Harvard Forest Research

Improving forecasts of species responses to climatic change: Hierarchical Bayesian analysis of tree distributions across space and time

Principal Investigator: Aaron Ellison
Harvard Forest: Apr 01 2009 - Mar 31 2011:

Abstract:
Objectives: The overall goal of this proposed research is to improve the scientific basis for projecting changes in the distributions of species by developing, comparing, and validating hierarchical Bayesian models that capture the spatially-structured, multilevel nature of species distributions while fully accounting for uncertainty.

Location: Research will be conducted at Harvard Forest, located in Petersham, Massachusetts. We will use present-day and fossil pollen-based distribution and abundance data for trees throughout eastern North America, with a focus on the northeastern region.

Questions: We will address three key questions: (1) To what extent does contemporary climate explain species distributions? (2) Do spatially-explicit models that account for spatial dependence generated by unmeasured abiotic and biotic processes predict current and historic distributions significantly better than non-spatial species-environment relationships? (3) How do forecasts differ between spatial and non-spatial models in terms of projected future distributions and their associated uncertainty?

Methods: We employ comprehensive databases describing current and historic distributions of tree species to develop and validate hierarchical Bayesian models. First, we develop and contrast non-spatial and spatially-explicit models to distinguish between species’ responses to measured climatic variables and unmeasured factors that generate spatial structure in species distributions. Second, to assess whether capturing spatial structure significantly improves forecasts of future distributions of species, we validate our models against historic changes in tree distributions as recorded in fossil pollen. Finally, we use our models to forecast potential shifts in tree distributions under climatic change and assess how these forecasts differ in terms of projected distributions and their degree of uncertainty. At all stages, we compare our models to “classical” approaches employed to model species distributions, such as generalized additive models.

Deliverables: We will reduce the considerable uncertainty associated with simple distribution models by examining how climatic and non-climatic factors shape tree distributions across different spatial and temporal scales. The net outcome will be an improved understanding of the degree to which observed species-climate relationships can be used to produce reliable forecasts of future distributions of species. We will produce spatially-explicit forecasts of changes in distributions of dominant tree species in eastern North America and will delineate areas where changes in distributions are considered most certain. Datasets, model code, and results will be made available online.
: