Ecological correlates of small mammal assemblage structure at different spatial scales in the savannah biome of South Africa

Study area

This study was conducted in Mkhuze Game Reserve (MGR; 27°35′S–27°44′S, 32°08′E–32°25′E) and the adjoining Kube Yini Private Game Reserve (KYGR; 27°42′S–27°45′S, 32°15′E–32°16′E), located at KwaZulu-Natal Province, South Africa. The two reserves have different histories. MGR was established in 1912, while KYGR was established in 1989. They also manage different species of large herbivores and predators: MGR hosts the “big five” (Du Toit et al. 2001), while KYGR only has two of them (leopards and white rhinos). The reserves fall in the savannah biome of southern Africa (Rutherford et al. 2006). The vegetation of the savannah biome is characterized by a continuous cover of perennial grasses with scattered shrubs and isolated trees (Rutherford et al. 2006). MGR covers 37,000 ha (Goodman 1990) and KYGR covers 1415 ha (Van Rooyen and Morgan 2007). The reserves are situated 40 km inland, between the Mkhuze River in the north and Phinda Private Game Reserve in the south. The Lebombo Mountains form the western border. The area is characterized by two distinct seasons: a warm and arid winter from April to September, and a hot and humid summer from October to March. Minimum annual average temperature ranges from 12°C to 20°C, and the maximum annual average temperature ranges from 25°C to 34°C. Mean average monthly rainfall ranges between 18 and 38 mm during the dry season and from 61 to 107 mm during the wet season (South African Weather Bureau, Makatini Weather Station).

We selected 20 local study sites at MGR and eight at KYGR in sample areas representing the overall habitat diversity of the reserves (Figure 1, Table 1). To choose the study sites, the maps of the reserves were divided into 1 km×1 km squares. The localizations of the sites were chosen at random within the 1 km×1 km squares falling in the major habitat types.

Figure 1

Map of the habitat types of Mkhuze and Kube Yini Game Reserves (after Van Rooyen and Morgan 2007). Black dots indicate local study sites.

Rodent and shrew sampling

We used both pitfall traps and live traps to capture rodents and shrews during the winter and summer months of 2007, 2008 (MGR), and 2009 (KYGR). At each local study site in MGR, we set up 15 live traps (Scientific Supa Kill CC) and 25 pitfall traps connected by drift fences; in KYGR, we used 15 live traps and four pitfall traps connected by drift fences. Pitfall traps consisted of 20 l buckets that were buried in the ground with the rim of each bucket at ground level. Pitfall traps were 3.5 m apart from each other and arranged at a 120° angle between each line (Figure 2). Live traps were arranged 10 m apart from each other in a 140 m line transect and at least 10 m away from the pitfall traps. Traps were checked every morning, and live traps were re-baited with a mixture of peanut butter and oats (McComb et al. 1991). Study sites were at least 500 m apart from each other to reduce the likelihood that rodents and shrews from one site would disperse to other study sites (Hurst et al. 2013).

Figure 2

Sample-based rarefaction curves of the rodent species richness (A=2007; B=2008) and the shrew species richness (C=2007, no shrew captured on site 10; D=2008, no shrew captured on site 13) at the local scale at Mkhuze. Winter and summer data are combined. See Tables 1–3 for habitat types of sites.

We identified rodents in the field by the following external characters: total length, tail length, ear length, and shape of the body (De Graaff 1981, Taylor 1998). We released individuals that could be identified after clipping their dorsal fur to avoid counting them twice. We euthanized shrews and rodent individuals that could not be identified in the field, as well as voucher specimens for each species at each site (these were hosted in the Durban Natural Science Museum, South Africa). Two specimens keyed out as Mus cf. indutus (DM 9265) and Mus cf. neavei (DM 9174) on external and cranial characters (see Lamb et al. 2014 for details) and were conservatively retained as rare species for this study. Their inclusion or exclusion did not materially affect the outcomes of the study.

All handling of living animals was carried out in accordance with guidelines of the American Society of Mammalogists (Gannon et al. 2007) and under a Durban Natural Science Museum collecting permit from Ezemvelo KZN Wildlife.

Microhabitat characteristics

We measured 17 microhabitat variables at each local study site in winter and summer. We quantified ground cover using the line-intercept method (Mueller-Dombois and Ellenberg 1974). Three 30-m transects were set along the pitfall trap lines. Every 50 cm along each transect, we recorded the following six ground cover variables: percentage bare soil, percentage plant cover, percentage rock cover, percentage shrub cover, percentage log cover, and percentage litter cover. In addition, we measured grass height to obtain a measure of vertical heterogeneity and used these data to classify grass into seven classes: % grass 0–5 cm, % grass 6–10 cm, % grass 11–20 cm, % grass 21–30 cm, % grass 31–40 cm, % grass 41–50 cm, and % grass >50 cm. We calculated the mean densities of trees and shrubs at the end of each pitfall trap line and at the center of the pitfall trap array using the point quarter method (Bonham 1989). Next, we obtained a mean value of canopy cover by measuring at those four points the amount of light coming through the vegetation at ground level using a photoelectric meter. Finally, from the center of the pitfall trap array, we visually assigned a value of slope inclination: 1=flat (∼0°), 2=intermediate (≤15°), and 3=steep (>15°).

We conducted a principal component analysis (PCA; SPSS version 15, LEAD Technologies, Inc., 2006) to reduce the number of variables and remove correlations between the microhabitat variables. Then, we used the principal components (PCs) as new microhabitat parameters in subsequent analyses. Although correspondence analysis can also be used to reduce the number of variables, this is more suitable for nominal data, and we wanted to emphasize the patterns in the actual values of the variables rather than the patterns of relative composition. We analyzed winter and summer data separately to account for the seasonal change in vegetation. We used five principal components in each season based on the percentage cumulative variance they explained.

Statistical analyses

Using EstimateS (version 8.2, Colwell 2009), we extrapolated expected species richness with two richness estimators, Chao 2 (Chao 1987) and Jackknife 2 (Palmer 1991) in order to assess the completeness of species inventories. These two estimators perform well with small numbers of samples (Colwell and Coddington 1994). By comparing the expected species richness with the observed one, the percentage completeness of a species inventory can be calculated.

Next, we plotted sample-based rarefaction curves using the software EstimateS (version 8.2, Colwell 2009) to compare species richness among study sites. The rarefaction curves standardize sampling effort at different study sites by taking into consideration the heterogeneity of the data (Gotelli and Colwell 2001).

We investigated the relationships between species richness and abundance, and PCs using general linear models (GLM, SPSS version 15, LEAD Technologies, Inc., 2006). Abundances and species richness were square root-transformed to meet the assumptions of normality and equal variances (Kolmogorov-Smirnov test, p>0.2). We used two-way analysis of variance (ANOVA) to determine the effects of season and habitat types, as well as the interaction effect between both factors, on rodent and shrew species richness and abundance. Post-hoc Tukey tests were performed on significant ANOVAs.

Using Ecosim (version 7.71, Gotelli and Entsminger 2012), we quantified species composition patterns with the following co-occurrence indices, in order to test predictions from competition theory: (1) the C-score (Stone and Roberts 1990), which measures the number of checkerboard units of all possible pairs of species (it should be significantly larger than expected by chance); (2) the number of checkerboards (Diamond 1975), which measures the number of species pairs that never coexist at any site (it should be significantly larger than expected by chance); (3) the number of unique species combinations (Pielou and Pielou 1968), which should be lower than expected by chance; and (4) the V-ratio (Schluter 1984), which measures the variability of the number of species per site and tests the niche limitation hypothesis (it should be smaller than expected by chance, Wilson et al. 1987). To test these predictions, observed co-occurrence indices were compared with 1000 expected values generated using the null models identified in MPower (see below). We considered co-occurrence patterns non-random if the observed co-occurrence indices were significantly different from 95% of the expected values. Co-occurrence patterns were analyzed in winter, summer, in both seasons and at different spatial scales. The analyses were performed within and across habitat types of MGR and KYGR, and at the broadest scale encompassing all habitat types of both reserves (MGR+KYGR).

In order to compare observed co-occurrence patterns with patterns expected by chance (Gotelli 2000), Ecosim software can be used to generate nine different null models (named SIM1 to SIM9, respectively; Gotelli and Entsminger 2012), each differing in the way rows and columns of the matrix are randomized. Because these null models may give conflicting conclusions when applied to the same data set, we used the software MPower (Ladau and Ryan 2010) to identify the most suitable SIM1 to SIM9 models for our data set, measured by size (Type 1 error), power (Type 2 error), robustness (dependence of a test’s error rates on assumptions), and bias (a measure of how much more likely the null hypothesis is to be rejected when it is false than when it is true). Species were assumed to have different probabilities of occurring at different sites and species as well as different probabilities of occurring at the same site (Ladau and Ryan 2010).

To test whether rodent and shrew assemblages were hierarchically structured, we quantified nestedness (i.e., whether species in species-poor assemblages are subsets of species in species-rich assemblages) using the nestedness temperature calculator BINMATNEST (Rodríguez-Gironés and Santamaría 2006). We also compared observed nestedness temperature with expected temperatures calculated for 1000 simulations. Nestedness temperatures range from 0 (a set of perfectly nested assemblages) to 100 (a set of completely disordered assemblages). We used Spearman rank correlations to test the respective correlations between the site rank order of study sites in the maximally packed matrix and site isolation, site area, and habitat heterogeneity. Each local study site was encompassed within a continuous unit composed of the two adjacent game reserves, MGR and KYGR, surrounded by disturbed areas. We quantified site isolation using the following indices: distance from the local study site to the nearest and the farthest borders of the unit, distance from the local site to the edge of the habitat patch where the site is found, distance from the site to the nearest patch of the same habitat as the one where it is found, and sum of the pairwise distances between sites (Cullingham et al. 2008). Furthermore, we quantified site area with two indices, namely, size of the habitat patch where the site is found and area of this habitat within the unit. The indices were measured with ArcMap (version 9.3, ESRI Inc., 2008) using the “Measure” tool. We quantified habitat heterogeneity with six indices measured at macrohabitat and microhabitat scales, including number of habitats adjacent to the habitat patch where the site is found and the principal components of the microhabitat variables, among others. Nestedness was analyzed within and across habitat types of each reserve, and across all habitat types of both reserves (MGR+KYGR).

Species richness and abundance

At MGR, a total of 215 rodents representing 14 species and ten genera in two families, Muridae and Nesomyidae, were captured over 102 trapping nights (Table 2). The two most common rodent species caught were Mus minutoides and Mastomys natalensis, representing 59% of all the captures. The least abundant species were Steatomys krebsii, Mus cf. neavei and M. cf. indutus. The species richness estimators indicated that our species inventory was between 64% (Chao 2) and 70% (Jackknife 2) complete. Sample-based rarefaction curves indicated that rodent species richness was the highest in the Acacia woodlands and the lowest in sand forests (Figure 2).

At MGR, we captured 96 shrews representing four species and two genera from the family Soricidae over 102 trapping nights (Table 2). The two most commonly species caught were Crocidura fuscomurina and C. hirta, representing 73% of all captures. Crocidura silacea was the least abundant species caught. Both estimators indicated that the species inventory of shrews was 100% complete. Species richness at a local scale was the highest in the Acacia woodlands and the lowest in sand forests (Figure 3

Figure 3

Sample-based rarefaction curves of the rodent species richness (A: no rodent captured on site 3) and the shrew species richness (B: no shrew captured on site 4) at the local scale at Kube Yini in 2009. Winter and summer data are combined. See Tables 1–3 for habitat types of sites.

).

At KYGR, a total of 63 rodents representing six species and six genera belonging to two families, Muridae and Nesomyidae, were captured over 20 trapping nights (Table 3). The two most common rodent species caught were Mus minutoides and Aethomys ineptus, representing 70% of all the captures. The least abundant species were Lemniscomys rosalia and Mastomys natalensis. The species richness estimators indicated that our rodent species inventory was between 83% (Jackknife 2) and 100% (Chao 2) complete. Meanwhile, sample-based rarefaction curves indicated that species richness was the highest in Lebombo wooded grasslands and the lowest in riverine woodlands (Figure 3).

At KYGR we captured 21 shrews representing four species and two genera from the family Soricidae over 20 trapping nights (Table 3), these included Crocidura hirta (six individuals), Suncus lixus (six individuals), S. infinitesimus (five individuals), and C. silacea (four individuals). Both estimators indicated that the species inventory of shrews was 100% complete. Suncus infinitesimus was collected at KYGR but not at MGR, while Crocidura fuscomurina was collected at MGR but not at KYGR. Species richness at a local scale was the highest in Lebombo wooded grasslands and the lowest in Ziziphus mucronata bushland (Figure 3).

At identical sampling efforts (i.e., for the same number of individuals caught), rodent species richness was higher at MGR than at KYGR, while shrew species richness was higher at KYGR than at MGR (Figure 4). There were also significant differences, but only in rodent species richness and abundance between seasons (F_1,8=7.05, p<0.05; F_1,8=7.80, p<0.05, respectively). Post-hoc Tukey tests revealed that rodent species richness and abundance were the highest in winter.

Figure 4

Sample-based rarefaction of the species richness of rodents (A) and shrews (B) at the reserve scale (i.e., Mkhuze or Kube Yini). Winter and summer data are combined.

Correlation with microhabitat characteristics

The PCA of the 17 microhabitat variables extracted five principal components that accounted for 79.48% of the total variance in winter and 77.11% in summer (Table 4). We interpreted the PCs as follows. In winter, PC1 was a measure of differences in the vertical height of grass and % plants: local study sites with a high % grass height 31–40 cm and 41–50 cm, % plant cover loaded high on the axis and sites with a high % grass height 0–5 cm loaded low on the axis. PC2 was a measure of differences in tree density and % litter: local study sites with a high density of trees and high % litter loaded high on the axis. PC3 was a measure of differences in canopy cover and % grass height >50 cm: local study sites with a high canopy cover loaded high on the axis and sites with a high % grass height >50 cm loaded low. PC4 was a measure of differences in the percentage of rocks: local study sites with a high percentage of rocks loaded high on the axis. Finally, PC5 was a measure of differences in the percentage of bare soil: local study sites with a high percentage of bare soil loaded low.

In summer, PC1 was a measure of differences in the vertical height of grass: local study sites with a high % grass height 31–40 cm and 41–50 cm loaded high on the axis, and those with a high % grass height 0–5 cm loaded low on the axis. PC2 was a measure of differences in % logs and % litter: local study sites with high % logs and high % litter loaded high on the axis. PC3 was a measure of differences in canopy cover and % grass height >50 cm: local study sites with a high canopy cover loaded high on the axis and sites with a high % grass height >50 cm loaded low. PC4 was a measure of differences in the percentage of shrubs and % grass height 6–10 cm: local study sites with a high % of shrubs loaded high on the axis and sites with a high % grass height 6–10 cm loaded low. Finally, PC5 was a measure of differences in tree density and % grass height 11–20 cm: local study sites with a high tree density loaded high on the axis and sites with a high % grass height 11–20 cm loaded low.

Only in winter, rodent species richness was significantly correlated with PC1 (r₂₇=0.4, p<0.05) and PC4 (r₂₇=0.6, p<0.05), and rodent abundance was significantly correlated with PC1 (r₂₇=0.6, p<0.05). Similarly, only in winter, shrew species richness was significantly correlated with PC1 (r₂₇=0.3, p<0.05), PC2 (r₂₇=0.4, p<0.05), and PC4 (r₂₇=0.5, p<0.05).

Non-random co-occurrence patterns

The influence of competition on rodent and shrew assemblages was tested with four co-occurrence indices and seven algorithms (Table 5). Based on the assessments of the quality of simulations, co-occurrence patterns within habitat types could only be tested in the Acacia woodlands and not in the shrew assemblages at KYGR.

In winter, we found non-random patterns consistent with competition theory in rodent assemblages at MGR and MGR+KYGR with SIM1 in combination with the number of checkerboards (Table 5). This algorithm allows the number of species at a site to vary, with the constraint that all sites have the same average number of species, and occurrence frequencies of each species vary with the same probability. In both seasons, we found non-random patterns consistent with competition theory in rodent assemblages at MGR and MGR+KYGR with SIM1 in combination with the number of checkerboards. We did not find patterns of competition in summer.

Rodent assemblages were significantly nested at MGR in winter, summer, and both seasons combined; they were significantly nested at KYGR in both seasons combined; and they were significantly nested at MGR + KYGR in summer and both seasons combined (Table 6). However, results from the correlation analyses were only significant at the broadest scale MGR+KYGR with both seasons combined. The site rank order in the maximally packed matrix was significantly correlated with the distance from the local study site to the nearest (r₂₇=0.4, p<0.05) and the farthest (r₂₇=0.4, p<0.05) borders of the unit formed by the two reserves, the distance from the local study site to the nearest patch of the same habitat as the one where the local study site can be found (r₂₇=0.5, p<0.05), the sum of the pairwise distances between sites (r₂₇=0.4, p<0.05), the size of the habitat patch where the local study site can be found (r₂₇=0.4, p<0.05), the size of this habitat in the unit formed by the two reserves (r₂₇=0.4, p<0.05), the PC2 (r₂₇=0.5, p<0.05), and the PC4 (r₂₇=0.4, p<0.05) of the microhabitat variables measured in winter.

We found non-random patterns consistent with competition theory in shrew assemblages in winter at MGR with SIM4 in combination with C-score, and in both seasons at MGR+KYGR with SIM1 in combination with the number of checkerboards (Table 5). In SIM4, the species occurrence totals were maintained, and the probabilities of occurrence in sites were proportional to the observed species richness per site.

Shrew assemblages were only significantly nested across habitats at MGR+KYGR when both seasons were combined (Table 6). The site rank order in the maximally packed matrix was only significantly correlated with the PC3 of the microhabitat variables measured in summer at the broadest scale MGR+KYGR (r₂₇=0.5, p<0.05).