*2.1. Study Location*

Our analysis is based on the 1.2-km<sup>2</sup> Mpala ForestGEO plot (0◦17- N, 36◦53- E) [22]. The plot is in a semi-arid savanna (~600 mm annual rainfall) that supports at least 17 wild large herbivore species (>5 kg). The ForestGEO plot includes at least 62 woody plant species out of >460 plant species that occur in the region [22,23]. It spans three habitats characteristic of the Laikipia Highlands: (*i*) 'plateau' habitat on poorly drained and nutrient-rich clay vertisols (black-cotton soil; 1775–1792 m asl); (*ii*) 'low plain' habitat on well-drained,

nutrient-poor, red sandy loams (red sands soil; 1669–1779 m asl); and (*iii*) a rocky 'slope' between the plateau and low plain (1679–1779 m asl).

We analyzed communities of woody plants and large herbivores at 33 sites within the ForestGEO plot (Figure 1). The sites were positioned at regular 100-m intervals across the plot, with 14 occurring in the plateau, 3 on the slope, and 16 in the low plain. Across the plot, we also measured fine-scale topographic habitat variables, including elevation, slope, and convexity, based on elevation data recorded in a 5 m × 5 m grid [22]. We associated our topographic values with a set of 20 m × 20-m vegetation quadrats using the *fgeo.analyze* package [24] in R [25]. After standardizing measures of elevation, slope, and convexity to a mean of 0 and standard deviation of 1, the three variables were used to classify the 3000 20 × 20-m quadrats into the three major habitats (plateau, slope, and low plain) using Ward hierarchical clustering (Figure 1). Finally, we calculated the topographic wetness index (TWI), which reflects the ability of a landscape to retain water and is a strong predictor of savanna wildlife distributions [26], by integrating the total water catchment area and slope of each grid cell using *build.layers* in the *dynatopmodel* package [27]. Cells with high TWI tend to be flat or concave.

**Figure 1.** The Mpala ForestGEO ecosystem. (**a**) The map shows the extent of the study plot, with white circles representing 33 sampling sites. The three habitats are distinguished by color (purple = "plateau", teal = "slope", yellow = "low plain"). (**b**) The locations of six woody plant species comprising 80% of stems in the plot are shown with different color points. Phylogenies show relationships between (**c**) woody plants and (**d**) large mammalian herbivores; scale bars represent 1 MY. Note *Hibiscus aponeurus* in the phylogeny represents *Hibiscus* spp. in the ForestGEO data; see Table S1 in Supplementary Materials for description of grafted or substituted taxa.

### *2.2. Woody Plant Distribution and DNA Barcoding*

The first comprehensive ForestGEO survey of woody plants began in 2010. It established a regular grid of 400-m<sup>2</sup> quadrats in which the main and auxiliary stems of woody trees and shrubs > 0.5 m tall were geolocated, tagged, and measured for diameter at knee height (dkh) [22]. Species were identified by researchers from the East African Herbarium

at the National Museums of Kenya. We obtained the complete dataset from the ForestGEO portal (12 March 2019) [22]. It included 363,798 total stems and 139,078 main stems (henceforth 'individual trees') representing 67 species and 22 families. The branching architecture of shrubs such as *Croton* and *Euclea* can make it difficult to identify discrete individuals, but we assumed the data were internally consistent. The dataset was filtered to include only living stems > 2 cm dkh, species with >2 main stems in the plot, and sufficient identification for phylogenetic analysis. The filtered data retained 355,461 stems, 136,297 individual trees, and 55 morphospecies (Table S1). We extracted a dataset for analysis that focused on trees within 25 m<sup>2</sup> of our 33 grid sites.

Our plant phylogeny was based on an extensive plant DNA barcode library and phylogeny for Mpala, which was constructed using a supermatrix approach [15,23,28]. The full DNA barcode library includes high-quality data from 1760 specimens representing at least 438 species sequenced at up to 5 markers (*mat*K, *rbc*L, *psb*A-*trn*H, *trn*L-F, and ITS). A subset of species missing from the phylogeny were grafted in three complementary ways. First, we obtained new *trn*L and *rbc*L DNA barcodes from 5 species [23], and we used these data to determine how to graft them into the phylogeny (Table S1). Second, we represented taxa with substitutes that were already in the phylogeny (e.g., congeners, such as *Hibiscus aponeurus* used to represent *Hibiscus* spp.; Figure 1, Table S1). Third, we grafted remaining species based on the literature (see Table S1 for details).

### *2.3. Large Mammalian Herbivore Community Data*

To assess herbivore distributions, we deployed camera traps from March 2018 to April 2019 (Bushnell, #11-9874C). We recorded date, time, and species using Wild-ID software [29]. Photos of large herbivores were extracted and filtered to independent detections (defined as >30 min apart) to reduce the impacts of temporal autocorrelation. For each species, we calculated a relative abundance index (RAI) as the total number of independent photographs divided by the total number of working camera days over the course of the survey. Simple RAI-based approaches yielded relative abundance estimates that correlated strongly with independent estimates of animal abundance for large mammals [30].
