Introduction

Stable isotope ratios of carbon (δ 13C) and nitrogen (δ 15N) are widely used as tracers of the origin and fate of organic matter in food webs (Haines and Montague 1979; Loneragan et al. 1997; Schwamborn et al. 2002; Bouillon et al. 2004; Giarrizzo et al. 2011; Layman et al. 2012). Furthermore, stable isotope ratios of nitrogen (δ 15N) are generally assumed to be good indicators of the trophic level (TL) of consumers in food webs (DeNiro and Epstein 1981; Minagawa and Wada 1984; Vander Zanden and Rasmussen 2001; Post 2002; McCutchan et al. 2003). This notion is mostly based on observations that values of δ 15N increase within food webs (Harrigan et al. 1989; Hobson et al. 1996; Pinnegar and Polunin 1999; Sweeting et al. 2007) or on laboratory feeding experiments (Minagawa and Wada 1984; Post 2002). Several studies have compared TL estimates based on stomach contents and TL estimates based on nitrogen isotopes, assuming a constant discrimination δN = 3.4 ‰ TL−1 (Kline and Pauly 1998; Polunin and Pinnegar 2000; Dame and Christian 2008; Nilsen et al. 2008; Milessi et al. 2010), but none were performed in mangrove food webs or any other tropical ecosystems, and none of these studies have tried to assess the magnitude and standard error of the change in isotope ratio with TL.

Trophic processes are known to cause an enrichment in the heavy isotope, or more precisely, a trophic discrimination (Δtissue-diet), generally calculated as the difference in stable isotope signature between a consumer’s tissue and its diet. Since the discovery of δ 15N as a proxy for TL, an increase of 3.4 ‰ per TL, based on one single study (Minagawa and Wada 1984), has been generally assumed as a universal value of ΔN. Almost two decades later, a comprehensive review of previously published ΔN data (Post 2002) confirmed the value of 3.4 ‰ as the average ΔN, with a standard deviation of 0.98 ‰ (n = 56). Furthermore, this review compared fractionation rates among different types of habitats and feeding modes (Post 2002). This exact value of ΔN = 3.4 ‰ has since then been used to estimate TL of consumers in a variety of ecosystems, based on consumer δ 15N, leading to a plethora of studies based on “precise” TL values (e.g., Kline and Pauly 1998; Vander Zanden et al. 1999; Eggers and Jones 2000; Polunin and Pinnegar 2000; Mantel et al. 2004; Vander Zanden and Rasmussen 2001; Urton and Hobson 2005; Delong and Thorp 2006; Dame and Christian 2008; Nilsen et al. 2008; Milessi et al. 2010). This exact value of ΔN has been widely used in spite of solid evidence that the magnitude of the enrichment in 15N may actually be very variable, ranging from less than −1 to more than 5 ‰ (Peterson and Fry 1987; Post 2002; McCutchan et al. 2003).

The stable isotope signature of carbon (δ 13C) is also known to change between diet and consumer, yet to a much lesser extent than δ 15N. Most food web studies based on isotope mixing models simply ignored possible changes in δ 13C with TL (e.g., Abed-Navandi and Dworschak 2005; Connolly et al. 2005; Benstead et al. 2006). Other recent studies based on food web mixing models simply applied a constant correction of 0.4 ‰ to all consumers (March and Pringle 2003; Delong and Thorp 2006), regardless of TL, or assumed a range for ΔC between 0 and 1.5 ‰ (Schwamborn et al. 2002). Even when a consistent value of ΔC was used in food web mixing models, other fundamental information, such as trophic level, was generally not considered (Phillips et al. 2005; Benstead et al. 2006). More sophisticated Bayesian mixing models are able to account for the uncertainty in trophic discrimination factors, but still do not consider its additive character and increase of uncertainty with TL (Parnell et al. 2010).

Only recently has the additive character of trophic discrimination been considered in the analysis of carbon sources in a mangrove food web (Giarrizzo et al. 2011). Instead of applying a constant correction factor, regardless of trophic level, this mixing model considered the discrimination that happens at each step within the food web. Furthermore, the effect of error propagation in food webs on the uncertainty associated to mixing model estimates was assessed (Giarrizzo et al. 2011).

Trophic discrimination (Δ) has been recognized as the result of several processes, mainly fractionation and routing. Fractionation during metabolic transformations is well known to lead to an enrichment in food webs (Peterson and Fry 1987), e.g., higher respiration and excretion rates for the lighter isotope lead to enrichment in the heavier isotope in consumer tissues. Isotopic routing, on the other hand, is the effect of the unequal distribution of macromolecules (and their isotopes) to different body tissues and metabolic pathways (Martínez del Río and Wolf 2005). One of the most severe and best understood forms of routing is the change in δ 13C due to the differential allocation of lipids, which are isotopically lighter than other macromolecules (DeNiro and Epstein 1977), to storage tissues and respiration, leading to a 13C enrichment of protein-rich tissues, such as muscle (Focken and Becker 1998). Several authors have considered the effects of lipid content on δ 13C, generally assessed as the carbon/nitrogen ratio of the samples (McConnaughey and McRoy 1979; Sweeting et al. 2006; Post et al. 2007; Smyntek et al. 2007). Yet, no previous studies have attempted to verify and compare the relative magnitude of the effects of trophic levels and changes in C/N ratio on stable isotope data in food webs.

Since the beginning of the use of stable isotopes as food web tracers (Fry et al. 1978; Peterson and Fry 1987), ecologists have argued about the importance of various factors in determining the isotope signature of a given consumer in a food web, such as source isotopic signal, trophic fractionation, and the biochemical composition (lipids, proteins, fatty acids, etc.) of the samples (Post et al. 2007; Tieszen et al. 1983; Bearhop et al. 2002; McClelland and Montoya 2002; Vanderklift and Ponsard 2003; McCue 2008). Some have even questioned the usefulness of the stable isotope method, based on the uncertainty of these factors, that may be interacting in complex ways (Gannes et al. 1997; Bond and Diamond 2011; Galván et al. 2012). A common plea in most studies is the urgent need for more knowledge on the processes forming a consumer’s isotope signature under realistic conditions. The objectives of this study were to verify a possible ecosystem-wide enrichment in 13C and 15N with trophic level and to analyze the potential relationships between tissue C/N ratios and stable isotope ratios of consumers in a tropical mangrove food web.

Materials and Methods

Sampling

Sampling was conducted in May 2004 (rainy season) in a macrotidal mangrove creek in the inner part of the Curuçá Estuary, approximately 160 km from Belém, northern Brazil (0° 10′ S, 47° 50′ W). Samples of primary producers included 12 samples of mangrove (Rhizophora mangle and Avicennia germinans) leaves in different stages of senescence, 10 samples of epiphytic macroalgae that grow on mangrove roots (Bostrychia sp., Catenella sp., and Enteromorpha sp.), and 3 samples of epibenthic microalgae that were collected at low tide by gently scraping the visible mats of benthic diatoms on the sediment surface, where they formed a conspicuous greenish-brown layer. Phytoplankton in the sediment-rich macrotidal waters of this estuary cannot be sampled separately (POM samples are mainly resuspended sediments) and is mainly composed of resuspended epibenthic microalgae. Consumers included 52 samples of benthic invertebrates, 6 samples of zooplankton, a bird, a piscivorous bat, and 92 samples taken from 31 fish species. Large animals such as crabs, shrimp, fish, and large vertebrates were sampled as muscle tissue, and microinvertebrates as several whole organisms. Detailed descriptions of the study area, sampling strategy, and sample preparation are provided by Giarrizzo et al. (2011).

Stable Isotope Measurements

To remove all inorganic carbonates and standardize sample treatment (Jacob et al. 2005), subsamples of all samples for δ 13C analysis were treated with 0.2 ml of 0.1 N HCl without rinsing before they were dried at 50 °C for 12 h. Subsamples for δ 15N analysis were not acidified because this would result in a 15N enrichment (Pinnegar and Polunin 1999).

Carbon and nitrogen isotope ratios were determined using a CN analyzer (ThermoFinnigan 1112 Series—Flash EA) interfaced with a mass spectrometer (ThermoFinnigan MAT Deltaplus). Stable isotope ratios were expressed in the standard δ notation (Craig 1953). The analytical precision of the instruments, assessed by the standard deviation of 39 measurements of the working standard, was ±0.11 and ±0.66 ‰ for carbon and nitrogen isotope ratios, respectively. Atomic C/N ratios were measured to assess the potential effect of lipids on δ 13C. C/N ratios are often a good proxy for lipid content in animal tissues (Schmidt et al. 2003; Sweeting et al. 2006; Smyntek et al. 2007), although several other factors (carbohydrates, varying size of amino acids, chitin content, etc.) can also affect C/N ratios. This ratio was measured simultaneously with δ 13C and δ 15N. A total of 177 samples were successfully analyzed for stable isotopes (13C and 15N) and C/N ratio.

Data Analysis

TLs were assigned to each consumer, based on stomach content analysis (for fishes) or published data (for invertebrates). TLs were initially defined as TL = 1 for plants, 2 for herbivores, 3 or higher for carnivores, and intermediate values were calculated for animals with mixed diets. For fish, TL was calculated based on quantitative stomach content data (wet weight) from the present study and on previous studies carried out in the same geographical area and habitat (Brenner and Krumme 2007; Krumme et al. 2005; Giarrizzo and Saint-Paul 2008; Monteiro et al. 2009; Giarrizzo et al. 2010; Giarrizzo et al. 2011; Froese and Pauly 2007). The TL of each of fish species k was calculated using the following equation (Cortés 1999):

$$ {\mathrm{TL}}_k=1+\left({\displaystyle \sum {P}_j \times {\mathrm{TL}}_j}\right) $$

where P j is the weight contribution of prey category j and TL j is the trophic level of each prey category j. Estimation of TL j (TL of prey) was based on stomach content data for each prey species (for fish) or from published data (for invertebrates).

Trophic discrimination ΔX for each isotope ratio δ iX was estimated as the slope of the relationship δ iX vs TL in linear models. To assess the robustness of linear models to potential issues of the data structure (heteroscedasticity, non-normality, and outliers), a comparison of the results of standard and nonparametric fitting methods was performed. Two methods of fit were used for comparison: the commonly used ordinary least squares regression (OLSR) and robust regression (RR), which is less prone to errors due to outlier effects, heteroscedasticity, and non-normality of residuals (Rousseeuw and Leroy 1987; Hubert et al. 2004). Significance of descriptor effects in RR models (assessment of p) was tested by bootstrap (10,000 runs, Faraway 2002). Homoscedasticity of all linear models was tested using the studentized Breusch-Pagan test (Breusch and Pagan 1979; Koenker 1981). Normality of the residuals of all models was checked using the Shapiro-Wilk test (Royston 1982). These linear regression analyses were performed for all data (consumers and primary producers), for consumers only, and for fish muscle samples only. Furthermore, comparisons of isotope signatures and C/N ratios were performed between trophic level groups (herbivores, omnivores, carnivores, etc.) using nonparametric Mann-Whitney U tests (Mann and Whitney 1947; Zar 1999).

Monte Carlo simulations were run to test the robustness of the regression analysis to possible artifacts of the sampling design and outliers. To check to what extent a random removal of samples from the data set affects the significance of the regression slope, a subsample of fixed size was randomly removed from the original data set, without replacement. The “sample” function implemented in “R” (R Foundation for Statistical Computing, 2009, version 2.15.1) was used to perform the random sampling. The remaining data set was then used to perform OLSR analysis of stable isotope and TL data, producing a slope and a p value for each run. This procedure was repeated 5,000 times, thus generating a distribution of 5,000 p values for each subsample size. The size of the remaining data set was then continuously reduced until more than 5 % of the runs yielded nonsignificant p values (p > 0.05).

To test whether variations in C/N ratio have an effect on ΔC, C/N normalization (i.e., lipid normalization) of δ 13C values was performed. Two methods of C/N normalization were used for comparison, based on the C/N ratios of samples: (1) the equation proposed by Post et al. (2007) (δ 13Cnorm1) and (2) the residuals of the simple OSLR linear model of δ 13C vs C/N (all consumers), added to the arithmetic mean of the original δ 13C values (δ 13Cnorm2). C/N normalization through residuals of the δ 15N vs C/N ratio linear models was also applied to consumer δ 15N for comparison with δ 13C.

Multiple linear models were built to explain the variability in stable isotope signatures, considering both continuous variables (TL and C/N ratio) and several covariates (consumer type, sample type, feeding mode, etc.). Variables and covariates were added to the model in a stepwise forward procedure. Inclusion of variables was accepted if their effects in the model were significant (p < 0.05) and if their inclusion increased the total R 2 and the adjusted R 2 of the model. Adjusted R 2 is a modification of R 2 that adjusts for the number of explanatory terms in a model. Unlike R 2, the adjusted R 2 increases only if the new term improves the model more than would be expected by chance (Zar 1999). The relative importance of each variable in the model (% variability explained) was assessed a posteriori through the “lmg” method (Lindeman et al. 1980) in the “R” package “relaimpo” (Groemping 2006). All analyses and tests were performed using the “R” language and environment at α = 0.05 (R Foundation for Statistical Computing, 2009, version 2.15.1).

Results

Trophic Discrimination

The relative amount of the heavy isotopes significantly increased with trophic level, thus indicating the existence of a positive trophic discrimination (i.e., trophic enrichment) for carbon and nitrogen in the Curuçá mangrove ecosystem (Fig. 1). When all samples (primary producers and consumers) were included in the analysis, a significant OLSR model was obtained with δ 13C vs TL, which explained 26 % of the variability in δ 13C (Table S1). Linear models for δ 15N yielded a much better fit (higher R 2) than the models built for δ 13C. Linear models of δ 15N vs TL were highly significant and explained 75 % of the total variability in ecosystem δ 15N. Trophic discrimination (Δ) values of 1.6 ‰ TL−1 (95 % confidence interval 1.2 to 2.0 ‰ TL−1) and 2.3 ‰ (95 % confidence interval 2.1 to 2.5 ‰ TL−1) were obtained for δ 13C and δ 15N, respectively, when including consumers and primary producers (Fig. 1).

Fig. 1
figure 1

Stable carbon and nitrogen isotope ratios (δ 13C and δ 15N) vs trophic level of primary producers and consumers in the Curuçá Estuary, Brazil. Prim. Prod. primary producers, Top. Pred. top predators. Algae: benthic macroalgae (except Bostrychia sp.) and epiphytic algae Enteromorpha sp. Mangrove: green and senescent (yellow) leaves of R. mangle and A. germinans and Bostrychia sp. Mangrove-based and algae-based food webs are indicated for illustration. Linear models were fitted using ordinary least squares regression (dashed lines) and by robust regression (continuous lines). Statistical significance of RR models (p values) was assessed with bootstrap. Linear models are shown for all data and for all consumers

Full size image

Primary producers showed a wide range in δ 13C, from −30.2 to −19.2 ‰ (Fig. 1). Consumers also showed a wide (9.8 ‰) range in δ 13C, from −26.8 to −17.0 ‰, indicating the existence of algae-based (13C-rich) and mangrove-based (13C-depleted) food webs (Fig. 1). Most consumers were feeding on the 13C-rich algae-based food web, especially carnivorous and top predatory fish (Fig. 1). The majority of herbivorous invertebrates, on the other hand, was concentrated in the mangrove food web (Fig. 1). The range of the residuals of the OLSR model for δ 13C (all data) was 11.5 ‰, similar to the ranges of primary producers and consumers. The variability in primary source δ 15N, on the other hand, was much narrower (from 1.6 to 6.4 ‰). Furthermore, the range of the residuals of OLSR models for δ 15N (7.2 ‰) was also much narrower than for δ 13C, which may, together with the strong ΔN, explain the much better fit (higher R 2) of the linear models for δ 15N than for δ 13C.

When considering consumers only, significant relationships with TL were still found for δ 13C and δ 15N. However, for δ 13C, only 3 % of the variability was explained by the linear model, and it was barely significant (p = 0.027). For δ 15N, the linear model for consumers also showed a lower R 2 than the model with primary producers and consumers, but this model still explained 56 % of the variability in δ 15N, and the significance level was still very high (p < 0.0001), indicating an excellent fit.

Significant relationships were detected between several variables (stable isotope ratios, TL, and C/N ratio), with p generally below 0.0001 (Figs. 1 and 2). Yet, for all bivariate models, essential assumptions of OLSR, such as normality of residuals and homoscedasticity, were not given. Thus, RR and bootstrap should be more appropriate to analyze the relationships of most of the pairs of variables than OLSR, although both fitting methods generally produced very similar models (Figs. 1 and 2, Table S1, Supporting information). Slopes estimated with OLSR and RR were 6–18 % apart from each other, estimates of R 2 and p were generally similar within 1 % (Table S1, Supporting information).

Fig. 2
figure 2

Stable isotope ratios (δ 13C and δ 15N) vs carbon/nitrogen ratios of consumers in the Curuçá Estuary, Brazil. Linear models were fitted using ordinary least squares regression (dashed lines) and by robust regression (continuous lines). Statistical significance of RR models (p values) was assessed with bootstrap. Symbols as in Fig. 1

Full size image

The exclusion of primary producers led to estimates of ΔC that were drastically reduced (by more than a half), to 0.6 ± 0.3 ‰ TL−1 (p = 0.027; R 2 = 0.03; N = 145), while consumer ΔN remained virtually unchanged (ΔN = 2.05 ± 0.15 ‰ TL−1; p < 0.0001; R 2 = 0.56; N = 152). RR produced similar estimates of consumer ΔN (1.97 ± 0.14 ‰ TL−1), but slightly (8 %) higher values of consumer ΔC (0.65 ± 0.3 ‰ TL−1). Thus, the estimates of Δ were always approximately 2 ‰ TL−1 for δ 15N, regardless of fitting method (OLSR or RR) and data set (with or without primary producers), while the estimates of ΔC were strongly affected by the inclusion of primary producers and, to a lesser extent, by the fitting method.

When considering fish muscle samples only (Table S1, Supporting information), highly significant relationships (p < 0.0001) were still found between δ 15N and TL, but the slope, and thus the estimate of ΔN, was reduced to 1.42 ± 0.17 (OLSR) or 1.46 ± 0.18 (RR). For δ 13C vs TL, the slopes of the linear models for fish samples only were not significantly different from zero (p > 0.05), i.e., fish muscle δ 13C was not significantly correlated to TL.

Monte Carlo simulations showed that for δ 13C vs TL, even a small reduction in sample size affects the significance of the regression slope. A random reduction of the original data set by only six samples (only 4 % of the original data) proved to have serious consequences for the results of the regression analysis. This slight diminution of the data set caused nonsignificance (p > 0.05) in more than 5 % of the simulations.

Conversely, the regression slope for δ 15N vs TL proved to be extremely robust to random removal of samples. A significant positive slope was still obtained with only 13 remaining samples in more than 95 % of the simulations. Thus, 136 samples (91 % of the data) could be randomly eliminated from the original data set in more than 95 % of the runs, and the slope was still significant (p < 0.05).

C/N Ratios

Carbon/nitrogen ratios of consumers clearly decreased with trophic level (Fig. 3). Herbivores and omnivores (TL = 2 to 2.5) showed higher medians and higher means of C/N ratios than carnivores (Fig. 3). Mann-Whitney tests confirmed that herbivores and omnivores (TL = 2 to 2.5) had significantly higher C/N ratios than carnivores (p = 0.021). Furthermore, stable isotope signatures of consumers were significantly correlated with C/N ratio (Fig. 2).

Fig. 3
figure 3

Carbon/nitrogen ratios vs trophic level of consumer groups in the Curuçá Estuary, Brazil. Group medians (horizontal lines) were compared by means of pairwise Mann-Whitney tests. Groups marked “a” and “b” indicate significantly different groups at α = 0.05. Arithmetic means are indicated as “x.” H herbivores, O omnivores, PC primary carnivores, IC intermediate carnivores, SCT secondary carnivores and top predators

Full size image

C/N ratios of primary producers showed a wide range, from 3.5 to 67, where extremely high C/N ratios (e.g., above 10) are due to the high amount of cellulose and long chain fatty acids in mangrove leaves. C/N ratios of consumers (Fig. 3) displayed a much narrower range, from 2.7 to 6.5, where high C/N ratios (e.g., above 3.5) indicate high lipid contents. Herbivores (TL = 2) were highly variable in their C/N ratio, while the remaining trophic levels spanned a very narrow range (Fig. 3). Most consumer samples, and virtually all fish samples, showed C/N ratios below 3.5. Only 15 % of all consumers and only 4 % of the fish samples showed C/N ratios above 3.5.

Significant relationships were also detected between stable isotopes and C/N ratios, with a depletion of the heavy isotope with increasing C/N ratio (Fig. 2). For δ 13C, the linear model with C/N ratio as single explanatory variable (Table S1, Supporting information) had a much better fit (R 2 = 18 %, p < 0.0001) than the model with TL only (R 2 = 3 %, p = 0.024), showing that C/N ratio (i.e., lipid content) was more relevant than TL. Conversely, for consumer δ 15N, the model with C/N ratio only (R 2 = 16 % and 17 % for RR and OLSR, respectively) was much worse than TL only (R 2 = 56 %), showing that for nitrogen isotopes, TL is much more influential than C/N ratio.

Effects of C/N Normalization

After C/N normalization (i.e., lipid normalization), δ 13C values were not any longer increasing with TL (Fig. 4). Both methods of C/N normalizing (δ 13Cnorm1 and δ 13Cnorm2) produced δ 13C data that were not significantly affected by TL and are thus also not correlated with TL or δ 15N. A significant effect of TL could not be obtained by the exclusion of selected single outliers or even selected pairs of outliers. Thus, the increase in the original raw δ 13C data set with TL could be fully explained by the change in C/N ratio (i.e., lipid content) with TL. Conversely, δ 15N values of consumers still significantly increased after C/N normalization. Yet, similar to the effects of building a model for fish muscle samples only (that have a consistently low C/N ratio), C/N normalization did reduce the estimate of ΔN (Fig. 4). C/N normalization of the δ 15N vs TL models (RR and OLSR) produced a reduction of R 2 (40 %) and slope (1.6 ‰ TL−1).

Fig. 4
figure 4

Lipid-normalized stable isotope ratios (δ 13Cnorm2 and δ 15Nnorm2) vs trophic level, including 31 fish species, in the Curuçá Estuary, Brazil. Linear models were fitted using ordinary least squares regression (OLSR) and by robust regression (RR). Statistical significance of RR models (p values) was assessed with bootstrap. Above: δ 13CNORM2 = δ 13C data corrected for lipid effects using the residuals of the OLSR linear model fitted to the C/N ratio vs δ 13C. Below: δ 15Nnorm2 = δ 15N data corrected for lipid effects using the residuals of the OLSR linear model fitted to the C/N ratio vs δ 15N. Symbols as in Fig. 1

Full size image

Multiple Effects

When considering all possible combinations of explanatory variables and covariates, the linear model that best explained the variability in consumer δ 13C was still the simple model with C/N ratio only. The fit of this simple model was not increased by the inclusion of any covariates (consumer type, sample type, feeding mode, etc.) or combinations of covariates, nor was there any significant difference between covariate levels (e.g., muscle vs whole sample) in the residuals of any of these simple models, indicating that none of these factors are relevant for sample δ 13C in this dataset, even when considering C/N ratio or TL.

Conversely, for δ 15N, the model with the largest amount of the total variability explained (best R 2) was a full model (Table S2, Supporting information) with both continuous variables (C/N ratio and TL) and two covariates (consumer type and sample type), all effects being highly significant (p < 0.0001). This full model explained 70 % of the variability in consumer δ 15N, a considerable increase from the linear model with TL only (56 %). When the effects of TL and C/N ratio were considered, vertebrates had a significantly higher (p = 0.015) δ 15N than invertebrates, and muscle tissue samples had a significantly lower (p = 0.0017) δ 15N than samples of whole animals.

When analyzing the relative importance (% of the variability explained by the model) of the four variables (Table S2, Supporting information), it became clear that TL (53 %) was most important in determining a consumer’s δ 15N, followed by consumer type (vertebrate vs invertebrae, 25 %), C/N ratio (14 %), and sample type (muscle vs whole body, 9 %).

Discussion

This is the first study to analyze and quantify isotopic enrichment in a complex mangrove food web across several trophic levels by comparing the relative magnitude of the effects of TL and C/N ratio on stable isotope data. Furthermore, it was possible to examine the relationships between TL, biochemical, and isotopic composition in this ecosystem. It presents a simple and straightforward method to estimate isotope enrichment within any given food web.

Usefulness of δ 15N as a Proxy for TL

Our data confirm the tight relationship between δ 15N and TL, with 75 % of the variability in ecosystem δ 15N explained by TL alone. Thus, our results confirm the accuracy of δ 15N as a proxy to assess TL in mangrove ecosystems. This isotope ratio has been widely used to assess TL (Vander Zanden et al. 1999; Post 2002; Carscallen et al. 2012), although estimates of ΔN had been mostly based on laboratory feeding experiments (Pinnegar and Polunin 1999; Bosley et al. 2002; Schwamborn et al. 2002; McCutchan et al. 2003; Sweeting et al. 2007). The widespread use of a single exact value of ΔN = 3.4 ‰ (Minagawa and Wada 1984) was not confirmed by the Curuçá data. Rather, a much lower ΔN value of 2.3 ± 0.20 ‰ TL−1 (all data) was observed in the present study, thus challenging the notion of a universal ΔN of 3.4 ‰ TL−1. The concept of a single, universally valid ΔN has been challenged by literature surveys (Post 2002; McCutchan et al. 2003) that revealed a huge variability in ΔN. The survey of McCutchan et al. (2003) listed ΔN values ranging from −2.1 to 5.4 ‰ TL−1, with an average of 2.0 ± 0.2 ‰ TL−1, which is within the ΔN obtained in the present study. Nevertheless, this uncertainty in ΔN (Parnell et al. 2010) has generally been ignored in food web models based on isotopes. Published ΔN values are mostly obtained by single-diet (and thus artificial) laboratory feeding experiments. Clearly, more ecosystem-wide studies under different climatic and trophic conditions are necessary to evaluate the factors that affect Δ in real ecosystems.

Isotope discrimination was recently quantified for selected coral reef fish species, based on the comparison of isotope ratios in tissues and stomach contents (Wyatt et al. 2010). Isotope ratios of stomach contents are prone to a series of modifications, such as contamination with mucus and digestive enzymes, and selective absorption of nutrients, which may limit a wider applicability of their results, although few data are yet available to assess the magnitude of these modifications (Schwamborn and Criales 2000; Guelinckx et al. 2008). In spite of these possible drawbacks that need further investigation, the study of Wyatt et al. (2010) provided first in situ ΔNstomach content-tissue estimates for 24 fish species. As expected for this kind of samples, they showed a huge variability in ΔNstomach content–tissue, which ranged from −1.1 to +5.6 ‰ with an average of +2.4 ‰,which is close to the ΔN obtained in the present study and to the average ΔN value given by McCutchan et al. (2003). Furthermore, they observed higher ΔN in carnivorous species than for herbivores. The present study gives an average ΔN for the food webs sampled in the Curuçá mangrove, but that does not mean that there is a single, constant value across all trophic levels. In the present study, ΔN in carnivores was not higher than in herbivores, which would have been observed as a change in the slope of δ 15N vs TL when including primary producers. Conversely, top predators did not show significantly higher δ 15N than intermediate carnivores. This indicates that in the Curuçá mangrove, ΔN of top predators was smaller than the ΔN of lower trophic levels.

Our findings, together with the high variability in ΔN described by several recent studies (Post 2002; Schmidt et al. 2004; Wyatt et al. 2010; Carscallen et al. 2012), imply that δ 15N cannot simply be transformed to TL for any given consumer in any ecosystem by applying a universal value of 3.4 ‰ TL−1, as done in a plethora of studies. Since ΔN is mostly positive (with few exceptions, see Post 2002), our data confirm that δ 15N could be useful as a relative measure of TL in many cases (Harrigan et al. 1989). Many factors have been shown to influence the magnitude of ΔN, such as the trophic status of the ecosystem (Bucci et al. 2007a, b), and even the elemental composition of baseline organisms (Wilson et al. 2009). Thus, a cautionary approach should be chosen when interpreting δ 15N data for each ecosystem and species, ideally together with other, independent estimates of TL (Carscallen et al. 2012).

Comparison of ΔC and ΔN

Our data confirm that the δ 13C of consumers is mainly determined by their biochemical composition (i.e., lipid content) and the δ 13C of their primary food sources, which showed a wide range (from −30.2 to −19.2 ‰). The variability in primary source δ 15N, on the other hand, was much narrower (about half the range found for δ 13C), which may be one main reason for the narrow range in δ 15N at all TLs, and consequently for the excellent linear fit and high R 2 for δ 15N. The stronger and more consistent enrichment for 15N than for 13C confirms the perceptions of previous studies (DeNiro and Epstein 1977; Peterson and Fry 1987; Michener and Kaufman 2007; Boecklen et al. 2011).

The high variability in consumer δ 13C in this mangrove ecosystem has led Giarrizzo et al. (2011) to the conception that it may be composed of three segregated food webs, a mangrove-based food web, with 13C-depleted consumers (e.g., the crab Ucides cordatus and zooplanktivorous food chains), an algal food web, where benthic algae are eaten directly by consumers, and a mixed food web, where consumers (mainly benthivorous fishes) use carbon from different primary sources (Giarrizzo et al. 2011). In their study, taxon-specific ΔC values of 0.6 to 1.5 ‰ TL−1 were used to estimate source contributions for invertebrates, fish, birds, and predatory mammals, which are close to the slopes of δ 13C vs TL presented in this study (0.6 to 1.6 ‰ TL−1).

The comparison of the traditional OLSR with the less common RR fitting method showed that both methods produce similar results and are apparently perfectly suitable for building useful linear models. Considering the heteroscedasticity and non-normality of residuals in all bivariate and multiple models, fitting through RR and a posteriori testing by bootstrap should be preferred to standard parametric (OLSR) methods. However, this important aspect has not been considered in the context of any previously published stable isotope studies.

There are several important differences between the δ 15N and δ 13C data obtained in the Curuçá mangrove estuary that have to be taken account in any interpretation of the observed regression slopes. The linear regression of δ 15N vs TL is well suited to quantify the discrimination (enrichment) of this isotope, because the primary sources have similar and highly overlapping δ 15N values, the discrimination per trophic level is relatively large (ΔN was about one third of the variability within TL), and the data are well distributed within and between TLs (balanced design). This becomes clear in the low p values and the high R 2: Most of the variability in the δ 15N data could be explained by trophic enrichment alone. The robustness of the significance of the slope to random removal of samples during Monte Carlo simulations further underlines the validity of this approach for the assessment of ΔN in food webs.

The linear regressions of δ 13C vs TL, on the other hand, must be interpreted with caution. The primary sources have very different δ 13C values, which produces a high initial variability in the ecosystem and sets off food webs with distinct carbon isotope signature. Trophic enrichment of 13C per trophic level, as expected from the literature (e.g., Post 2002; Schwamborn et al. 2002; March and Pringle 2003; Delong and Thorp 2006), is very low, with less than one tenth of source variability for δ 13C. This leads to high p values (close to the significance level) and low R 2 (only 3 % of the variability in the data are explained by trophic enrichment, when considering consumers only). The random removal of only six samples, or including a few more mangroves and a few less algae samples, has drastic consequences on the results of the analyses of the δ 13C vs TL slope. The vulnerability of this slope to a random removal of only six samples further underlines the weakness of the relationship between δ 13C and TL and its susceptibility to possible artefacts of the sampling design. Also, a straightforward interpretation of this weak slope as being caused by trophic discrimination does not seem possible for several reasons, mainly since it is fully explained by the change of C/N ratio with TL and the highly significant (p < 0.0001) change of δ 13C with C/N ratio.

An important and generally neglected issue of food web modeling (e.g., Christensen and Pauly 1993; Wolff et al. 2000) is the inherent uncertainty in TL estimates that adds uncertainty to any analyses and models. Clearly, there is an urgent need for the development of standardized methods to assess uncertainty in TL estimates. Furthermore, linear regressions of stable isotope data vs TL may be affected by the unbalanced distribution of the species among trophic levels and food webs, which is an unavoidable by-product of using in situ data from a real ecosystem, in contrast to perfectly balanced feeding experiments under artificial conditions (e.g., Steffan et al. 2013). For instance, predatory fish in this ecosystem were mostly feeding on the 13C-rich algae-based food web, leading to distortions in the distribution of the δ 13C data in the higher TLs. This unequal distribution of top predators could generate an artificially positive slope of δ 13C vs TL that is not necessarily due to trophic enrichment. This is one more reason why the slope of δ 13C vs TL cannot be used to estimate ΔC directly from these data. Yet, for ΔN, this issue does not apply, as demonstrated by the results of the Monte Carlo analysis. This issue might become even more complex when multiple species and interacting ecosystems have to be considered (Harrigan et al. 1989; Benstead et al. 2006). The Curuçá ecosystem, on the other hand, is a relatively simple mangrove estuary without any adjacent nearshore macrophytic habitats (e.g., seagrass beds, coral reefs, salt marshes, or macroalgae beds). Nevertheless, it produces a complex system of interacting food webs that impedes a simplistic interpretation of its δ 13C data.

Thus, multiple effects may hamper a straightforward assessment of ΔC from ecosystem-scale regression analysis of δ 13C vs TL, even when the slope is statistically significant. The method of assessing Δ from in situ data, as used in the present study, is valid only when the abovementioned methodological limitations (e.g., source variability, C/N effects, and unbalanced distribution of the data) are taken into account. Some possible issues, such as the C/N effect described above, may also affect the interpretation of laboratory feeding experiments, but have been generally neglected in such studies.

Lipid Effects Vs Trophic Fractionation—Implications for Carbon Mixing Models

The present study shows the importance of considering C/N ratios in the interpretation of stable isotope data. The already weak and barely significant increase of δ 13C with TL in the Curuçá data can be fully explained by the change in C/N ratio with TL.

In spite of the easiness and straightforwardness of calculation (McConnaughey and McRoy 1979; Schwamborn et al. 1999; Schwamborn et al. 2002; Sweeting et al. 2006; Post et al. 2007; Logan et al. 2008) and the immediate availability of C/N ratios as a regular by-product of mass spectrometry, lipid effects or C/N effects have been widely ignored in isotope analyses, with few, notable exceptions (Kiljunen et al. 2006; Bodin et al. 2007; Hoffman and Sutton 2010). More seriously, none of the published estimates of Δ based on feeding experiments consider any lipid effects, probably leading to an overestimation of fractionation for δ 13C. It has been suggested that these effects could be less relevant when lipid contents of diet and consumer tissue are similar and consistently low (C/N ratio <3.5) for aquatic animals (Post et al. 2007), although this is generally not checked for in feeding experiments. In our data, C/N ratios were mostly below 3.5, which is a characteristic for low lipid contents, especially in fish muscle samples. In spite of being a good proxy in many situations, C/N ratios cannot be used to calculate lipid content using a single, universally valid factor. Rather, species-specific equations have to be used to quantitatively assess lipid content from C/N ratio (Mintenbeck et al. 2008). Nevertheless, significant relationships between C/N ratio and both isotopes were found, showing the huge relevance of this parameter for δ 13C, even when most samples have low C/N ratios.

Are C/N Ratios Relevant for Nitrogen Isotope Discrimination?

C/N normalization (i.e., lipid normalization) did not affect the validity and significance of the ΔN given by the strong and highly significant increase of δ 15N with TL. However, ΔN decreased from 2.0 to only 1.6 ‰ TL−1 after removing the C/N effect (i.e., after the δ 15N values were C/N normalized), indicating that a part of the original ΔN was due to C/N effects (i.e., changes in C/N ratio with TL). One possible explanation for this relatively weak, but significant C/N effect, is that tissues with specific amino acid profiles (and thus, specific δ 15N values) concomitantly also may have different C/N ratios, which would explain a correlation of C/N ratio and of δ 15N in this and other studies (e.g., Schmidt et al. 2004). Vanderklift and Ponsard (2003) found that 15N enrichment increases significantly with the C/N ratio of the diet, suggesting that a nitrogen-poor diet can have an effect similar to that already documented for fasting organisms. Yet, no physiological models are available that would predict an effect of the C/N ratio of a consumer’s tissue on its ΔN, as suggest by our data. The weakness of this effect as shown in this and other studies and the lack of straightforwardly applicable physiological models still prohibit a regular C/N normalization of δ 15N in food web models.

A Plea for Standardized Sampling of Prey and Predators

One relevant aspect related to the effect of C/N ratio on δ 15N may be the type of sample (muscle tissue vs whole samples). Several studies suggest that muscle tissue has a distinct δ 15N signature, as compared to other tissues (Schmidt et al. 2004; MacAvoy et al. 2005). In the present study, invertebrate muscle tissue samples showed a significantly lower δ 15N than invertebrates sampled as whole animals. Conversely to the present results, Schmidt et al. (2004) measured the δ 15N values of specific amino acids of Antarctic krill (Euphausia superba) and found that δ 15N values of abdominal muscle were higher than those of the digestive gland. As in the present study, C/N ratios of krill tissues were also clearly related to δ 15N and δ 13C. In the recent study of Wyatt et al. (2010), ΔN was measured as the difference between whole (partially digested) diet items and fish muscle tissue and, therefore, could also have been influenced by the above mentioned effects.

The use of fish muscle samples only (that have consistently low C/N ratios) and C/N normalization both neutralize these deleterious, and possibly methodological, effects. This also explains why both procedures (building linear models for fish only and C/N normalization of data) had similar effects in reducing Δ for C and N. Conversely, the large meta-analysis of Caut et al. (2009) did not find any significant difference in ΔN between muscle tissue and whole body samples, nor between any taxonomic groups, showing the need for further studies on these controversial subjects. As a precautionary measure, a standardized sampling of prey and predator tissues (muscle tissue or whole animals) should be considered when determining Δ from experimental or in situ data.

Conclusions

Our results demonstrate the importance of considering trophic levels and C/N ratios in stable isotope analyses of food webs. To account for these effects, C/N normalization should be used in mixing models based on δ 13C data. This can be easily accomplished using the simple linear method proposed in this study (using the residuals of the linear regression of 13C vs C/N), or using the methods proposed by McConnaughey and McRoy (1979) and by Post et al. (2007). Another possibility to improve the accuracy of δ 13C mixing models is to account for Δ per TL using taxon-specific Δ values and species-specific TL values, as in Giarrizzo et al. (2011). Since there are no well-established values of C/N-normalized Δ values in the literature, it has not yet been possible to perform a dual (TL and C/N) normalization of δ 13C data for regular use in mixing models. Furthermore, the high variability in Δ, the fact that both effects are relatively weak (as compared to the variability in source signatures), impedes a regular dual normalization. The Curuçá data also suggest that a dual normalization of δ 13C would not make sense, since both (C/N and TL) effects are strongly correlated and thus virtually extinguish each other, at least for the present data set.

A regular C/N normalization of δ 15N in food web models cannot yet be recommended, due to the relative weakness of this effect, the physiological basis of which is not yet well understood, in contrast to the huge and well-documented importance of C/N effects for δ 13C. The indiscriminate use of a global ΔN value of 3.4 ‰ should be replaced by other, more precise estimates of TL, such as stomach contents, ecosystem-specific ΔN values (as provided in this study), or compound-specific ΔN values (Chikaraishi et al. 2007; Steffan et al. 2013).

A better understanding of the interactions and processes leading to trophic fractionation and routing on δ 13C and δ 15N will emerge when more data on stable isotopes in specific biochemical compounds, such as amino acids and fatty acids, will increasingly become available (Hammer et al. 1998; Fantle et al. 1999; Fogel and Tuross 2003; Schmidt et al. 2004; Morrison et al. 2010; Steffan et al. 2013).