Publication bias
Funnel plots were produced from the final main model and all the
environmental stressor models, and therefore account for any covariates
that were retained in the model. Funnel plots were produced using the
‘funnel’ function in the metafor package . In addition, the impact of
small-study effects and decline effects/time-lag bias were tested (see
Supplementary Materials, Appendix 2).
All analyses were performed in R . The dataset used in the analysis is
available: https://doi.org/10.5281/zenodo.6903152. The R code for
data preparation and analysis, as well as the protocol for paper
screening and data extraction can also be accessed:
https://github.com/helenphillips/GCimpactsSB.
Results :
In total, 3,173 cases were available for modelling, from 627 published
articles (see Supplementary Figure 2 for PRISMA diagram). The cases were
distributed across the globe, however, large numbers of cases were from
the USA and China (Figure 1). Land-use intensification (n = 876),
pollution (n = 769), and nutrient enrichment (n = 789) were the most
represented GCs, with habitat fragmentation (n = 104) having the least
number of cases. Most cases were in relation to meso-fauna, which
included micro-arthropods (n = 1,293). However, macro-fauna and
micro-fauna were also well represented (n = 1,001 and n = 731,
respectively; Figure S3).
For the main model with the six main GCs, body size was removed from the
model, as both the interactive and the main effect. Thus, the effect of
the different GCs was the only remaining term in the model (Figure 2).
The I2 of the model was 86.31% (total variance due to
heterogeneity). 48.34% of the total variance was estimated to be due to
between-cluster heterogeneity, 35.89% due to within-cluster
heterogeneity, and only 1.06% due to the crossed effects of measurement
type. The remaining 13.69% was sampling variance.