probabilities using Markov chain Monte Carlo (MCMC). We find that without dealing with sampling bias, using raw data is almost certain to provide inaccurate results, under even the mildest perturbations to idealized sampling. Among the two methods for addressing bias, the empirical estimate of the probability density of a change provided by MCMC was essential for accurately rejecting false changes. The performance of the weighting method, while an improvement over the use of raw data, was limited to cases in which the bias was weak. We conclude that species distribution changes are most robustly estimated using binomial probabilities, in which the probability space is explored empirically by MCMC.