Global experiment and arthropod samples
Experiments at all 52 sites were conducted following a standardized protocol. Each experiment followed a randomized block design; for each site, we randomly chose 10 to 20 paired trees, shrubs or herbs (blocks). Paired plants were at least 2 m apart from each other. Each block was at least 6 m apart from each other. In most sites, the experiment was conducted using a single native plant species, typically the most common species locally. However, in some tropical forests, due to the high species diversity and low relative density, we used different plant species among blocks, but the plants within each pair always belonged to the same species. We only used broadleaf native plant species that did not bear any apparent type of indirect defence (e.g., domatia, extra-floral nectaries, glandular trichomes) (Romero and Koricheva, 2011).
One plant in each pair was randomly chosen as a control whereas the second plant was used for the leaf-rolling (shelter addition) treatment. On each plant, we selected 5-10 fully expanded leaves without apparent damage. Prior to the experiment, any arthropods present were removed by hand from leaves. Then, the leaves on the treatment plant were manually rolled from the adaxial to the abaxial surface transversally to the leaf axis to form a cylinder of roughly 0.6 cm in diameter (Fig. S1) to mimic shelters built by caterpillars from at least 17 moth and butterfly families, including Hesperiidae, Nymphalidae, Gelechiidae, Oecophoridae, Lasiocampidae, Pyralidae, Gracillariidae, Tortricidae, Geometridae, Erebidae (Fitzgerald et al ., 1991; Fukui, 2001; Lill and Marquis, 2007). The rolls were secured with a metal hairpin (see Fig. S1 and Vieira and Romero, 2013). The control plant of the pair had 5-10 unrolled leaves marked with a metal hairpin. Rolled and control leaves were exposed for 10 days in the field. This was deemed sufficient as previous bioassays showed that leaf shelters can be colonized very quickly (within 24 hours) (Vieira and Romero, 2013). Maximum width of both control and rolled leaves was measured as an estimate of leaf size.
After 10 days of the experiment, we collected rolled and control leaves, and stored them grouped per replicate and treatment in small zip-lock plastic bags. The leaves were either frozen for later sorting or immediately sorted to collect the invertebrates. We collected all the invertebrates visible to the naked eye (except mites) and stored them in ethanol. We identified the invertebrates to the lowest taxonomic level possible and classified them into morphospecies and feeding guild (i.e., predator, parasitoid, herbivore, detritivore, omnivore). Individual body size (dry body mass) was estimated from the dry mass (dried at 70oC for 24h) or by measuring total body length and then calculating the dry mass through published taxon-specific allometric equations (Hódar, 1996). Four dependent variables were used for analyses: arthropod abundance, species richness, biomass and mean body size. Arthropod biomass represented the sum of all individual body masses. All variables were weighted by the number of sampled leaves per plant.