Literature search and compilation of data set
To identify relevant papers, we searched Web of Science for papers published between January 1945 and May 2012, using combinations of key words related to spatial habitat structure (patch AND size, isolation, connectivity, arrangement; edge effect; habitat fragmentation; shape complexity; landscape AND matrix, fragmentation, connectivity) in combination with those related to trophic interactions (food web, trophic interaction, detritivory, herbivory, predation, parasitism, disease). To this extensive list, we added papers with which we were already familiar and those citing recent reviews of habitat fragmentation.
For a publication to be included, it had to empirically evaluate how habitat fragmentation or natural spatial structure affected levels of resource consumption. Response variables included the proportion of plant material consumed, herbivores parasitised or detritus processed. Studies were included if resource consumption was measured (1) as a function of the proportion of natural habitat in a landscape; (2) in continuous habitats vs. habitat fragments (3) in contiguous vs. fragmented landscapes (fragmentation per se; sensu Fahrig 2003); (4) with respect to other measures of landscape fragmentation (e.g. amount of edge, number of patches, average connectivity); (5) as a function of patch size or degree of connectivity or isolation or (6) at the centres vs. edges of patches. To make the broadest inference possible, we include each of these spatial variables as measures of habitat fragmentation in the broad sense (Lajeunesse & Forbes 2003; Didham 2010; see also van Nouhuys 2005; Aguilar et al. 2006; Tylianakis et al. 2008). Studies were limited to those in which at least one trophic level involved terrestrial arthropods.
For each study, we recorded details of study location and design, the type of spatial variable and trophic interaction, and details about the interacting species. To compare the sensitivity of resource consumption to fragmentation across habitats or biomes, we recorded the location of each study (locality, country, latitude, longitude) and the biogeographical realm (Olson et al. 2001). We also recorded whether each study was conducted on an island or a mainland, whether it used experimental or observational methods, whether the cause of the spatial variation was natural (e.g. naturally patchy host plants), anthropogenic (e.g. due to land use change) or experimental in nature, and how long the habitat had been fragmented. Finally, we recorded the type of habitat in which the trophic interaction was evaluated, whether the habitat was composed of a single species or many, and the composition of the matrix, when possible. We categorised studies comparing large to small patches as well as those comparing continuous habitat to habitat fragments jointly as studies of patch size.
Pursuant to our focus on terrestrial arthropod systems, we categorised the studied trophic interactions into five broad interaction types (detritivory, herbivory, parasite–host, parasitoid–host and predator–prey), and recorded details of the specific measures used to quantify resource consumption. To test hypotheses about sources of variation among studies, we also recorded characteristics of the species involved in the trophic interactions. For each consumptive interaction, we defined a resource taxon and a consumer taxon and recorded trophic rank, diet breadth, habitat affinity, guild, taxonomic family and order, and taxonomic resolution (single species or entire assemblages). Trophic rank was assigned as follows: herbivores, pollinators and detritivores were considered primary consumers; parasitoids, predators and parasites of herbivores, as well as predators of unspecified trophic position were considered secondary consumers; hyperparasitoids and predators or parasites of secondary consumers were considered tertiary consumers. Omnivorous taxa in our data set were relatively rare (n = 11) and were omitted in analyses of trophic rank. Diet breadth for phytophagous insects was often given in the original publication as monophagous (feeding on one genus of plants), oligophagous (feeding on a limited number of species or genera) or polyphagous (feeding across several plant families). For predators and parasitoids, prey consumption or host use was considered specific to one or a few species, or generalist on many species. To achieve common currency across trophic levels, we considered species that were monophagous, oligophagous or specific to have a ‘narrow’ diet breadth, and species that were polyphagous, generalist or omnivorous to have ‘broad’ diet breadth. We recorded habitat affinity as ‘specific’ for species confined largely to the focal patch habitat and ‘general’ for species known to utilise both the focal patch habitat as well as the matrix surrounding the patch, according to the original publication or through literature searches. These two measures of niche breadth were highly related (χ2 = 90.14; d.f. = 1; P < 0.0001). We used habitat affinity as our measure of niche breadth for the analyses, as it was more frequently reported in the original studies. Guild was recorded as plant, herbivore, parasitoid, parasite, predator or detritivore; where available, details on subguild, such as galling herbivore or external parasite, were also recorded.
Finally, we recorded taxonomic family and order for each interacting species for two purposes: (1) to characterise whether trophic interactions for particular taxa are more sensitive to gradients in spatial structure, and (2) because many life history and behavioural traits are correlated with phylogeny. For example, aphids and butterflies use different mechanisms to locate host plant patches, and this can lead to contrasting density–area relationships for the two groups (Hambäck & Englund 2005).
We considered each combination of spatial variable and trophic interaction as a separate study in our data set. Most publications investigated the effects of multiple spatial variables on several consumer–resource pairs, and we had no a priori reasons to include particular spatial variables or consumptive interactions and exclude others. We also expected that trophic interactions within a publication may respond differently to the same spatial variable, as outlined in our hypotheses. We found that variation in effect size within publications was nearly as large as variation among publications (see Appendix S1). For example, the range of effect sizes (ES) within publications of patch size effects encompassed up to 78% of the range of ES among publications (for publications reporting > 3 ES: largest range in ES = 3.59, mean range in ES = 1.76, n = 10; for publications reporting 1 ES: range in ES = 4.61, n = 54). We therefore considered effect sizes to be independent observations for the meta-analysis (Gurevitch & Hedge 1999; see also Aguilar et al. 2006; Chaplin-Kramer et al. 2011; Magrach et al. 2014).
Calculation of effect sizes
We quantified the effects of variation in spatial habitat structure on resource consumption using the log response ratio, LRR = ln (Xe/Xc), where Xe is the mean response under the experimental condition and Xc is the mean response under the control condition (Hedges et al. 1999). For each spatial variable, we defined the control condition as the more spatially continuous condition (e.g. large patches, continuous landscapes, interior of patch) and the experimental as the more spatially disrupted condition (e.g. small patches, fragmented landscapes, edge of patch). Thus, negative effect sizes indicate that fewer resources were consumed in more fragmented settings, and positive effect sizes indicate that more resources were consumed. Where means were not reported, we digitised points from graphs and calculated means for the highest and lowest values of the spatial variable examined. When multiple comparisons were possible, we compared the most extreme treatment effects.
Effect sizes are unitless measures used to quantify the strength of the relationship between two variables (Arnqvist & Wooster 1995; Borenstein et al. 2009). The LRR is an appropriate effect size for this study because it represents the proportional change in resource consumption with fragmentation, regardless of the original unit of measurement (Hedges et al. 1999). This effect size can be calculated from studies reporting the mean responses alone, even if sample sizes and measures of variability are not reported (Lajeunesse & Forbes 2003). Furthermore, LRR is statistically well-distributed and weights deviations in the numerator and denominator equally (Hedges et al. 1999; Lajeunesse & Forbes 2003).
We also evaluated several other measures of effect size for our analysis. The relative interaction intensity effect size of Armas et al. (2004) and the difference ratio of Sorte et al. (2013) were both highly correlated with LRR (for both comparisons, r = 0.98; t417 = 95.71; P < 0.001) and were perfectly correlated with each other, as should be expected because these effect sizes are all mathematically related (Armas et al. 2004). Overall effect sizes and patterns of variation among potential explanatory variables were qualitatively similar regardless of the effect size used (data not shown), and we present all results using LRR as the effect size.
When possible, we also recorded the values for the control and experimental spatial variable, e.g. the size of the largest and smallest categories of patch size. We calculated the spatial extent of the study as the difference between ValC, the median value of the control spatial variable, and ValE, the median value of the experimental, or more fragmented, spatial variable. We categorised studies with spatial extents greater than or equal to the median value as having a ‘large extent’ and those less than the median as having a ‘small extent.’ Categorising studies based only on the maximum spatial extent (e.g. the largest patch size) typically provided the same designations of large vs. small spatial extent.