All study sites were located within a 600 km2 area around and including the Stability of Altered Forest Ecosystems (SAFE) project site in Sabah, Malaysian Borneo (117.5°N, 4.6°E). The area is a mixture of twice-logged lowland dipterocarp rainforest, acacia, and oil palm plantations, in which palms were planted between 1998 and 2011. Further details are given in Ewers et al. (2011).
We selected 23 focal sites along river banks. Seven were located in logged forest (areas of continuous forest at least 500 ha in size), seven in areas of continuous oil palm with no riparian reserve adjacent to the river, and eight in areas of oil palm with a riparian reserve adjacent to the river (Fig. S1 shows these sites). One site was located in Maliau Basin primary forest reserve (70 km from the SAFE project) as a reference point, but there were no other primary forest sites near enough to allow spatial interspersion of replicate primary forest sites. As all the remaining primary forest in Sabah is already protected (Reynolds et al. 2011), evaluating the ecological characteristics of large areas of logged forest versus a network of smaller forest strips is more informative for future conservation action. All our sites were separated by at least 1.5 km, and the riparian reserve sites were at least 900 m (mean distance 3.3 km, standard deviation (SD) = 2.5 km) from the logged forest boundary. Given that dung beetle movement within a 48-h period is thought to be less than 500 m (Roslin 2000; Larsen and Forsyth 2005), it is therefore unlikely that dung beetles were drawn to the riparian reserve traps from the logged forest areas.
At each site, we set up a sampling grid of 12 points, consisting of four transects perpendicular to the river (Fig. S2). Transects were 100 m apart, with sampling points at 0 m, 50 m and 100 m from the high water line. The spacing of the grid conforms to standard methods of dung beetle sampling (Larsen and Forsyth 2005). Due to variation in width of the riparian reserves (mean 49 m, SD = 30 m, referring to forest width on one side of the river), where the riparian reserve was narrow, some points in these grids fell in the surrounding oil palm area.
Data collection and analysis
All data were collected between the end of February and the beginning of July 2011. Seasonal changes in the lowland dipterocarp forests of Borneo are very limited (Walsh and Newbery 1999; Kumagai et al. 2005), and these months all fall in the slightly drier half of the year (Hamer et al. 2005).
All analyses were carried out in R (R Core Team 2012) using the packages vegan (Oksanen et al. 2013), lme4 (Bates et al. 2012) and nlme (Pinheiro et al. 2013).
Dung beetle community and land use
Dung beetles were collected using pitfall traps baited with 25 g of human dung. Human dung attracts a wide variety of species (Davis et al. 2001; Larsen et al. 2006) and is recommended as a standardized bait in tropical forests (Marsh et al. 2013). Each trap consisted of a plastic cup (8-cm top diameter, 5.5-cm bottom diameter, and 12.5-cm depth) half-filled with a solution of water, detergent, and salt. The traps were protected from the rain with a cover and collected after 48 h. The order of sites was randomized, and traps were set at no more than two sites in each 48-h period.
We could not obtain sufficient human dung to supply both the traps and the dung piles, so we used cattle dung for the dung removal experiment. Preliminary work in similar forest sites in Sabah shows that large cattle dung baits attract a similar species composition to smaller human dung baits, with the exception of some carrion feeding species found in higher abundances in human dung (Slade et al. 2011, E. Slade and D. Mann, unpubl. data). To compare species and dung removal results, we removed data on these carrion feeding species (n = 13, highlighted in Table S1) from all analyses apart from those testing for the effect of riparian reserve structural features on the entire dung beetle community.
For each sampling point (trap), we calculated dung beetle abundance, the number of functional groups present (using classifications based on diurnal vs. nocturnal activity, body length, and method of dung removal after Slade et al. (2007)), α diversity (Shannon index), and total biomass. We weighed beetles from 24 species taken from across the whole range of body sizes (between 7 and 51 individuals per species, average = 27, SD = 8) and used a polynomial regression to estimate biomass for the remaining species (Log10(mass) = −1.64 + 5.61*Log10(length) − 4.39*Log10(length)2 + 1.99*Log10(length)3, R2 = 0.982).
For each site, we calculated β diversity (mean Sørensen's similarity index) and species richness (using coverage-based rarefaction methods (Chao and Jost 2012) through the iNEXT online software (Hsieh et al. 2013). Coverage-based methods of rarefaction provide a more informative comparison of richness among multiple samples than individual or sample-based methods of rarefaction as the ratio of species richness is not compressed (Chao and Jost 2012). Rarefied species richness could not be calculated at the trap level due to four traps having only one or two beetles.
Wherever possible, we retained data at the highest spatial resolution (trap level) for analyses. For response variables where this was the case (abundance, functional group richness, diversity, and biomass), we analyzed the effect of land use (logged forest, riparian reserve, or oil palm) with generalized linear mixed models, using transect nested within site as a random factor. Where response variables could only be calculated at the site level (β diversity and rarefied species richness), we analyzed the effect of land use with a generalized least squares model. For all models, appropriate error distributions were specified and transformations or weight structures (varIdent function as described by Zuur et al. (2009)) applied where necessary. For all analyses testing for an effect of land use, we excluded data from points that fell outside of the forest strip at the narrowest riparian reserve sites so that we were carrying out a true test for differences between the three land uses. Some traps were lost due to flooding or other disturbances, so data were only obtained from 201 traps in total (82 from logged forest sites, 43 from inside riparian reserves, and 76 from oil palm sites).
Differences in community composition across land uses were explored using de-trended correspondence analysis (DCA, vegan function “decorana”), which performs well as an ordination method for displaying similarity of tropical insect communities along an environmental gradient (Brehm and Fiedler 2004). We tested for significant differences in community composition using a permutational analysis of variance (vegan function “adonis”) with 999 permutations and site as a grouping variable.
To determine whether the relative abundance of the functional groups differed between logged forest and riparian reserves, we ran a mixed model with abundance as a response variable and functional group, land use, and their interaction term as predictors.
Dung removal and land use
To record dung removal activity, uniform pats of 700 g of cow dung were set out at each sampling point (n = 12 at each site) and collected after 24 h. Large herbivores, such as the tembadau or wild cow (Bos javanicus d'Alton), Asian elephant (Elephus maximus L.), and bearded pig (Sus barbatus Müller), occur within the study area so the experimental dung pats resemble those occurring naturally. Dung removal experiments were carried out at least 1 month after pitfalls traps were collected, in order to avoid interference but also remain close enough for dung beetles assemblages to be similar (Slade et al. 2011). The order in which sites were visited was randomized. The dung was frozen for a minimum of 24 h before the experiment to kill any invertebrates already present. Data on mass loss were corrected for evaporation using estimates from three evaporation controls set at each site. For the controls, the cow dung was placed in a flat-bottomed sieve with mosquito netting sealed around the top (both 1-mm mesh), to prevent entry of any dung beetles.
The effect of land use on the mass of dung removed was analyzed with a general linear mixed model, with transect nested in site as random factors and a log transformation for the response variable. As with the data for beetle communities, for the riparian reserve sites, we only used data from within the forest strips (total n = 212: 84 from logged forest sites, 44 from within riparian reserve vegetation, 84 from oil palm sites).
We assessed whether the relationship between dung beetle community characteristics and dung removal was consistent across all land use types with a generalized linear model including land use, rarefied species richness (correlated with diversity R2 = 0.6, P = 0.004), biomass (correlated with abundance, R2 = 0.7, P = 0.0002), functional group richness, and all two-way interactions. As this analysis included coverage-based rarefied species richness all other data were averaged to the site level.
Dung beetle community structure and riparian reserve characteristics
To analyse the effect of riparian reserve width and vegetation complexity on the dung beetle community, we included data on all dung beetle species (both carrion and dung feeders). In order to test whether increasing the proportion of area in the riparian zone left as native vegetation impacts the dung beetle community, we included all the points in the sampling grid at each riparian reserve site (n = 95). Data were combined for each transect as this was the resolution at which riparian forest width could be measured (using GIS software [ArcMap version 10.1, ESRI, Redlands, CA]). A similar approach was used by Viegas et al. (2014) to test whether reserve width affects dung beetle communities in the Amazon.
To assess the vegetation structure at each sampling point, we measured humus depth, canopy density (using a spherical densitometer), and basal area (using the angle point method (Bitterlich 1984)). We estimated the height of the tallest tree to the nearest 5 m using a ruler held at arm's length and a known reference height at the base of the tree. We scored the understorey vegetation density (below 2 m) and midstorey vegetation density (between 2 m and 5 m) on an ordinal scale of sparse (fewer than 20 stems or branches), medium (20–60 stems or branches), and dense (few patches of light and 60–100 + stems or branches). To obtain one numerical index summarizing the greatest variation in these data, we ran a metric scaling analysis on all these measurements. The first axis was positively correlated with canopy density, tree height, humus depth, basal area, and midstorey density. Because this output is therefore capturing variation in the three-dimensional structure of the habitats, we refer to it as a vegetation complexity index.
We analyzed the effect of vegetation complexity on dung beetle abundance, biomass, diversity, functional group richness, and species richness using only data from sampling points falling within the riparian reserve forest. To test for any effects of reserve width or vegetation complexity, we used generalized linear mixed models, with site as a random factor and specified error families where appropriate.
Provisioning of dung removal services by riparian reserves
We analyzed the effect of riparian reserves on dung removal rates in the surrounding oil palm area in two ways. First, we compared the dung removed in oil palm adjacent to a riparian reserve (i.e., from sampling points at riparian reserve sites that fell outside the riparian forest, n = 52) and in oil palm without an adjacent riparian reserve (and also at least 50 m from the river bank, n = 56). Second, using only the data from sampling points in oil palm adjacent to a riparian reserve, we analyzed the effect of distance from the riparian reserve boundary on the mass of dung removed. We used a generalized linear mixed model with the presence/absence of riparian reserve or distance from the reserve boundary as a fixed factor for the two analyses, respectively. In both cases, we specified transect nested within site as a random factor and applied log transformations to meet model assumptions.