study area
In Argentina, the Yungas extend along discontinuous mountain ranges between 22° to 29°S, and include three altitudinal forest belts: premontane lowland, lower montane and upper montane forest (Cabrera 1976). The northern premontane lowland forest type is the most speciesrich as well as the most endangered because of clearcutting (Grau & Brown 2000; Brown et al. 2001). This study took place in grapefruit plantations adjacent to altered premontane lowland forest in the upper Bermejo River Basin, near Orán city (23°08′ S; 64°20′ W) in Salta province.
In this region, there is an annual deforestation rate of 2·3% (Brown & Malizia 2004). In the northern sector of the Argentine Yungas, near Orán, citrus plantations occupy 13·500 ha (Danza 2001). Most of them are grapefruit (57%) and represent more than 30% of the national grapefruit production, which is mainly exported (Danza 2001).
study sites and experimental design
We considered five ‘distance treatments’ to the forest edge at each of four different grapefruit plantations: 0 m or ‘edge’ (the first grapefruit tree row in the plantation from the edge), 10 m (the third row), 100, 500 and 1000 m. Grapefruit in these plantations were of the red and very red seedless varieties (Foster Seedless, at Peña Colorada, Rio Red at Citrusalta and Rouge La Toma at Manero and La Toma). At the beginning of the 2000 flowering season, 20 focal plants were selected at each distance for the visit frequency censuses across the whole study. Flower visitors of each plant were observed over 3 consecutive years (2000–02) and throughout each flowering season (middle August–early October). The plantations were embedded within a citrus agricultural matrix (either grapefruit or orange), and were all adjacent to the forest on one side and adjacent to other citrus plantations on the side furthest from the forest.
None of the study citrus plantations had domestic beekeeping activity during the study period (personal communication of the owners; N. Chacoff, personal observation). Thus, we inferred that all the A. mellifera visiting grapefruit flowers were feral africanized honeybees (Schneider, DeGrandiHoffman & Smith 2004). Over the 3 years of this study, we found no bee tree nests and only one ground nest of Bombus atratus Franklin found in just one plantation. Thus, we inferred that most bees observed in the plantation were nesting in the nearby forest. Heavy use of herbicides in these plantations ensured that citrus flowers were almost the exclusive and dominant source of nectar and pollen present in the plantations. Distances between plantations range from 5 to 50 km, which exceeded the expected flight distances of most foraging bees and other invertebrate flowervisiting taxa (Osborne et al. 1999; WaltherHellwig & Frankl 2000; Gathmann & Tscharntke 2002; SteffanDewenter & Kuhn 2003; Klein, SteffanDewenter & Tscharntke 2004). Thus, the four plantations included in this study were considered true replicates (Hulbert 1984).
sampling of flower visitors
Insect visits to flowers were observed between 0930 and 1700 hours under typical weather conditions, i.e. sunny and slightly cloudy days with low wind velocity. Different sites were sampled on different days. Censuses of visit frequencies were restricted to grapefruit trees in full bloom (i.e. > 50% of their flowers opened). For each sampling day and distance, three censuses were made on different trees at least 30 m apart and observation times were randomized among distances. Each census consisted of a 15min observation period to a flowering branch or group of branches, so a similar number of flowers (mean ± 1 SD = 10 ± 5 flowers observed per census) was observed on each sample tree. After counting the number of open flowers, we carefully recorded the number of visitors and identified them to the lowest taxonomic level possible. However, some native visitors, particularly some congeneric bee species, could not be identified from our observation posts; thus, for diversity estimations they were recorded only as morphospecies. These censuses were carried out during 8–10 days per plantation and per year distributed throughout the flowering season. In total, 941 censuses were performed over 3 consecutive years in the four grapefruit plantations.
During the 2002 flowering season, insects visiting grapefruit flowers were also recorded along transects to assess edge effects on the composition of the flowervisiting assemblage. These transects consisted of 10min walks at a given distance (i.e. 0, 10, 100, 500 or 1000 m) along which all presumed legitimate flower visitors were identified to the lowest taxonomic level possible. On a given sampling day, two transects per distance were made and different distances were surveyed in a random order. Transect surveys were conducted during a total of 20 days and over a total of 33 hours.
After frequency censuses or along transects, flowervisiting insects were collected with sweepnets or with an entomological aspirator for taxonomic determination. Because we collected for vouchering, only new species were captured. Bees and syrphid flies, the dominant grapefruit flower visitors, were identified to the species or at least to the genus level. Specimens of some Halictidae from the genera Augochlora, Dialictus and Augochloropsis and some Andrenidae could not be identified to the species level and were classified into morphospecies.
data analysis
From censuses of visit frequencies, we estimated the following four variables: (1) species richness; (2) total visit frequency; (3) visit frequency by honeybees only; and (4) visit frequency by other visitors. Variable 1 was expressed as no. morphospecies/census and variables 2–4 were expressed as no. visits × 15 min^{−1} × flower^{−1}. Only insect visitors actually or potentially performing pollination (i.e. those contacting anther and/or stigma) have been included in the analysis.
The effect of distance to the forest edge was assessed using general (variables 2 and 3) and generalized (variables 1 and 4) linear mixed models (Littell, Stroup & Freund 2002). The latter models are used when variables cannot achieve normality even after transformation. In all models, distance was considered as a fixed effect and site, year and two and threeway interactions as random effects. As this study focused on the effect of distance to the forest edge, we did not consider significance tests associated with random effects. However, variance and covariance estimates > 0 are provided for comparison purposes (Appendix 1).
To analyse variables 2 and 3 we used the mixed procedure in SAS (SAS 1999) that implements a generalization of the standard linear model allowing the incorporation of random effects (for further details see Littell et al. 1996). To analyse variables 1 and 4, we used the macro program glimmix (available at http://ftpsas.com/techsup/download/stat/glmm800.html) that iteratively calls SAS procedure mixed (SAS 1999). In all cases Satterthwaite's approximation method was used to estimate degrees of freedom of the models because our data sets were not completely balanced (Littell et al. 1996). We assessed further the relationship between variables 1–4 and distance using a linear and quadratic regression approach applying contrasts. A summary of all fitted models to the census data is shown in Table 1.
Table 1. Type of general and generalized linear models applied to the different dependent variables estimated from the census data (see Data analysis) and results of the models for the distance (fixed) effect including linear and quadratic contrasts. For each model, information on the transformation of the dependent variable, distribution of the error term and link function are included Dependent variable  Measurement units  Transformation  Error distribution  Link function  General linear model  Linear contrast  Quadratic contrast 

d.f.§  F  d.f.  F  d.f.  F 


1 Species richness  No. morphospecies/census  None  Poisson  Log  4,9·4  1·60  1,10·6  5·64**  1,10·0  0·04 
2 Total visit frequency  No. visits × 15 min^{−1} × flower^{−1}  Square root  Normal−  4,11·8  3·17*  1,12·1  11·68***  1,11·9  0·60  
3 Visits by Apis mellifera  No. visits × 15 min^{−1} × flower^{−1}  Square root  Normal−  4,11·9  2·18  1,12·2  7·61**  1,12  0·97  
4 Visits by other visitors  No. visits × 15 min^{−1} × flower^{−1}  None  Binomial  Logit  4,7·69  5·71**  1,12·2  17·16***  1,12·2  2·04 
For insect data from transect surveys we conducted a nonmetric multidimensional scaling (NMDS) ordination using PcOrd version 4·0 (McCune & Mefford 1999). The rows of this input matrix represented the morphospecies (50 rows), and the columns the five distances for each of the four sites (20 columns). To standardize observations by sampling effort (sample sizes varied between four and 16, with an average of 9·5 transect surveys per distance and site), for each cell of the matrix the number of times each pollinator species was observed was divided by the total number of transects made at each distance in each site.
We ran two separate analyses: one using the complete matrix (50 × 20) and the other using a reduced matrix (28 × 20) in which rare species (species that appear only once in the census) were deleted. The NMDS ordination was based on a Bray–Curtis distance coefficient matrix (Legendre & Legendre 1998). We used a twodimensional configuration in both because the final stress (19·74 and 24·96, respectively) did not decrease significantly in the threedimensional configuration (McCune & Mefford 1999). Because the results of some ordination methods can be affected spuriously by unequal sample sizes despite standarization (Legendre & Legendre 1998), we also rarefied samples by the minimum number of transects surveyed (four) using ecosim (Gotelli & Entsminger 2001). Here we report results from the analysis of the complete matrix only, because they were very similar to those of either the reduced and rarefied matrices.
To assess the relative importance of the plantation (= site) factor and the distance to the edge factor on species composition we conducted a permutation test following Vazquez & Simberloff (2003). To conduct this test, we constructed two additional matrices with the same dimensions as the Bray–Curtis matrix (i.e. 20 × 20). In the first matrix, pairs of distances surveyed in the same site were represented by zeros whereas pairs of distances surveyed in different sites were represented by the number 1, independently of the distance to the forest edge. In the second matrix, pairs of the same distance to the edge (in the same or different plantations) were represented by zeros whereas pairs of different distances to the edge were represented by 0·25, 0·5, 0·75 and 1, according to how many distance treatments separated a given pair. We then calculated the standardized Mantel correlation statistics (r_{m}) to measure the independence of the entries of the dissimilarity matrix from those of the other two. We randomly permuted the elements of one of the matrices, and recalculated the statistics 1000 times to estimate a Pvalue for each observed statistic.
To investigate further whether, within each plantation and for the whole data set, poorer species assemblages (usually those far from the forest edge; see Results) represent nonrandom, nested subsets of those present near the edge; presence/absence matrices were assembled by listing distances as rows and insect species as columns for each site separately and for all sites together. A nestedness index was computed for each matrix with the Nestedness Calculator software (AICS Research, University Park, MN, USA; Atmar & Patterson 1993). The estimated metric measures the disorder of the matrix using the distribution of unexpected presences and absences of species compared to those in a perfectly nested matrix (Atmar & Patterson 1993). This measure of disorder is called temperature (T) and ranges between 0 and 100°. The level of nestedness is defined as n = (100T)/100, with values ranging from 0 and 1, maximum nestedness being 1 (Bascompte et al. 2003). For each matrix, Nestedness Calculator generates a population of n = 1000 random matrices and estimates P, the probability of a random replicate being more or equally nested than the observed matrix. It has been argued, however, that Nestedness Calculator may overestimate the degree of nestedness and its statistical significance, because the null model assumes that all species are equally common (Fischer & Lindenmayer 2002). In spite of this potential drawback, all significant results presented in this paper (but one) are highly nested at the level of P≤ 0·005; thus, they are unlikely to be the result of a type I error. Nestedness indexes were estimated for the complete insect assemblage and for bee species only (superfamily Apoidea), as bees represented the most diverse and functionally important component of the flowervisiting insect assemblage.