Fitting and assessing models of interspecies competitive effects on the growth of trees with the forestecology R package

Albert Y. Kim, Statistical & Data Sciences, Smith College, Northampton, MA, David N. Allen, Biology, Middlebury College, Middlebury, VT and Simon P. Couch, Mathematics, Reed College, Portland, OR

Background/Question/Methods
Many models for the growth of trees that incorporate interspecies competition are based on a neighborhood effect assumption whereby all trees within a fixed distance of a focal tree are considered competitors. Methods and tools are needed to quantify this competitive effect and assess the quality of all resulting models.
We present the forestecology R package providing methods for both 1) evaluating the effect of competitor species identity using permutation tests and 2) evaluating model performance while accounting for the spatial autocorrelation of forest data via spatial cross-validation. The package features 1) tidyverse-like structure whereby verb-named functions can be modularly "piped" in sequence, 2) functions with standardized inputs/outputs of simple features sf package type that encode spatial vector data, and 3) an object-oriented implementation of the Bayesian linear regression model in Allen (2020).
We demonstrate the package's functionality using data from the Smithsonian Conservation Biology Institute's (SCBI) large forest dynamics plot in Front Royal VA USA, part of the ForestGEO global network of research sites.
Results/Conclusions
We demonstrate that for SCBI forest data both 1) competitor species identity matters when modeling the growth of trees and 2) that model error estimates that do not incorporate the inherent spatial autocorrelation of forest data are overly optimistic.
The package was designed with the following in mind: clear articulation of all steps in the sequence of analysis, easy wrangling and visualization of geospatial data, and cross-site compatibility given ForestGEO's data collection protocols and data formatting standards. While the package currently only has a Bayesian linear regression model implemented, it can easily be extended to other modeling methods.