Computational protocol: Effects of mammalian herbivore declines on plant communities: observations and experiments in an African savanna

Similar protocols

Protocol publication

[…] We examined effects of variation in wildlife abundance on plant community structure and species richness using Generalized Linear Models (GLMs). GLMs were constructed using Poisson errors and log-link functions for analysis of species richness and using Gaussian errors and identity link functions for analyses of per cent aerial cover, total cover and mean vegetation height. For each modelled response, we first constructed sets of candidate regression models in statistical software r v 2.14.2 (R Development Core Team ) using the following variables as factors: wildlife abundance (estimated by per cent wildlife dung cover), livestock abundance (estimated by per cent livestock dung cover), mean annual rainfall, recent rainfall in 3 months prior to the survey, soil sand:silt ratio and a categorical classification of ‘experimental status’ as ‘experimental’ for experimental sites (both exclosures and their paired controls) and ‘landscape’ for all other sites. Candidate models included all interactions between wildlife abundance and livestock abundance with soil, annual rainfall, and experimental status and the interaction of wildlife and livestock abundance. To compare effects of changes in wildlife abundance as opposed to all herbivore abundance, similar models were also run (separately) with all herbivores (and all interactions between herbivore abundance and abiotic gradients) rather than with wild herbivores and livestock separated. Tables with these (all herbivores) results are reported as supplementary tables and not in main text except where specified.To focus directly on the effects of wildlife decline and interactions with abiotic gradients, we performed a second analysis of effect size on the response of vegetation to wildlife declines using only the subset of sites that were (i) spatially paired across different land-use types and (ii) sampled simultaneously (n = 24 pairs, 12 landscape pairs and 12 experimental pairs). Effect sizes between high wildlife and low wildlife designations (i.e. our two categories of landscape sites) were calculated using the formula ln (plant metric low wildlife/plant metric high wildlife) (Hedges, Gurevitch & Curtis ). We examined drivers of this effect size using GLMs (as described above) with the factors experimental status, annual rainfall, soil sand:silt ratio and the interactions of these two abiotic variables with experimental status. Based both on data structure and on biological evidence for nonlinear relationships between productivity and plant responses to herbivore removal, we applied models that included a square-transformed rainfall term as well as models that included only a linear rainfall term. A square-transformed soil term (sand:silt) was initially included but was dropped due to lack of support in the models. Prior to all GLM analyses, we tested all factors and found no substantial colinearity (VIF < 2) using variance inflation factors.We compared all possible GLM models using AICc and Akaike weights (Burnham & Anderson ). Because there were multiple candidate models that received substantial empirical support (AICc < 2), we used model averaging to more directly compare competing models. With this approach, we calculated model-averaged parameter estimates for all models with ΔAICc < 2, with each model's contribution to parameter estimates being proportional to its Akaike weight (MuMIN, Barton ). To visualize the responses of plant community data, linear regressions and partial residual plots were created based on best-fit parameters across all the averaged model parameters.To examine changes in plant community composition and growth form (all species classified as either forb, grass, sedge, succulent and woody) as a function of wildlife abundance, livestock abundance and abiotic factors, we used nonmetric multidimensional scaling (NMDS) on both the species-level composition data and on life-form composition. We analysed the effects of wildlife abundance and environmental factors on plant community composition NMDS results using nonparametric multivariate anovas (npmanova; McArdle & Anderson ) and calculated P values using general permutation procedures (Manley ). We also compared best-fit models of plant community species and growth form responses using a multivariate AIC. We used the following reduced set of main factors: wildlife abundance, annual rainfall, soil sand:silt ratio and a categorical classification of experimental vs. landscape treatments. We considered all interactions of wildlife abundance with soil, annual rainfall and experimental status. As the best-fit model for both growth form and species composition had much stronger support than competing models (ΔAICc > 4), we present only the best model (r package vegan, v. 2.0–3, Oksanen et al. ). A summary of analytical approaches is provided in . […]

Pipeline specifications

Software tools MuMIn, vegan
Application Phylogenetics
Organisms Homo sapiens
Diseases Pulmonary Fibrosis