Field-Based Response Data
We obtained ecosystem field data from plot surveys of forested sites
following methods detailed in Keys et al. (2023). Surveys involved
measuring biotic (e.g., species coverage) and abiotic (e.g., humus
depth) forest ecosystem properties (Figure 1) (additional survey
information in S1). Following surveys, we calculated several aggregate
ecosystem properties from plot data, including biotic (e.g.,
productivity) and abiotic variables (e.g., brown carbon stocks) (Figure
1; S1). For each of our three models, we employed a response data
selection workflow (see S1) to satisfied model assumptions and to
account for data errors or survey biases (e.g., missing data
attributes).
We generally sought to build parsimonious models incorporating
constituent and aggregate ecosystem properties drawn from biotic and
abiotic domains. In our ecosystem model (see S1), we strove to ensure
the number of biotic and abiotic response variables were relatively
similar. For example, representing biota at the species level (664 taxa)
would have skewed the ratio of biotic to abiotic response variables in
this model, because fewer abiotic variables were available in the plot
database. This imbalance would unduly weight the influence of biotic
variation on model outcomes. To help address this inequity, we
represented biota with ten functional groups (see S1) adapted from the
Canadian National Vegetation Classification (Baldwin et al. 2019).
Lastly, different measurement units (e.g., percent cover for species,
centimeters for humus depth) were employed for quantifying the relativeabundance of biotic and abiotic properties (see S1). Mixing
these variables introduced dissimilar data structures in the response
data pool, with a potential bearing on model outcomes. We applied a cube
root transformation to a subset of response variables (see S1) (Cox
2011) to help resolve this issue.