We analysed the stomach contents and stable carbon isotopes of muscle tissue of four species of frogs from a savannah formation in south-eastern Brazil locally known as Cerrado (Oliveira & Marquis 2002). There is marked seasonality in the area, with a wet/warm season (henceforth ‘wet season’) from September to March and a dry/mild season (henceforth ‘dry season’) from April to August (Rosa, Lima & Assunção 1991). Specimens of four species [Physalaemus cuvieri Fitzinger, 1826, Eupemphix nattereri Steindachner, 1863, Chiasmocleis albopunctata (Boettger, 1885), and Elachistocleis bicolor (Guérin-Meneville, 1838); n = 60, 65, 51, and 54 individuals respectively] were obtained from the collection of the Museu de Biodiversidade do Cerrado of the Universidade Federal de Uberlândia (MBC-UFU). Specimens were collected in the municipality of Uberlândia (18°55′S, 48°17′W, 850 m), in the state of Minas Gerais, south-eastern Brazil, at the Clube de Caça e Pesca Itororó de Uberlândia (P. cuvieri, E. nattereri, and E. bicolor), and at the Estação Ecológica do Panga (C. albopunctata). Details on the study area are provided in Araújo et al. (2007b). Frogs were collected weekly in the wet season and once every two weeks in the dry season, for a period of 2 years. Specimens of P. cuvieri were collected from October 1999 to November 2000; E. nattereri from October 1999 to October 2001; C. albopunctata from November 2000 to October 2001; and E. bicolor from October 1999 to March 2001. Frogs were immediately killed upon collection, preserved in 5% formalin and later transferred to 70% ethanol.
Five measurements were taken from each specimen with digital calipers (nearest 0·01 mm) always by the same person (M.S.A.): snout-vent length (SVL), mouth width, lower jaw length, head length, and eye–nostril distance. We did not measure the mass of individuals, because preservative absorption was likely to bias our results.
Diets were quantified by the analysis of stomach contents of the preserved specimens, which were dissected to obtain stomach contents. Prey items were counted, and identified to the lowest taxonomic level possible (Order and Family, in most cases). We recognize that by lumping prey into such broad taxonomic groups, we are ignoring substantial variation in resources, which may in turn lead to an underestimation of the degree of individual specialization (Bolnick et al. 2002). Our results therefore may be seen as a conservative estimate of the degree of individual specialization in these species. Upon dissection, individuals were sexed by examination of gonads and classified as juveniles or adults (see Araújo et al. 2007b for details).
We measured carbon stable isotopes from the preserved frogs. Araújo et al. (2007a) measured the carbon isotopes of the prey consumed by four other frog species that inhabit the same areas as the species studied here and were collected at the same time period. The prey taxa found in the present study (see Results) were the same found by Araújo et al. (2007a). We therefore assigned the same δ13C signatures reported by those authors to the prey taxa found here.
Carbon isotopic signatures of animal tissues can be altered by ethanol and formalin preservation (Kaehler & Pakhomov 2001; Sweeting, Polunin & Jennings, 2004). However, since we are interested in estimating the variance among individual isotopic ratios (see below, Data analyses) and all our samples were subject to the same preservation conditions, preservation should not be a problem in our study.
The processing of samples follows Araújo et al. (2007a). The abundances of 13C and 12C were determined at the Centro de Energia Nuclear na Agricultura of the Universidade de São Paulo (CENA/USP) in Piracicaba. Samples were analysed in a Micromass 602E mass spectrometer (Finnegan Mat, Bremen, Germany) fitted with double inlet and collector systems. Organic standards (BBOT) were run every 12 samples and their mean ± standard error was –26·9 ± 0·06. Additionally, five randomly picked samples were duplicated. Experimental precision was measured as the mean ± standard error of the repeatability of duplicates and was 0·1 ± 0·02‰. The 13C/12C compositions are reported using conventional delta notation, showing differences between the observed concentration and that of Pee Dee Belemnite (PDB).
Due to the small sample sizes in the dry season (P. cuvieri, n = 2 individuals; E. nattereri, n = 10; C. albopunctata, n = 1; and E. bicolor, n = 2) we analysed only the individuals collected in the wet season (when diet variation tends to be stronger; Araújo et al. 2007b). In the analyses of diet data, only the individuals having any content in their stomachs were analysed, which explains the differences in sample sizes between diet and isotope analyses.
Individual-level diet variation may be confounded with other forms of intrapopulation variation if individuals of different age classes or different sexes are analysed. Therefore, before investigating individual specialization we tested for ontogenetic and sex-related diet differences. In order to investigate ontogenetic diet shifts, we used Schoener's (1968) proportional similarity index (PS),
in which pik and pjk are the proportions of prey category k in the diets of i and j. PSij is the overlap between the diets of i and j, varying from 0 (no overlap) to 1 (total overlap). We tested for age and sex-related differences in diet using the PS index. In the comparisons between sexes, only adults were analysed. For the sake of statistical power, when no ontogenetic and/or sex-related differences in diets were found, we pooled individuals of different age-classes and/or sexes in the analyses of individual specialization.
In order to measure individual-level diet variation, we used the proposed adaptation of the proportional similarity index, PSi, which measures the overlap between an individual i's diet and the population diet. Details on this index can be found in Bolnick et al. (2002). Briefly, the PSi values of all individuals in the population can be calculated and summarized as a population-wide measure of individual specialization, which is the average of PSi values, IS (Bolnick et al. 2002). IS varies from near 0 (maximum individual specialization) to 1 (no individual specialization). In order to make this measure more intuitive, we use V = 1 – IS, so that higher values now indicate higher individual specialization.
The calculation of all indices was performed in indspec 1·0, a program to calculate indices of individual specialization (Bolnick et al. 2002). In indspec 1·0, the proportion of diet categories in the population diet can be calculated either by adding up the prey counts of all individuals for each resource and dividing it by the total count of prey for the population, or by converting the counts of each individual to proportions and averaging them across all individuals for each resource (Bolnick et al. 2002). We used the latter method, whereas Bolnick et al. (2007) used the former, which explains the slight differences between the measures of V in Bolnick et al. (2007) and in the present study (see Diet variation vs. population niche width below). We also used indspec 1·0 to calculate the significance of the PS measures between age classes, sexes, and the V measures of individual specialization. indspec 1·0 uses a nonparametric Monte Carlo procedure to generate replicate null diet matrices drawn from the population distribution (Bolnick et al. 2002), from which P values can be computed. The null model relies on the assumption that each prey item in the diet corresponds to an independent feeding event, which we acknowledge is probably untrue in the case of termites and ants. We used 10 000 replicates in Monte Carlo bootstrap simulations to obtain P values for these indices.
Many studies focusing on individual specialization have relied on gut contents as a source of diet information (Bryan & Larkin 1972; Roughgarden 1974; Robinson et al. 1993; Schindler 1997; Fermon & Cibert 1998; Warburton, Retif & Hume 1998; Svanbäck & Bolnick 2007). However, gut contents are a ‘snapshot’ of an individual's diet and do not necessarily reflect long-term preferences (Warburton et al. 1998). This sampling problem may make one believe that individuals are more specialized than they really are, leading to an overestimate of the degree of individual specialization in the population (Bolnick et al. 2003). Therefore, in studies using gut-content data, it is desirable to have some measure of temporal consistency in food resource use by individuals (Bolnick et al. 2003). Several studies have measured stable isotopes (Fry, Joern & Parker 1978; Gu, Schelske & Hoyer 1997) to infer temporal consistency in the diets of individuals. Due to their slow turnover (Tieszen et al. 1983), isotopes will not be subject to the same stochastic sampling effects as gut contents and can be a more reliable way to infer individual temporal consistency in resource use. In fact, the among-individual variation in δ13C signatures can be interpreted as a measure of individual-level diet variation (Fry et al. 1978; Angerbjörn et al. 1994; Gu et al. 1997; Sweeting, Jennings & Polunin 2005). If the individuals in a given population all have similar diets, they will also show similar isotopic signatures, so that the population isotopic variance will be low. On the other hand, if individuals vary in their isotopic signatures, this can be taken as evidence of long-term interindividual diet variation.
However, using isotopes to estimate diet variation has some drawbacks. If the number of food resources is higher than one can discriminate with isotopes (Phillips & Gregg 2003), isotope variation can underestimate diet variation (Matthews & Mazumder 2004). On the other hand, if resources vary in their isotopic composition in space or time and consumers are sampled in different places and/or times, there will be isotopic variation that cannot be attributed to diet variation. In the same line, variation in fractionation among individual consumers and isotopic variation within food resources themselves may also increase variation in consumers (Moore & Semmens 2008). Finally, for a given level of diet variation, populations using more isotopically variable prey (e.g. –34, –32, –30, –28, –26‰) will show higher isotopic variances than populations using less variable prey (e.g. –31, –30, –29, –28, –27‰; Matthews & Mazumder 2004). Consequently, measures of population isotopic variance per se can be a misleading guide to diet variation if the prey isotopic variance is not taken into account. These caveats, although not invalidating the use of stable isotopes in studies of individual specialization, underscore the necessity of using other sources of diet information (e.g. gut contents) as complementary approaches.
Bearing those caveats in mind, we used a method that converts a measure of δ13C variance in consumers into an estimate of the V index of individual specialization (Araújo et al. 2007a). This method uses empirical diet data of consumers and isotope data of prey to generate an expected relationship between the δ13C variance and the V measure of diet variation. We then use this relationship to convert the empirical variance in the isotopes of consumers into an estimate of V. Readers are referred to Araújo et al. (2007a) for the details on the method. The parameters used in the model were the population diet (estimated empirically from gut contents), the prey δ13C signatures and dry masses (taken from Araújo et al. 2007a) and the δ13C variance of consumers (also estimated empirically). Simulations were run in the program variso 1·0 (Araújo et al. 2007a).
Morphology vs. diet
We tested for the correlation between morphology and diet. Such correlations would be indicative of the presence of biomechanical trade-offs, which would offer a mechanistic explanation for diet variation. In order to test this correlation, we took an approach that relates body shape to interindividual diet overlap. Within each species, we first did a principal component analysis (PCA) on the five log-transformed morphological measurements. We then calculated a matrix of pairwise Euclidean morphological distances based on the PC2–PC5 scores (interpreted as body shape) among all individuals. Next, we calculated a matrix of pairwise diet overlap among individuals, using the PS index, in which pik and pjk are the proportions of prey category k in individual i's and j's diet, respectively. If there were an effect of functional morphology on diet, we would expect that morphologically similar individuals (small Euclidean distances) also show similar diets (high diet overlap), and vice-versa. If this were true, we would expect a negative correlation between the matrices of morphological distance and diet overlap. We tested the correlation between matrices with a simple Mantel test with 10 000 simulations. The PCAs were performed in systat 11 and the Mantel tests were carried out using the software poptools 2·6·9 (Hood 2005).
Diet variation vs. population niche width
The total niche width of each population (TNW) was quantified using the Shannon–Weaver diversity index, following Roughgarden (1979). This index will yield a value of zero when the entire population uses only a single category of prey, increasing with both the number of prey categories and the evenness with which they are used. We then took the data for the other four species of frogs reported in Bolnick et al. (2007) – Leptodactylus fuscus, Leptodactylus sp., Proceratophrys sp., and Ischnocnema penaxavantinho (Giaretta, Toffoli & Oliveira 2007); note that in Bolnick et al. (2007), there are two samples for each frog species (wet and dry season)-combined with our data and regressed V on TNW. A significant positive slope would confirm a positive relationship between TNW and V. Following Bolnick et al. (2007), we tested whether this positive slope could be accounted for by a null model involving only stochastic sampling. We used a resampling procedure to recreate this artefact as a null expectation for the relationship between population niche width and among-individual diet variation. For each sample, we took the population niche (the proportions of each prey category in the population diet) and let each individual observed to have consumed some number n of prey items to randomly sample n items from the population diet frequencies via multinomial sampling. The null degree of diet variation (V) was calculated once all individuals were assigned random diets. For each sample, we carried out 10 000 such resampling estimates. We then regressed the mean resampled V against the observed TNW to evaluate the null hypothesis that limited individual diet data also generate a positive relationship between these measures. The resampling procedures were carried out in indspec 1·0 and the regression analyses were performed in systat 11. The NVH is only supported if the empirical slope of V as a function of TNW is significantly steeper than the null slope.