823

REVIEW
Best practices for use of stable isotope mixing models in
food-web studies
Donald L. Phillips, Richard Inger, Stuart Bearhop, Andrew L. Jackson, Jonathan W. Moore,
Andrew C. Parnell, Brice X. Semmens, and Eric J. Ward
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

Abstract: Stable isotope mixing models are increasingly used to quantify consumer diets, but may be misused and misinter-
preted. We address major challenges to their effective application. Mixing models have increased rapidly in sophistication.
Current models estimate probability distributions of source contributions, have user-friendly interfaces, and incorporate com-
plexities such as variability in isotope signatures, discrimination factors, hierarchical variance structure, covariates, and con-
centration dependence. For proper implementation of mixing models, we offer the following suggestions. First, mixing models
can only be as good as the study and data. Studies should have clear questions, be informed by knowledge of the system, and have
strong sampling designs to effectively characterize isotope variability of consumers and resources on proper spatio-temporal
scales. Second, studies should use models appropriate for the question and recognize their assumptions and limitations.
Decisions about source grouping or incorporation of concentration dependence can influence results. Third, studies should be
careful about interpretation of model outputs. Mixing models generally estimate proportions of assimilated resources with
substantial uncertainty distributions. Last, common sense, such as graphing data before analyzing, is essential to maximize
usefulness of these tools. We hope these suggestions for effective implementation of stable isotope mixing models will aid
continued development and application of this field.
Key words: consumers, diet, food chain, isotopic ratios, trophic level.
For personal use only.

Résumé : Si les modèles de mélange d’isotopes stables sont de plus en plus utilisés pour quantifier les régimes alimentaires des
consommateurs, ils pourraient être mal utilisés ou mal interprétés. Nous nous penchons sur d’importants défis associés à leur
application efficace. La complexité des modèles de mélange a augmenté rapidement. Les modèles actuels estiment les distribu-
tions de probabilités des contributions de différentes sources, ont des interfaces conviviales et intègrent des éléments complexes
comme la variabilité des signatures isotopiques, des facteurs de discrimination, la structure de variance hiérarchique, les
covariables et la dépendance à la concentration. Nous formulons les suggestions suivantes pour assurer la bonne utilisation de
ces modèles de mélange. Premièrement, la qualité des modèles de mélange est limitée par celle de l’étude et des données
sous-jacentes. Pour permettre la caractérisation efficace de la variabilité isotopique des consommateurs et des ressources à des
échelles spatiotemporelles adéquates, les études doivent poser des questions claires et reposer sur une bonne connaissance du
système et une conception soigneuse de l’échantillonnage. Deuxièmement, les études doivent utiliser des modèles adaptés à la
question posée et tenir compte des hypothèses sous-jacentes et des limites de ces modèles. Les décisions concernant le regroupe-
ment de sources ou l’inclusion de la dépendance à la concentration peuvent influencer les résultats. Troisièmement, la prudence
est de mise dans l’interprétation des sorties des modèles. Les modèles de mélange estiment généralement les proportions de
ressources assimilées avec d’importantes distributions d’incertitude. Finalement, il est nécessaire de faire preuve de bon sens,
par exemple en affichant graphiquement les données préalablement à leur analyse, pour maximiser l’utilité de ces outils. Nous
espérons que ces suggestions pour une application efficace des modèles de mélange d’isotopes stables contribueront à leur
développement et à leur application continus. [Traduit par la Rédaction]
Mots-clés : consommateurs, régime alimentaire, chaîne alimentaire, rapports isotopiques, niveau trophique.

Introduction isotopic composition of consumers and their foods to make infer-

In recent years, stable isotope analysis has become an increas- ences about the composition of the animal’s assimilated diet. This is
ingly important tool in the study of food webs (McKechnie 2004; accomplished by the use of isotopic mixing models to convert the
Crawford et al. 2008). A common application is to use the stable isotopic data into estimates of food source contributions from

Received 8 May 2014. Accepted 12 August 2014.

D.L. Phillips. U.S. Environmental Protection Agency, National Health and Environmental Effects Research Laboratory, Western Ecology Division,
200 SW 35th Street, Corvallis, OR 97330, USA.
R. Inger and S. Bearhop.* Environment and Sustainability Institute, School of Biosciences, University of Exeter, Cornwall Campus, Penryn, Cornwall,
TR10 9EZ, UK.
A.L. Jackson.* Department of Zoology, School of Natural Sciences, Trinity College Dublin, Dublin 2, Ireland.
J.W. Moore.* Department of Biological Sciences, Simon Fraser University, Burnaby, BC V5A 1S6, Canada.
A.C. Parnell.* School of Mathematical Sciences (Statistics), Complex and Adaptive Systems Laboratory, University College Dublin, Dublin 4, Ireland.
B.X. Semmens.* Scripps Institution of Oceanography, University of California – San Diego, San Diego, CA 92093, USA.
E.J. Ward.* Northwest Fisheries Science Center, National Marine Fisheries Service, National Oceanic and Atmospheric Administration, Seattle,
WA 98112, USA.
Corresponding author: Donald L. Phillips (e-mail: [email protected]).
*The last six authors contributed equally and are listed in alphabetical order in the list of authors above.

Can. J. Zool. 92: 823–835 (2014) dx.doi.org/10.1139/cjz-2014-0127 Published at www.nrcresearchpress.com/cjz on 27 August 2014.
 824 Can. J. Zool. Vol. 92, 2014

Fig. 1. Number of citations per year from 1990 to 2012 for papers matching the search terms (a) “stable isotopes” and “mixing model” or
(b) “stable isotopes” and “mixing models” from ISI Web of Knowledge. Search conducted June 2013.
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14
For personal use only.

the various components of an animal’s diet (Phillips 2012). One of there was no unique solution (Phillips and Gregg 2003). The Iso-
the earliest studies employing this method was Evelyn Haines’ Source mixing model (Phillips and Gregg 2003) was developed to
classic work on the relative importance of the C4 marsh grass allow calculation of distributions of possible diets, including a larger
Spartina alterniflora Loisel. (smooth cordgrass) and other C3 plants number of sources, and has been widely used. Several other similar
for Atlantic fiddler crabs (Uca pugnax (S.I. Smith, 1870)) in a Georgia mixing models using different algorithms were subsequently devel-
salt marsh (Haines 1976). Since that time, there has been a steady oped as well (SOURCE/STEP: Lubetkin and Simenstad 2004; Moore–
increase in the number of studies utilizing isotopic mixing mod- Penrose pseudoinverse: Hall-Aspland et al. 2005; LP_Tracer: Bugalho
els to examine animal trophic relations, with a noticeable accel- et al. 2008) but have been less frequently used (Layman et al. 2012).
eration in their use in recent years (Fig. 1). More recently, Bayesian mixing models have been developed
At the same time, mixing model capabilities and sophistication that allow flexible model specification in a rigorous Bayesian sta-
have continued to progress in recent years. Initially, mixing mod- tistical framework to incorporate some or all of these features—
els only provided point estimates for dietary contributions (e.g., uncertainties, concentration dependence, larger numbers of
40% A, 33% B, and 27% C), with no uncertainty levels specified to sources—as well as others. These models include MixSIR (Moore
account for variability in consumer or food isotopic values, mea- and Semmens 2008; Ward et al. 2010), SIAR (Parnell et al. 2010),
surement error, etc. Sources of variation and uncertainty were first MixSIAR (Stock and Semmens 2013), IsotopeR (Hopkins and
explicitly addressed in the IsoError mixing model (Phillips and Ferguson 2012), and FRUITS (Fernandes et al. 2014). Erhardt and
Gregg 2001a, 2001b). In certain cases, large differences in elemental Bedrick (2013) also recently published a description for an un-
concentrations among food sources may need to be considered in named Bayesian mixing model that is an extension of the earlier
estimating dietary composition. This feature was incorporated in SISUS model (http://statacumen.com/sisus/, accessed 23 July 2014),
concentration-dependent equations in the IsoConc model (Phillips and Kadoya et al. (2012) developed IsoWeb, which solves for diet
and Koch 2002). Traditionally, mixing models could only partition proportions for entire food webs rather than just a focal con-
the dietary proportions of two or three food sources based on the sumer. Bayesian approaches use statistical distributions to char-
use of one or two isotopic values (e.g., ␦13C, ␦15N), respectively; acterize the uncertainties in food source and consumer isotopic
otherwise the models were mathematically underdetermined and values and in estimated source contributions. Alternative linear

Published by NRC Research Press

 Phillips et al. 825

programming approaches use ranges of isotopic values to deter- may still be useful for characterizing food-web components and
mine ranges of source contributions even in underdetermined giving a rough idea of trophic levels, for example, but not neces-
situations (Evrard et al. 2010, 2012). Our focus here is on general sarily for using mixing models to quantify diets of specific con-
principles and practices for use of mixing models in food-web sumers with so little background information.
studies; we do not attempt to comprehensively review all the
features and capabilities of various individual mixing models 2. Consider what is known about the animal’s diet
here, but Hopkins and Ferguson (2012) provide a table that ad- A good understanding of your study system is essential before
dresses many of these. Readers are also referred to the papers using stable isotope mixing models to determine the diet of con-
cited above and review papers that address this topic (Boecklen sumers. Isotopic mixing models are not a “magic bullet” and re-
et al. 2011; Layman et al. 2012; Phillips 2012; Parnell et al. 2013). searchers cannot simply sample an animal’s tissue, measure the
With the availability of these tools, the number of isotopic mix- isotopic ratios, and hope to determine its diet. This is because you
ing model applications to food-web problems has proliferated in need to know some aspects of an animal’s diet a priori; specifi-
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

the published literature. Hopefully, there have been gains in our cally, you need to know what different food sources might be in
knowledge about food webs and animal diets as a result. However, the diet of your study species. Identification of food sources may
as with any tools, the best results come when they are the right be achieved by traditional means such as direct observation, stom-
ones for the job, they are used properly, and their limitations are ach content analysis, fecal pellet analysis, or via the literature,
understood. As the developers of a number of the mixing models although we advise caution with the latter approach as recent
cited above, collectively we have had considerable opportunity to stable isotope studies suggest that there is considerable inter- and
see the details of how the models are applied in food-web studies intra-population variability in diet (Votier et al. 2003; Semmens
and we consult on those uses when asked. The purpose of this et al. 2009; Willson et al. 2010a; Layman et al. 2012). Once the
paper is to put forth some “best practices” guidelines for judicious dietary sources have been identified, stable isotope mixing mod-
use of isotopic mixing models in these types of studies in the hope els offer an excellent way to quantify the diet (in proportional
that they will help researchers maximize the models’ usefulness, terms) of both individuals and populations and offer significant
but be cognizant of their limitations and assumptions as well. advantages over other methods. Stable isotope analyses provide a
picture of diet integrated over a period of time, while conven-
Suggested guidelines tional dietary analysis by stomach contents, regurgitates, and fe-
cal pellets give only a short-term picture of diet; in addition, stable
1. Use prior knowledge to identify questions and spatial or isotopes represent assimilated diet rather than what is ingested,
temporal scales which may or may not be digested and contribute to a consumer’s
Clearly defined questions or sets of hypotheses should always
For personal use only.

nutrition (Bearhop et al. 1999; Votier et al. 2003). Conventional

be the starting point for any well-designed study. Often isotopic methods may also be biased toward particular types of prey
mixing models are employed to address general questions such as (Hobson et al. 1994).
“what is the diet of consumer X”? However, as discussed more Microbial food sources can be particularly challenging to discern
thoroughly in later sections, it must be realized that (i) there is an and characterize, especially in aquatic systems. While allochthonous
inherent temporal scale to this question that depends on the organic matter may be consumed directly by consumers, it may also
choice of samples and (ii) the diet estimates are not precise due to be consumed indirectly through microbes that decompose this ma-
a number of uncertainties. Beyond static estimates of diet for a terial. Isotopic characterization of microbial food sources is difficult
given time window, interesting questions may be posed about and is often done indirectly using primary consumers or other prox-
quantifying dietary variation owing to factors such as temporal, ies (e.g., bulk or size classes of organic matter, respired CO2), or
spatial, and demographic segregation within or between popula- compound-specific isotope analysis of biomarkers, although all of
tions. There may be population hierarchical structures that are these approaches have their own drawbacks; for further discussion
either categorical (e.g., different regional populations, different of these issues see Middelburg (2014).
social groups in the same region) or continuous (e.g., differences
in zooplankton diet across gradients of light penetration in lakes), 3. Sample collection
and which could either be eliminated from the sampling design or A considered and well-planned sampling design is critical to any
incorporated into the model framework (Semmens et al. 2009). stable isotope mixing model study. A critical consideration is the
The mixing model technique is particularly powerful when the time period over which the diet is to be estimated and what tis-
results of dietary studies are used in subsequent analysis to ad- sues best reflect this period. Bone collagen or fish otoliths may
dress questions about causes and consequences of the observed provide dietary information that spans the entire life of an indi-
patterns, and determine how diet links to life-history parameters vidual (Radtke et al. 1996; Dalerum and Angerbjörn 2005), while
of individuals or dynamics of populations. other more metabolically active tissues may turn over in days
Researchers should also consider whether isotopic mixing mod- (e.g., liver) or months (e.g., muscle) (Tieszen et al. 1983). Sampling
els are the right tool for the job. While in some systems isotopic of some tissues (e.g., bone) requires that the consumer is sacri-
analysis can provide unique opportunities and insights, it may not ficed. Destructive sampling also has the advantage that any tis-
always be the appropriate choice. Examples of this might include sue(s) of interest can be sampled. However, while it may be an
the following: (i) where there is little isotopic variation among option for harvested populations (e.g., fisheries or game birds), it
various food sources; (ii) where the isotopic composition of spe- is often not feasible due to the nature of the study question, where
cific diet items varies substantially on spatial scales over which the aim of the research is to identify how diet affects other param-
the consumers move; or (iii) where there is large temporal varia- eters in the future, or on moral, ethical, or conservation grounds.
tion either in diet composition or in isotopic values of diet items Where destructive sampling is not an option (or undesirable),
relative to the integration time of the consumer tissues sampled. there are a number of nondestructive options available. Blood is a
A sound appreciation of the assumptions and limitations of stable particularly useful metabolically active tissue because it can be
isotope analysis and the application of mixing models is key here. separated into two components: plasma, which rapidly turns over
Novel study systems where very little is known about dietary pref- and so reflects diet over the previous few days (depending on the
erences are unlikely to be suitable for stable isotopic mixing mod- study species metabolic rate), and red blood cells, which turn over
els, unless there are well-described isotopic gradients across more slowly and represent diet over the previous few weeks.
which the consumer is thought to forage such as the gradient Hence, a single sample can yield details on diet over two temporal
between marine and terrestrial habitats. Stable isotope analyses time periods, which may be particularly useful in determining

Published by NRC Research Press

 826 Can. J. Zool. Vol. 92, 2014

diet switches (Hobson and Clark 1993; Votier et al. 2003; Phillips When using Bayesian mixing models that fully incorporate
and Eldridge 2006; Klaassen et al. 2010; Heady and Moore 2013). sources of uncertainty, the amount of data collected for both
Tissues that are metabolically inert after formation such as hair, consumers and sources can have a substantial impact on the pre-
claw, or feather also offer scope for constructing time series. cision of model-based parameter estimates. Bayesian models use
In addition to the isotopic values of consumers, mixing models means and variances (and, for more advanced models, covari-
require data on the food sources consumed. These data are of ances) supplied by the user to characterize uncertainty for each
equal importance to the consumer data and equal consideration source. However, traditional model formulations assume that
and planning should be attached to collection of food sources as these means and variances are known without error, while in
well as to the sampling of consumers. There are two key points to reality they are estimated from samples. When the number of
consider here: firstly the number of sources to include in the samples per source is small (<20), there may be considerable un-
model, and if necessary how to aggregate sources, and secondly certainty in the mean and variance values, and Ward et al. (2010)
temporal and geographic variation in sources. recommend a “fully Bayesian approach” that models them as ran-
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

The number of sources to be included in the model is an initial dom variables conditioned on the observed sample data. When
and critical consideration and will of course be shaped by the using hierarchical models to incorporate structure in a modelling
study system. Mixing models are sensitive to missing sources and framework (e.g., subpopulations, social groups, individual differ-
may produce erroneous results if all sources are not included; ences in consumers; Semmens et al. 2009), the number of samples
hence, it is critical that all food sources are sampled (see section 6 within each level of the model can have substantial influence on
below). However, the discriminatory power of mixing models gen- the precision of the level-specific variance terms. Limitations of
erally decreases with the number of sources, although this is also time and expense often limit the number of samples that are
strongly influenced by the isotopic separation of sources. The possible to collect and analyze, so there are trade-offs among sam-
number of sources therefore should be kept as low as possible ple size, model complexity (number of parameters to estimate),
without excluding sources. In our experience, the discriminatory and the precision of estimation.
power of mixing models starts to decline markedly above six or
seven sources. But how exactly do we define a food source? We 4. Use appropriate diet–tissue discrimination factors
will consider a study examining the diet of Brent Geese (Branta The saying “you are what you eat plus a few per mil” (DeNiro
bernicla (L., 1758)) (Inger et al. 2006), which alter their diet during and Epstein 1976) refers to the fact that even at isotopic equilib-
the nonbreeding period from feeding almost exclusively on inter- rium with its diet, an animal’s tissues generally have somewhat
tidal marine resources during the early season before increasing higher (usually) isotopic values for C and N than its diet because of
feeding on terrestrial grasses as the season progresses. During the discrimination during assimilation and excretion processes (Olive
For personal use only.

intertidal foraging period, the birds feed on two species of the et al. 2003). Before applying a mixing model, these systematic
seagrass genus Zostera L. (eelgrass (Zostera marina L.) and dwarf differences must be corrected. This is generally done either by
eelgrass (Zostera noltii Hornem.)), and as this becomes depleted, subtracting a diet–tissue discrimination factor (DTDF) from the
they eat increasing amounts of the green algae Ulva lactuca (L., consumer tissue isotope values or adding it to the food source
1753) and various species of the genus Enteromorpha Link, 1820 (= isotope values. The latter approach is more flexible, allowing dif-
Ulva L., 1753) before moving inland to feed on terrestrial grasses, ferent factors for different food sources (Ben-David et al. 1997b).
which are dominated by perennial rye grass (Lolium perenne L.) and Whereas most studies focus on the dietary contributions of prey
common timothy (Phleum pratense L.). So we can identify at least six items directly consumed, some studies aim instead to measure
sources, although this number would increase if we tried to iden- the importance to the consumer of food webs based on different
tify Enteromorpha to species level and include all the different basal resources (generally producers or detritus). One example of
species of terrestrial plants on which the geese feed. Aggregation this is the study of Reid et al. (2008) on the importance of biofilm,
of isotopically similar sources is one way to reconcile the compet- allochthonous detritus, macrophytes, and algae basal resources
ing needs for including all utilized sources but retain discrimina- for stream consumers. This requires additional information about
tory power. For the Brent Goose system, the isotopic data for each trophic structure, as the DTDFs must be multiplied by the number
species of food revealed clear patterns leading to the aggregation of trophic levels between the consumer and the basal resources to
of the samples into four sources by combining the different species appropriately correct the isotope values used in the mixing model
of Zostera, Enteromorpha, U. lactuca, and terrestrial grasses because analyses.
these groups were clearly clustered and separated in isotopic Early studies indicated that DTDFs were around +1‰ for C
space (Fig. 1 in Inger et al. 2006). In addition, these source group- (DeNiro and Epstein 1978), +3‰ for N (DeNiro and Epstein 1981), and
ings make intuitive sense biologically and were pertinent to the negligible for S (Peterson and Fry 1987). More recently, researchers
study question. (For further discussion of source aggregation see have documented considerable variation in DTDFs. Several review
section 7 below.) and meta-analysis papers (McCutchan et al. 2003; Vanderklift and
If mixing models are used to analyze diet variation in a hierar- Ponsard 2003; Dalerum and Angerbjörn 2005) summarized this
chical population structure, data collection must assess the isoto- variation as a function of a number of environmental and physi-
pic composition of sources across these strata. The isotopic values ological factors that include environment (terrestrial, freshwater,
of sources may vary widely both temporally and spatially, so it is marine), trophic level, taxon, tissue, metabolic rate (poikilotherm,
also important to ensure that the samples taken for analysis geo- homeotherm), nitrogenous excretion (ammonia, urea, uric acid),
graphically match the food the consumers are eating and reflect and sample treatment procedures. Other authors have examined
the time periods over which the tissues of interest were synthe- the effects of diet quality on trophic enrichment (Webb et al. 1998;
sised, which is likely to mean that source samples are collected Adams and Sterner 2000; Robbins et al. 2005; Miron et al. 2006;
before those of the consumer. Temporal matching of sources and Florin et al. 2011). Caut et al. (2008b, 2009) proposed that DTDFs
consumers can be particularly problematic for tissues that inte- vary systematically based on the isotopic values of the diet, but
grate over longer terms, such as bone collagen, as the isotopic this conclusion has been criticized for the lack of a theoretical and
ratios of these tissues represent the average of all food sources mechanistic basis and also because of mathematical artefacts and
consumed and any diet switches that may have occurred during experimental biases (Auerswald et al. 2010; Perga and Grey 2010;
the period of synthesis. There may also be considerable isotopic Codron et al. 2012). Because of the multiplicity of factors that can
variation within sources even from the same geographic and tem- affect DTDFs, this is probably one of the biggest sources of uncer-
porally sampled population and it is important that the sampling tainty in using mixing models to assess diet, and DTDF estimates
regime adequately captures this variation. obtained for the particular situation (e.g., taxon, tissue, diet)

Published by NRC Research Press

 Phillips et al. 827

Fig. 2. Sample ␦15N vs. ␦13C biplot of a consumer (MINK) and its food sources corrected for trophic enrichment. For simplicity, only mean
values are shown here. The dotted lines show the minimum bounding rectangle that corresponds to the lowest and highest food source
values for ␦15N and ␦13C. The solid lines show the convex hull defined by the sources. Note that shrimp is not a vertex of the convex polygon
created because its inclusion as a vertex would create a concave side. The consumer must fall inside this mixing polygon for its isotopic
composition to be explained as a mixture of just these seven food sources. Data are from Ben-David et al. (1997b).
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14
For personal use only.

are likely to produce the most accurate results (Spence and within the range of the food source isotopic values. For example,
Rosenheim 2005; Caut et al. 2008a; Bond and Diamond 2011). Par- if the consumer had a ␦15N value of 16‰ but all its food sources
ticularly useful in this regard are captive feeding studies (DeNiro ranged from 10‰ to 15‰ (after correction for DTDFs), then there
and Epstein 1978, 1981; Tieszen et al. 1983; Hobson and Clark 1992; is no combination of those food isotopic values that could result
Hilderbrand et al. 1996) and studies of wild populations where in the observed consumer isotopic value. While there technically
their diets are well-known (Fox-Dobbs et al. 2007; Harper 2007; could be mathematical solutions that involve negative diet pro-
Newsome et al. 2010). It should be clear, however, that in the vast portions and possibly proportions >100%, these make no physical
majority of cases there will always be some uncertainty and vari- or biological sense and should not be considered, as assumed by
ability associated with the DTDFs. This was one of the rationales current mixing model applications.
behind the production of Bayesian mixing models such as MixSIR Thus, consumer isotopic values within the range of corrected
(Moore and Semmens 2008) and SIAR (Parnell et al. 2010). These food source isotopic values for each element (e.g., C and N) is a
models have the advantage of allowing one to specify a priori a necessary condition for mixing models to work. This is illustrated
standard deviation around the DTDF estimate to account for this by a bounding rectangle in the ␦15N vs. ␦13C biplot (Fig. 2). How-
uncertainty in the mixing analysis. Despite incorporating this ever, this condition is not sufficient because there may be corners
uncertainty because DTDFs are one of the key parameters in mix- of this space that are outside the range of any “combination” of
ing models, it is not surprising that results are still sensitive to the
␦13C and ␦15N values. For example, in Fig. 2 if the mink had a ␦13C
values used for DTDFs. We strongly recommend the use of sensi-
value equivalent to that of the rodent and a ␦15N value equivalent
tivity analysis to test the robustness of model outputs to variation
to that of the duck, it would fall within the bounding rectangle
in DTDFs.
but there still would be no set of dietary proportions for these
5. Plot your data seven food sources that could result in the observed isotopic val-
It is always a good idea to do some exploratory data analysis to ues for the mink. The only way to account for the mink’s ␦13C is for
make sure that you understand the data that you will be process- the diet to consist entirely of rodents, but this would not account
ing with a mixing model. After applying appropriate DTDF correc- for the high ␦15N value. If two isotopic values are used (e.g., ␦13C
tions (see section 4), the first thing that should be done is to plot and ␦15N), a useful graphical two-dimensional geometric proce-
the data. If, for example, one has ␦13C and ␦15N values for con- dure is to use the food source data points to define a “convex hull”.
sumer tissues and for various food sources, then plot those on a In essence, one connects the food source data points with lines to
graph of ␦15N vs. ␦13C or vice versa (Fig. 2). Food source isotopic create a convex polygon with these data points as vertices. If in-
values can be plotted as mean values, along with confidence in- cluding a source as a vertex would lead to a concave side (e.g.,
tervals around those values if desired. Consumer values can be shrimp), then it would not be connected but allowed to just lie
plotted similarly, or isotopic values for individual consumers may inside the convex polygon defined by the other sources (Fig. 2). If
be shown. In order for there to be solutions for diet composition the consumer lies within the convex hull in isotope space defined
from these food sources, the consumer isotopic values must fall by the food sources, then one or more solutions exist that can

Published by NRC Research Press

 828 Can. J. Zool. Vol. 92, 2014

explain the consumer’s isotopic values as a combination of those the extent possible, researchers should use information from
dietary components. If it does not, then the reason for this must other sources (e.g., fecal analyses, behavioral observations, previ-
be sought. Are there data transcription errors? Are there other ous studies) in selecting the food sources for an animal dietary
food sources that should have been included? Have DTDF correc- study. Excluding sources should be based on knowledge that they
tions been applied or are their values inappropriate? If mean iso- are not important in an animal’s diet, not just that they are isoto-
topic values of food sources are used for the vertices but source pically dissimilar; this example demonstrates that they may still
sample variability is being considered as in Bayesian models, then contribute and exclusion from the model still biases the results.
the question of whether the consumer lies within the convex hull Residual error terms provide one approach to account for this
or not becomes fuzzy. Smith et al. (2013) propose a Monte Carlo (Parnell et al. 2010, 2013). An alternative to excluding sources to
procedure to construct many possible convex hulls, taking into simplify the mixing model is to group sources together, which is
account the variability in isotopic values and DTDF values. They the next topic.
then calculate the proportion of these hulls that contain the con-
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

sumer data points and recommend using a mixing model with a 7. Consider grouping sources
particular set of sources only if this proportion is greater than If sources of uncertainty (e.g., sample variability, measurement
some specified level (e.g., 95%). error, diet–tissue discrimination factor uncertainty) are not con-
Fry (2013) proposed some graphical procedures for analyzing sidered, then in theory mixing models using n isotope values (e.g.,
mixing problems as an alternative to mixing models. These in- ␦13C, ␦15N) can be used to find a unique set of assimilated diet
volve subdividing mixing diagrams into zones and using several proportions for n + 1 food sources. Models with greater than n + 1
semiquantitative indices to indicate the relative degree of uncer- sources are underdetermined and multiple solutions exist, often
tainty in the source (e.g., dietary component) proportions for mix- leading to diffuse ranges of possible diet proportions for each source
tures (e.g., consumers) that fall in these areas. It has long been (Phillips and Gregg 2003; Moore and Semmens 2008; Parnell et al.
recognized that the geometry of sources and mixtures in a mixing 2010). While valid food sources should not be discarded from the
diagram determines how constrained or diffuse the possible model as discussed above, combining some sources to reduce the
source contributions may be (e.g., Fig. 6 in Phillips and Gregg total number may sometimes lead to more constrained, less dif-
2003). Fry’s (2013) procedures may be useful for getting a qualita- fuse solutions (Phillips et al. 2005) and should be considered to be
tive feel for mixture data, but Semmens et al. (2013) were critical an option. Phillips et al. (2005) proposed alternative a priori and a
of their subjectivity and Fry’s dismissal of the statistical probabil- posteriori approaches for combining sources, described below.
ity foundation of Bayesian mixing models. Gannes et al. (1998) stated that when using stable isotope data to
reconstruct animal diets, the sources examined should have iso-
For personal use only.

6. Include all sources in an informed way topically distinct signatures. The a priori approach involves per-
One of the implicit assumptions of any mixing model is that all forming statistical tests for equality of means and combining
the sources are included and researchers should strive to ensure sources before running the mixing model if they are not signifi-
that this is the case to get the most reliable results. If one food cantly different (Ben-David et al. 1997a, 1997b). Rosing et al. (1998)
source that contributes to the diet is excluded from the analysis,
provided a K-nearest-neighbor randomization test for this specific
then that necessarily will lead to a bias in the estimates of the
purpose, but other statistical tests might be used as well. Interpre-
dietary contributions for the other sources since they must still
tation of the results will be more sensible if the sources that are
sum to 100%. The sum of these biases over the remaining food
combined are logically related in some way (e.g., same taxon or
sources will equal the actual contribution of the excluded source.
trophic guild) so that the combined source has some biological
It may not seem too important if a food source that contributes 5%
meaning (Phillips et al. 2005). If several sources are not signifi-
to the diet of the consumer is excluded because the remaining
cantly different in their isotopic composition, then they may be
food sources will simply be attributed to 100% of the diet rather
than the actual 95%. However, the magnitude and direction of the represented in the mixing model by a single set of isotopic values
biases for the individual food sources may vary greatly depending that is computed as the means of the individual source isotopic
on the geometry of the mixing space. Figure 3 shows a hypothet- values or as weighted means if weighting by sample size is de-
ical example of four food sources in the diet of a consumer. sired. What if several sources are similar in their isotopic values,
Source D is the most different from the consumer in its isotopic but are still statistically different? The threshold for significance
values and one might be tempted to consider excluding it from of differences varies with sample size in statistical tests, and if
the analysis on that basis. Table 1 shows the effect of this exclusion sample size is large, then a small difference (e.g., a small fraction
on the results of mixing model analyses for the remaining food of 1‰) may test as significant even though it is of no practical
sources under different hypothetical diets that are all consistent biological significance. Alternatively, the objective may be to con-
with the data. Even when the deleted source actually only ac- trast general groups of food sources whose isotopic differences
counts for 5% of the diet (diet No. 5), this causes biases of +10%, are large compared with the variation within the groups (e.g., C3
−15%, and +10% (which sum to 5%) on the dietary proportion esti- vs. C4 plants, marine vs. terrestrial foods). So some judgement
mates for the other three sources. However, in this particular may be applied, but sources should only be combined a priori if
example, it is also possible that source D, which isotopically is the they are not significantly different in their isotopic composition
most dissimilar from the consumer, could represent up to 25% of using either strict statistical criteria or modifying those by judg-
the diet (diet No. 1); in this case, its deletion would lead to extreme ments of biological significance if sample sizes are large, or if the
biases of +50%, −75%, and +50% (which sum to 25%) for the other within-group isotopic variation is small compared with the among-
three sources. group variation. But combining sources with somewhat larger
This example was designed as a worst case scenario for demon- isotopic differences (e.g., several ‰) implies that the individual
strating the potential impact of excluding a source. It may be sources each contribute equally if simple means are used, or con-
tempting to drop a source because of its dissimilarity in isotopic tribute relative to their respective sample sizes if weighted means
composition from the consumer, but that dissimilarity also mag- are used. These implied contributions may not be the case and
nifies the effect on estimates for the other sources. There are may lead to biases in the mixing model estimates for some or all of
obvious limits to including everything as remotely possible food the sources. This concern is not an issue if there is no significant
sources from the standpoint of sample collection logistics, expense difference among sources because the mean (or weighted mean)
of stable isotope analyses, and the general broadening of source isotopic values are common to the sources being combined. Fortin
contribution distributions as the number of sources increases. To et al. (2007) and Edwards et al. (2011) provide recent examples of

Published by NRC Research Press

 Phillips et al. 829

Fig. 3. Hypothetical mixing diagram for four food sources where one (source D) is deleted from the mixing analysis. The solid lines indicate
the mixing polygon without D and the broken lines indicate how it would be extended by including D. For the effects of deleting source D on
the proportion estimates for the other sources see Table 1.
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14
For personal use only.

Table 1. Effect of deleting a utilized food source (source D) on the dietary proportion estimates for
the remaining food sources (see Fig. 3).
Hypothetical actual diets (%)
Food source Diet 1 Diet 2 Diet 3 Diet 4 Diet 5 Mixing model estimate
A 0 10 20 30 40 50
B 75 60 45 30 15 0
C 0 10 20 30 40 50
D 25 20 15 10 5 —
Note: Five different hypothetical diets are shown, all of which are consistent with the data. If any one of these
represents the actual diet, then the difference between the mixing model estimates and the actual source propor-
tions represent the bias induced by dropping source D out of the model.

this a priori approach to combining sources in their studies of sible solutions for each source can be fairly wide and diffuse,
bear diets. while those for combined sources may be much narrower and
Alternatively, in the a posteriori approach, all the candidate lead to easier interpretation. Figure 4 shows an example of this
sources may be included in the mixing model, providing the usual from a study of paleohuman diets in northern California from
distribution of proportions for each source. Then for each individ- Newsome et al. (2004) and discussed further in Phillips et al.
ual solution, the proportions can be summed for sources that are (2005). Three marine food sources with a broad range of ␦13C and
to be grouped together (Phillips et al. 2005). Since this approach ␦15N values were included in the mixing model for Early Holocene
utilizes each individual solution, it preserves the covariation struc- humans and had broad distributions of possible dietary contribu-
ture among source proportions. In some cases, especially where tions: pinnipeds (4%–48%), shellfish (0%–36%), and marine fish
there are more than four or five sources, the distribution of pos- (0%–68%). These ranges were so broad that they did not provide

Published by NRC Research Press

 830 Can. J. Zool. Vol. 92, 2014

Fig. 4. Ranges of feasible dietary biomass contributions of individual marine food sources for an early Holocene (about 7000 years ago) group
of humans on the central California coast (Newsome et al. 2004). Distributions shown reflect 11 345 dietary solutions found by the mixing
model. While broad ranges of use are possible for each individual marine food source, the combined marine food group is tightly constrained
and represents 70%–84% of the diet. (From Phillips et al. 2005, reproduced with permission of Oecologia, vol. 144, p. 524, © 2005 Springer
Science + Business Media.)
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14
For personal use only.

much information at all about diet. However, when these sources concentrations are incorporated into mixing models. The original
were combined a posteriori into a general “marine foods” source, IsoConc model (Phillips and Koch 2002) performs these calcula-
the combined marine source had a dietary contribution range of tions for a two-element three-source system, but concentration
70%–84%, which clearly indicates the primacy of marine foods dependence has also been incorporated into SIAR (Parnell et al.
versus terrestrial foods in the diet of these humans. Recent exam- 2010), IsotopeR (Hopkins and Ferguson 2012), MixSIAR (Stock and
ples of this a posteriori approach for combining sources in animal Semmens 2013), and FRUITS (Fernandes et al. 2014), which can
studies include Painter et al. (2009) and MacArthur et al. (2011) deal with any numbers of isotopic values and sources. When sta-
using IsoSource, Cherry et al. (2011) using MixSIR, and Vaslet et al. ble isotope analyses are performed, elemental concentrations are
(2011) using SIAR. Ward et al. (2011) proposed a procedure using also determined along the way, so researchers should be sure to
the MixSIR Bayesian mixing model to evaluate the posterior prob- ask the laboratory for these data to consider concentration depen-
abilities of various combinations of sources for the a posteriori dence.
approach. Researchers may choose group membership strictly Since isotopic mixing models assess the proportions of food
based on maximum posterior probability, or they may weigh sources assimilated by the consumer, not just ingested by the
the trade-offs between slightly lower posterior probability with consumer, elemental concentrations in the ingested foods may
greater biological relevance and interpretability (e.g., combining not be the most appropriate measure. Rather, the macromolecu-
sources with similarities in functional guild, taxon, habitat).
lar composition (e.g., % protein, lipid, carbohydrate) and the di-
8. Consider concentration dependence and isotopic routing gestibility and C and N concentrations in these macromolecular
If there are large disparities among food sources in the concen- components will determine the amounts of assimilated C and N
trations of elements associated with the isotopic values being from each food source; Koch and Phillips (2002) show how these
used (e.g., C for ␦13C, N for ␦15N), then the possibility of using a calculations can be made. Thus, the data needs for using a
concentration-dependent mixing model might be considered concentration-dependent model in this way are considerably
(Phillips and Koch 2002). In a typical mixing model application higher and the requisite information may not always be available.
employing ␦13C and ␦15N values for food sources and consumer On the other hand, after going through all these calculations, it
tissues, there is an implicit assumption that the proportion of C may turn out that the assimilated C and N concentrations are
that each source contributes to the consumer is the same as the much more similar among sources than their raw C and N con-
proportion of N that it contributes. But if C and N concentrations centrations indicate. A plant food source may have a low C:N ratio
vary widely so that the sources have very different C:N ratios, then compared with an animal food source, but when the C in indigest-
this may be an untenable assumption. This may be particularly ible materials like lignin, cellulose, and fiber are subtracted out,
applicable to omnivores that consume both low N (and high C:N) the result may be more similar to the animal food. In this case, the
plant foods and high N (and low C:N) animal foods. A concentration- results of a concentration-dependent mixing model would be sim-
dependent model allows for different C contributions, N contri- ilar to those of a standard model, although this would not be
butions, and overall biomass contributions from any food source, known a priori without going through these calculations (Koch
depending on elemental concentrations and isotopic values. and Phillips 2002). In other cases, the results for contributions of
Readers are referred to Phillips and Koch (2002) for details on how C, N, and overall biomass from various food sources to the con-

Published by NRC Research Press

 Phillips et al. 831

sumer’s tissues may be quite distinct when sizable source differ- these factors and provides statistical confidence limits around the
ences in assimilated C and N concentrations exist (Phillips and proportion estimates. However, this model does not cover under-
Koch 2002). Gauthier et al. (2003) first applied the concentration- determined systems where there were greater than n + 1 sources
dependent model to examine the use of endogenous vs. exogenous when using n isotopic values. IsoSource (Phillips and Gregg 2003)
food resources in birds. More recent examples of concentration- addresses underdetermined systems and provides distributions of
dependent mixing model use include applications to mice possible sets of source proportions, but IsoSource does not in-
(Ramírez-Hernández and Herrera M 2010), birds (Rutz et al. 2010), clude sampling and measurement sources of uncertainty in a
fish (Wilson et al. 2010b; Vaslet et al. 2011), and shellfish (Wilson straightforward way. Recent Bayesian mixing models such as MixSIR
et al. 2010b; MacArthur et al. 2011). (Moore and Semmens 2008; Ward et al. 2010), SIAR (Parnell et al.
A major assumption in all aforementioned mixing models is 2010), MixSIAR (Stock and Semmens 2013), IsotopeR (Hopkins and
that the food sources are homogenised fully in the consumer Ferguson 2012), FRUITS (Fernandes et al. 2014), and that of Erhardt
where they are reassembled to generate new tissues (Wolf et al. and Bedrick (2013) explicitly deal with variability among con-
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

2009). It is well known that the metabolic pathways for creating sumer and source isotopic values (which implicitly include mea-
new tissues, even within an organism, are physiologically differ- surement errors). Furthermore, these models can incorporate
ent and hence fractionate heavy and light isotopes differently uncertainty in the assigned values of DTDFs as well, which is one
(Caut et al. 2009; Martínez del Rio et al. 2009; Wolf et al. 2009). of the greatest sources of uncertainty in applying mixing models
Furthermore, food sources that differ considerably in the com- to food webs (Caut et al. 2008a; Bond and Diamond 2011). While
pounds containing N and C are likely to be processed differently Bond and Diamond (2011) criticized the sensitivity of Bayesian
within the consumer. This process is readily dealt with by apply- model results to assumed DTDF values for their particular data
ing appropriate DTDFs for a given consumer tissue and source examples, this reflects the inability of any mixing model to ade-
combination. Where this becomes more difficult is where the diet quately resolve source proportions where the isotopic values are
itself affects the DTDFs so that they become a function of the in a narrow range, particularly when that range is of similar mag-
proportion in which they and the other sources are consumed. nitude to the uncertainty in DTDF values. The ability to specify
This can occur when consumer diets comprise sources with very uncertainties for DTDFs in Bayesian models at least allows dem-
different balances in protein and carbohydrates such as nectar- onstration of the lack of precise solutions possible, which is a
and fruit-eating animals, and omnivores (Martínez del Rio et al. more honest result than the more constrained results that would
2009). In this case, not only can the relative masses of C and N be found by simply relying on a single DTDF point estimate. We
differ between the sources (which can be handled by including urge researchers to use these Bayesian mixing models and take
concentration dependence in the model), but also consuming low advantage of their ability to incorporate various sources of uncer-
For personal use only.

N or low C diets can alter the balance between exogenous (diet- tainty in their estimates of source proportions.
derived) and endogenous (recycled from other tissues) com-
pounds (Martínez del Rio et al. 2009). 10. Report distributions of results
The problem from a statistical modelling point of view is that Just as it is important to incorporate sources of variability and
the functional form describing how fractionation is affected by uncertainty and not just rely on point estimates in mixing model
diet is unknown (Martínez del Rio et al. 2009), and in reality, this inputs, researchers should report the distributions of source pro-
will likely differ by species and from case to case. Without any portion estimates and not just summary values (e.g., means or
general process to govern how this should behave, at present this medians) in mixing model outputs. As Phillips and Gregg (2003)
means that incorporating this feature may require tailor-made stated in their concluding sentence, “To avoid misrepresenting
models to be developed based on detailed empirical studies spe- the uniqueness of the results, users should report the distribution
cific to that system. Explicitly separating the amino acids in con- of feasible solutions rather than focusing on a single value such as
sumer tissues that have arisen exogenously from those derived the mean.” Even when mixing models are not underdetermined,
endogenously is one promising avenue of study for dealing with sample variability, measurement error, and uncertainty in DTDF
this issue in cases where it is suspected a priori of being of impor- values impart uncertainty in source proportion estimates. Re-
tance (Budge et al. 2011). In many cases, however, the diet compo- searchers should always strive to provide statistical bounds on
nents of consumers are not so radically different in their composition, these estimates, such as the 95% confidence intervals provided by
so this routing issue is likely to be of less concern. IsoError (Phillips and Gregg 2001a, 2001b) and the corresponding
As a practical matter, how should these issues of possible con- 95% credible intervals (Bayesian confidence intervals) provided by
centration effects and isotopic routing be handled when the food Bayesian models. Outputs from Bayesian models are true proba-
sources do vary substantially in their composition, e.g., in om- bility distributions that may be plotted and summarized with any
nivory? In such cases, we suggest using a concentration-dependent number of descriptive statistics, as well as compared with each
mixing model after correcting for differences in digestibility of other, with hypothesized distributions, or with other parameters
food sources and their constituents, unless those corrections in- of interest such as fitness (Parnell et al. 2010). Although it is intu-
dicate that there is little remaining difference in digested elemen- itive, making histograms of the estimated source proportions is
tal concentrations. Where one is concerned that routing is an extremely important—doing so helps identify posteriors that are
issue but cannot obtain detailed empirical data on the physiology relatively flat (not informed by data) and situations where poste-
of diet-affected DTDFs, we would recommend increasing the un- rior sources are multimodal (such as from a data set with two
certainty on the DTDFs (see section 4 above). In this manner, one sources that are not identifiable, because they have similar or
is treating the process as unobservable nuisance and pushing it to an opposing source signatures).
error term, rather than modelling it directly, and the uncertainty of Many authors fail to consider the impact of joint uncertainty in
the source contribution estimates will increase accordingly. the source proportions. These are provided in the output of Bayes-
ian models such as MixSIR and SIAR and allow users to spot where
9. Consider and incorporate uncertainties the isotopic arrangement of sources leads to unavoidable model
Initially, mixing models could only provide point estimates for inadequacy. In Fig. 5, the matrix plot shows that Enteromorpha
two or three source proportions as a function of point estimates and Zostera are strongly negatively correlated. This means that, if
(means) of one or two isotopic values for both the consumer and Enteromorpha is being consumed at the top of its probability range,
its food sources, with no consideration of variability among sam- then Zostera is likely to be at the bottom of its probability range,
ples or measurement error. IsoError (Phillips and Gregg 2001a, and vice versa. Thus, the marginal uncertainty (shown in the his-
2001b) was an early mixing model development that incorporates tograms) in each source proportion is artificially inflated and no

Published by NRC Research Press

 832 Can. J. Zool. Vol. 92, 2014

Fig. 5. Matrix plot of food sources for the Brent Goose (Branta bernicla) example (discussed in section 3) during their intertidal foraging period.
The diagonal cells show the posterior probability distributions for each of the four food sources. The cells below the diagonal show the
correlations between contributions for pairs of food sources. The cells above the diagonal show contours of the joint posterior probability
distribution for contributions for pairs of food sources.
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14
For personal use only.

amount of extra data collection could have reduced this uncer- Beyond these sources of uncertainty, however, the degree to
tainty. If desired, extremely negatively correlated source propor- which a mixing model is underdetermined (having many more
tions may be combined (as in the a posteriori manner described sources than isotopic tracers) is an overarching consideration.
above) with a potential gain in precision. This is a fundamental and structural source of uncertainty, as
mathematically there are a large number of solutions where the
11. Limitations
number of variables exceeds the number of constraints. In gen-
Stable isotope mixing models can be a useful tool for unraveling
eral, the more food sources there are in an animal’s diet, the less
trophic relationships in food webs and understanding the causes
precise will be the estimates of their importance. This effect can
and consequences of variation in diets, but they are just a tool and
the picture that they provide of dietary composition is not neces- be magnified by the effects of mixing space geometry as described
sarily a precise one. Common sense dictates that their limitations above. In some cases, the distributions of food source contribu-
be recognized. Variability in the isotopic composition of food tions may be so wide that it is only useful in a very qualitative
sources and individual consumers blurs the picture, as does un- sense, ruling out contributions that are less than the minimum
certainty in DTDFs. The degree to which nice constrained distri- value and more than the maximum value (Fry 2013). For such
butions of source contributions are found depends greatly on the cases, Benstead et al. (2006) stated that the most useful results are
geometry of the mixing space—where the consumer falls within those with low maxima or high minima.
the range of its food sources and how different those food sources Besides not always providing a very precise determination of
are from each other. One cannot expect a precise and unambiguous dietary importance, mixing models may not necessarily always
sorting of the contributions of dietary sources based on isotopic val- provide an accurate one if their assumptions are not met. Two of
ues if they are not fairly distinctive in their isotopic composition. those assumptions are that all sources are included and that there

Published by NRC Research Press

 Phillips et al. 833

is complete mixing. The existence of unsampled sources, concen- isotopically similar, and Chiaradia et al. (2014), which incorpo-
tration differences among sources, and metabolic routing may rated DNA analysis data into model prior distributions to further
compromise the validity of the results. The turnover rates of the constrain dietary estimates. The recent MixSIAR model (Stock and
consumer tissues sampled define the time frame over which the Semmens 2013) allows for incorporation of hierarchical popula-
diet is being integrated. Within this time frame, temporal vari- tion structure and other continuous covariates in its estimation of
ability in diet, or spatial variability if the consumer shifts loca- consumer diets. Phospholipid-derived fatty acid (PLFA) analyses,
tions, make for a “moving target” and may muddle the results, compound-specific isotopic analyses, and pulse-chase experi-
although Francis et al. (2011) modified the Bayesian mixing model ments using enriched isotopic tracers to label certain food-web
framework to account for continuous diet gradients. Similarly, components are other powerful and promising techniques that
mixing models generally do not consider any differences in prey are being increasingly used to supplement bulk natural abun-
availability, although several methods have been proposed to in- dance isotopic analyses in food-web studies (Van den Meersche
corporate this information to some extent (Phillips et al. 2006; et al. 2009; Evrard et al. 2010, 2012; Middelburg 2014).
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

Yeakel et al. 2011). As with simple linear regression, mixing mod- The mixing models described in this paper for the most part
els generally assume that errors are normally distributed. Recent focus on a single consumer species in a food web based on isotopic
mixing model tools, such as MixSIR and SIAR (Moore and composition. In contrast, linear inverse models (Vezina and Platt
Semmens 2008; Parnell et al. 2010), assume that all consumers 1988) attempt to solve matter and energy flows throughout the
have the same diet (more advanced models, like the hierarchical food web based on an array of estimates of biomass, flux, etc., and
model of Semmens et al. (2009) and Parnell et al. (2013), can ex- other constraints that is generally highly underdetermined. There
plicitly account for individual differences). Finally, the distinction is potential for some merging of these two approaches to get
must also be kept in mind that any inferences from mixing mod- better estimates of food-web processes utilizing both types of data
els refer to the assimilated diet, which can be considerably differ- and some steps have already been taken in this direction. From
ent from the ingested diet, depending on the relative digestibility the mixing model side, Kadoya et al. (2012) have developed the
of various foods. IsoWeb model, which uses stable isotope data along with a user-
As with many statistical approaches, the fitting of the model to specified food-web topology to estimate diet proportions for the
the data creates some further issues of which users should be entire food web. From the linear inverse model side, Eldridge et al.
aware (Parnell et al. 2013). The model fitting algorithm used by (2005) first incorporated carbon stable isotope data as constraints
SIAR (Parnell et al. 2010), IsotopeR (Hopkins and Ferguson 2012), in an estuarine food-web model. Recently, Pacella et al. (2013) used
and Erhardt and Bedrick’s (2013) model is called Markov chain results from the stable isotope mixing model SIAR as additional
Monte Carlo (MCMC) and involves repeatedly guessing values of constraints input into a linear inverse model of an estuarine food
For personal use only.

the dietary proportions, discarding those which are not probabi- web. These first steps show the potential for synergy between
listically consistent with the data (Parnell et al. 2010). The new these two different modelling approaches.
guesses are required to be close to the older guesses, creating a
Markov chain. At the end of the run, the user is supplied with a Conclusions
sample of posterior dietary proportions, from which they can Mixing models can quantify resource contributions to consum-
form means, standard deviations, or any other function they re- ers through analysis of stable isotopes. However, they are not the
quire. An unfortunate side effect of the MCMC algorithm is that it best tool for every job. For instance, many study systems will have
can become stuck and not find suitable high-probability values. source isotope signatures that overlap broadly in isotope space,
For this reason, the posterior sample proportions are checked to preventing source discrimination. Some consumers may feed on
see if they “converge” to some suitable stable high-probability such a diversity of prey that stable isotopes can only provide a
values. If this check is not performed, the model may be produc- blurry picture of patterns of consumption. On the other hand,
ing incorrect dietary estimates. stable isotopes may provide insights into consumer–resource re-
An alternative (but strongly related algorithm) to MCMC used lationships that would otherwise be difficult to impossible to
by MixSIR is sampling importance resampling (SIR) (Moore and quantify. We hope that the suggestions offered above provide a
Semmens 2008). Here, a large number of plausible dietary propor- helpful set of guidelines for best practices in stable isotope mixing
tions are chosen and then weighted according to their probabilis- models.
tic consistency with the data. These weights are rescaled to sum to
1 and the original plausible dietary proportions are resampled Acknowledgements
according to their weights. The downside to the SIR approach is The information in this review has been funded in part by the
that if the initial plausible set of dietary proportions does not U.S. Environmental Protection Agency (EPA). It has been subjected
contain many good values, then there will be very few large to the Agency’s peer and administrative review and has been ap-
weights (also known as a low effective sample size). If this is the proved for publication as an EPA document. Mention of trade
case, then the uncertainty in the dietary proportions will be esti- names or commercial products does not constitute endorsement
mated poorly. Both disadvantages (lack of convergence for MCMC, or recommendation for use by the EPA. We thank K.A. Hobson
badly distributed weights for SIR) can be easily discerned with and A.L. Bond for constructive comments on an earlier draft.
suitable checks provided as part of the packages.
References
12. New horizons Adams, T.S., and Sterner, R.W. 2000. The effect of dietary nitrogen content on
A number of advances have been made in stable isotope mixing trophic level 15N enrichment. Limnol. Oceanogr. 45(3): 601–607. doi:10.4319/
model methods for food-web studies in recent years, but this is a lo.2000.45.3.0601.
dynamic field that continues to move ahead. One general area in Auerswald, K., Wittmer, M.H.O.M., Zazzo, A., Schäufele, R., and Schnyder, H.
2010. Biases in the analysis of stable isotope discrimination in food webs.
which we expect further advances is bringing additional nonisotopic
J. Appl. Ecol. 47: 936–941. doi:10.1111/j.1365-2664.2009.01764.x.
constraints to bear in mixing analyses. Earlier work included Bearhop, S., Thompson, D.R., Waldron, S., Russell, I.C., Alexander, G., and
Phillips et al. (2006), which proposed a general postfiltering pro- Furness, R.W. 1999. Stable isotopes indicate the extent of freshwater feeding
cedure for any type of constraints, and Sinisalo et al. (2006), which by cormorants Phalacrocorax carbo shot at inland fisheries in England. J. Appl.
used intestinal parasite presence to further constrain estimates of Ecol. 36: 75–84. doi:10.1046/j.1365-2664.1999.00378.x.
Ben-David, M., Flynn, R.W., and Schell, D.M. 1997a. Annual and seasonal changes
the diet of seals. Later work with Bayesian models included Yeakel in diets of martens: evidence from stable isotope analysis. Oecologia, 111:
et al. (2011), which used prey abundance data to weight source 280–291. doi:10.1007/s004420050236.
estimates in model posterior distributions when sources were Ben-David, M., Hanley, T.A., Klein, D.R., and Schell, D.M. 1997b. Seasonal changes

Published by NRC Research Press

 834 Can. J. Zool. Vol. 92, 2014

in diets of coastal and riverine mink: the role of spawning Pacific salmon. Carnivore-specific stable isotope variables and variation in the foraging ecol-
Can. J. Zool. 75(5): 803–811. doi:10.1139/z97-102. ogy of modern and ancient wolf populations: case studies from Isle Royale,
Benstead, J.P., March, J.G., Fry, B., Ewel, K.C., and Pringle, C.M. 2006. Testing Minnesota, and La Brea. Can. J. Zool. 85(4): 458–471. doi:10.1139/Z07-018.
IsoSource: stable isotope analysis of a tropical fishery with diverse organic Francis, T.B., Schindler, D.E., Holtgrieve, G.W., Larson, E.R., Scheuerell, M.D.,
matter sources. Ecology, 87(2): 326–333. doi:10.1890/05-0721. PMID:16637358. Semmens, B.X., and Ward, E.J. 2011. Habitat structure determines resource
Boecklen, W.J., Yarnes, C.T., Cook, B.A., and James, A.C. 2011. On the use of stable use by zooplankton in temperate lakes. Ecol. Lett. 14: 364–372. doi:10.1111/j.
isotopes in trophic ecology. Annu. Rev. Ecol. Evol. Syst. 42: 411–440. doi:10. 1461-0248.2011.01597.x. PMID:21314881.
1146/annurev-ecolsys-102209-144726. Fry, B. 2013. Alternative approaches for solving underdetermined isotope mix-
Bond, A.L., and Diamond, A.W. 2011. Recent Bayesian stable-isotope mixing ing problems. Mar. Ecol. Prog. Ser. 472: 1–13. doi:10.3354/meps10168.
models are highly sensitive to variation in discrimination factors. Ecol. Appl. Gannes, L.Z., Martínez del Rio, C., and Koch, P. 1998. Natural abundance varia-
21(4): 1017–1023. doi:10.1890/09-2409.1. PMID:21774408. tions in stable isotopes and their potential uses in animal physiological
Budge, S.M., Wang, S.W., Hollmén, T.E., and Wooller, M.J. 2011. Carbon isotopic ecology. Comp. Biochem. Physiol. A Mol. Integr. Physiol. 119(3): 725–737.
fractionation in eider adipose tissue varies with fatty acid structure: impli- doi:10.1016/S1095-6433(98)01016-2.
cations for trophic studies. J. Exp. Biol. 214: 3790–3800. doi:10.1242/jeb. Gauthier, G., Bety, J., and Hobson, K.A. 2003. Are Greater Snow Geese capital
057596. PMID:22031744.
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

breeders? New evidence from a stable-isotope model. Ecology, 84(12): 3250–

Bugalho, M.N., Barcia, P., Caldeira, M.C., and Cerdeira, J.O. 2008. Stable isotopes 3264. doi:10.1890/02-0613.
as ecological tracers: an efficient method for assessing the contribution of Haines, E.B. 1976. Relation between the stable carbon isotope composition of
multiple sources to mixtures. Biogeosciences, 5: 1351–1359. doi:10.5194/bg-5- fiddler crabs, plants, and soils in a salt marsh. Limnol. Oceanogr. 21: 880–883.
1351-2008. doi:10.4319/lo.1976.21.6.0880.
Caut, S., Angulo, E., and Courchamp, F. 2008a. Caution on isotopic model use for Hall-Aspland, S.A., Hall, A.P., and Rogers, T.L. 2005. A new approach to the
analyses of consumer diet. Can. J. Zool. 86(5): 438–445. doi:10.1139/Z08-012. solution of the linear mixing model for a single isotope: application to the
Caut, S., Angulo, E., and Courchamp, F. 2008b. Discrimination factors (⌬15N, case of an opportunistic predator. Oecologia, 143: 143–147. doi:10.1007/s00442-
⌬13C) in an omnivorous consumer: effect of diet isotopic ratio. Funct. Ecol. 004-1783-0. PMID:15599768.
22: 255–263. doi:10.1111/j.1365-2435.2007.01360.x. Harper, G.A. 2007. Detecting predation of a burrow-nesting seabird by two in-
Caut, S., Angulo, E., and Courchamp, F. 2009. Variation in discrimination factors troduced predators, using stable isotopes, dietary analysis and experimental
(⌬15N and ⌬13C): the effect of diet isotopic values and applications for diet removals. Wildl. Res. 34(6): 443–453. doi:10.1071/WR07037.
reconstruction. J. Appl. Ecol. 46: 443–453. doi:10.1111/j.1365-2664.2009.01620.x. Heady, W.N., and Moore, J.W. 2013. Tissue turnover and stable isotope clocks to
Cherry, S.G., Derocher, A.E., Hobson, K.A., Stirling, I., and Thiemann, G.W. 2011. quantify resource shifts in anadromous rainbow trout. Oecologia, 172: 21–34.
Quantifying dietary pathways of proteins and lipids to tissues of a marine doi:10.1007/s00442-012-2483-9. PMID:23183819.
predator. J. Appl. Ecol. 48(2): 373–381. doi:10.1111/j.1365-2664.2010.01908.x. Hilderbrand, G.V., Farley, S.D., Robbins, C.T., Hanley, T.A., Titus, K., and
Chiaradia, A., Forero, M.G., McInnes, J.C., and Ramírez, F. 2014. Searching for the Servheen, C. 1996. Use of stable isotopes to determine diets of living and
true diet of marine predators: incorporating Bayesian priors into stable isotope extinct bears. Can. J. Zool. 74(11): 2080–2088. doi:10.1139/z96-236.
mixing models. PLoS ONE, 9(3): e92665. doi:10.1371/journal.pone.0092665. PMID: Hobson, K.A., and Clark, R.G. 1992. Assessing avian diets using stable isotopes I:
24667296. turnover of 13C in tissues. Condor, 94: 181–188. doi:10.2307/1368807.
Codron, D., Sponheimer, M., Codron, J., Newton, I., Lanham, J.L., and Clauss, M. Hobson, K.A., and Clark, R.G. 1993. Turnover of 13C in cellular and plasma
For personal use only.

2012. The confounding effects of source isotopic heterogeneity on consumer– fractions of blood: implications for nondestructive sampling in avian dietary
diet and tissue–tissue stable isotope relationships. Oecologia, 169: 939–953. studies. Auk, 110(3): 638–641. doi:10.2307/4088430.
doi:10.1007/s00442-012-2274-3. PMID:22349754.
Hobson, K.A., Piatt, J.F., and Pitocchelli, J. 1994. Using stable isotopes to deter-
Crawford, K., McDonald, R.A., and Bearhop, S. 2008. Applications of stable iso- mine seabird trophic relationships. J. Anim. Ecol. 63: 786–798. doi:10.2307/
tope techniques to the ecology of mammals. Mammal Rev. 38(1): 87–107. 5256.
doi:10.1111/j.1365-2907.2008.00120.x.
Hopkins, J.B., III, and Ferguson, J.M. 2012. Estimating the diets of animals using
Dalerum, F., and Angerbjörn, A. 2005. Resolving temporal variation in verte- stable isotopes and a comprehensive Bayesian mixing model. PLoS ONE, 7(1):
brate diets using naturally occurring stable isotopes. Oecologia, 144(4): 647–
e28478. doi:10.1371/journal.pone.0028478. PMID:22235246.
658. doi:10.1007/s00442-005-0118-0. PMID:16041545.
Inger, R., Ruxton, G.D., Newton, J., Colhoun, K., Mackie, K., Robinson, J.A.,
DeNiro, M.J., and Epstein, S. 1976. You are what you eat (plus a few permil): the
Jackson, A.L., and Bearhop, S. 2006. Temporal and intrapopulation variation
carbon isotope cycle in food chains. Geol. Soc. Am. (Abstracts with Programs),
in prey choice of wintering geese determined by stable isotope analysis.
8: 834–835.
J. Anim. Ecol. 75: 1190–1200. doi:10.1111/j.1365-2656.2006.01142.x. PMID:16922855.
DeNiro, M.J., and Epstein, S. 1978. Influence of diet on the distribution of carbon
Kadoya, T., Osada, Y., and Takimoto, G. 2012. IsoWeb: a Bayesian isotope mixing
isotopes in animals. Geochim. Cosmochim. Acta, 42: 495–506. doi:10.1016/
model for diet analysis of the whole food web. PLoS ONE, 7(7): e41057. doi:10.
0016-7037(78)90199-0.
1371/journal.pone.0041057. PMID:22848427.
DeNiro, M.J., and Epstein, S. 1981. Influence of diet on the distribution of ni-
Klaassen, M., Piersma, T., Korthals, H., Dekinga, A., and Dietz, M.W. 2010. Single-
trogen isotopes in animals. Geochim. Cosmochim. Acta, 45: 341–351. doi:10.
point isotope measurements in blood cells and plasma to estimate the time
1016/0016-7037(81)90244-1.
since diet switches. Funct. Ecol. 24(4): 796–804. doi:10.1111/j.1365-2435.2010.
Edwards, M.A., Derocher, A.E., Hobson, K.A., Branigan, M., and Nagy, J.A. 2011.
Fast carnivores and slow herbivores: differential foraging strategies among 01689.x.
grizzly bears in the Canadian Arctic. Oecologia, 165(4): 877–889. doi:10.1007/ Koch, P.L., and Phillips, D.L. 2002. Incorporating concentration dependence in
s00442-010-1869-9. PMID:21153738. stable isotope mixing models: a reply to Robbins, Hilderbrand and Farley
Eldridge, P., Cifuentes, L., and Kaldy, J. 2005. Development of a stable-isotope (2002). Oecologia, 133: 14–18. doi:10.1007/s00442-002-0977-6. PMID:24599364.
constraint system for estuarine food-web models. Mar. Ecol. Prog. Ser. 303: Layman, C.A., Araujo, M.S., Boucek, R., Hammerschlag-Peyer, C.M., Harrison, E.,
73–90. doi:10.3354/meps303073. Jud, Z.R., Matich, P., Rosenblatt, A.E., Vaudo, J.J., Yeager, L.A., Post, D.M., and
Erhardt, E.B., and Bedrick, E.J. 2013. A Bayesian framework for stable isotope Bearhop, S. 2012. Applying stable isotopes to examine food-web structure: an
mixing models. Environ. Ecol. Stat. 20(3): 377–397. doi:10.1007/s10651-012- overview of analytical tools. Biol. Rev. Camb. Philos. Soc. 87(3): 545–562.
0224-1. doi:10.1111/j.1469-185X.2011.00208.x. PMID:22051097.
Evrard, V., Soetaert, K., Heip, C.H.R., Huettel, M., Xenopoulos, M.A., and Lubetkin, S.C., and Simenstad, C.A. 2004. Multi-source mixing models to quan-
Middelburg, J.J. 2010. Carbon and nitrogen flows through the benthic food tify food web sources and pathways. J. Appl. Ecol. 41(5): 996–1008. doi:10.1111/
web of a photic subtidal sandy sediment. Mar. Ecol. Prog. Ser. 416: 1–16. j.0021-8901.2004.00957.x.
doi:10.3354/meps08770. MacArthur, L.D., Phillips, D.L., Hyndes, G.A., Hanson, C.E., and Vanderklift, M.A.
Evrard, V., Huettel, M., Cook, P.L.M., Soetaert, K., Heip, C.H.R., and 2011. Habitat surrounding patch reefs influences the diet and nutrition of
Middelburg, J.J. 2012. Importance of phytodetritus and microphytobenthos the western rock lobster. Mar. Ecol. Prog. Ser. 436: 191–205. doi:10.3354/
for heterotrophs in a shallow subtidal sandy sediment. Mar. Ecol. Prog. Ser. meps09256.
455: 13–31. doi:10.3354/meps09676. Martínez del Rio, C., Wolf, N., Carleton, S.A., and Gannes, L.Z. 2009. Isotopic
Fernandes, R., Millard, A.R., Brabec, M., Nadeau, M.J., and Grootes, P. 2014. Food ecology ten years after a call for more laboratory experiments. Biol. Rev.
reconstruction using isotopic transferred signals (FRUITS): a Bayesian model Camb. Philos. Soc. 84: 91–111. doi:10.1111/j.1469-185X.2008.00064.x. PMID:
for diet reconstruction. PLoS ONE, 9(2): e87436. doi:10.1371/journal.pone. 19046398.
0087436. PMID:24551057. McCutchan, J.H., Lewis, W.M., Jr., Kendall, C., and McGrath, C.C. 2003. Variation
Florin, S.T., Felicetti, L.A., and Robbins, C.T. 2011. The biological basis for under- in trophic shift for stable isotope ratios of carbon, nitrogen, and sulfur.
standing and predicting dietary-induced variation in nitrogen and sulphur Oikos, 102: 378–390. doi:10.1034/j.1600-0706.2003.12098.x.
isotope ratio discrimination. Funct. Ecol. 25: 519–526. doi:10.1111/j.1365-2435. McKechnie, A.E. 2004. Stable isotopes: powerful new tools for animal ecologists.
2010.01799.x. S. Afr. J. Sci. 100(3–4): 131–134.
Fortin, J.K., Farley, S.D., Rode, K.D., and Robbins, C.T. 2007. Dietary and spatial Middelburg, J.J. 2014. Stable isotopes dissect aquatic food webs from the top to
overlap between sympatric ursids relative to salmon use. Ursus, 18(1): 19–29. the bottom. Biogeosciences, 11: 2357–2371. doi:10.5194/bg-11-2357-2014.
doi:10.2192/1537-6176(2007)18[19:DASOBS]2.0.CO;2. Miron, M.L.L., Herrera, M.L.G., Ramirez, P.N., and Hobson, K.A. 2006. Effect of
Fox-Dobbs, K., Bump, J.K., Peterson, R.O., Fox, D.L., and Koch, P.L. 2007. diet quality on carbon and nitrogen turnover and isotropic discrimination in

Published by NRC Research Press

 Phillips et al. 835

blood of a New World nectarivorous bat. J. Exp. Biol. 209: 541–548. doi:10. protein quality on nitrogen isotope discrimination in mammals and birds.
1242/jeb.02016. PMID:16424104. Oecologia, 144: 534–540. doi:10.1007/s00442-005-0021-8. PMID:15800751.
Moore, J.W., and Semmens, B.X. 2008. Incorporating uncertainty and prior in- Rosing, M.N., Ben-David, M., and Barry, R.P. 1998. Analysis of stable isotope data:
formation into stable isotope mixing models. Ecol. Lett. 11: 470–480. doi:10. a K nearest-neighbors randomization test. J. Wildl. Manage. 62(1): 380–388.
1111/j.1461-0248.2008.01163.x. PMID:18294213. doi:10.2307/3802302.
Newsome, S.D., Phillips, D.L., Culleton, B.J., Guilderson, T.P., and Koch, P.L. 2004. Rutz, C., Bluff, L.A., Reed, N., Troscianko, J., Newton, J., Inger, R., Kacelnik, A.,
Dietary reconstruction of an early to middle Holocene human population and Bearhop, S. 2010. The ecological significance of tool use in New Caledo-
from the central California coast: insights from advanced stable isotope nian crows. Science, 329(5998): 1523–1526. doi:10.1126/science.1192053. PMID:
mixing models. J. Archaeol. Sci. 31: 1101–1115. doi:10.1016/j.jas.2004.02.001. 20847272.
Newsome, S.D., Bentall, G.B., Tinker, M.T., Oftedal, O.T., Ralls, K., Estes, J.A., and Semmens, B.X., Ward, E.J., Moore, J.W., and Darimont, C.T. 2009. Quantifying
Fogel, M.L. 2010. Variation in ␦13C and ␦15N diet–vibrissae trophic discrimi- inter- and intra-population niche variability using hierarchical Bayesian sta-
nation factors in a wild population of California sea otters. Ecol. Appl. 20(6): ble isotope mixing models. PLoS ONE, 4(7): e6187. doi:10.1371/journal.pone.
1744–1752. doi:10.1890/09-1502.1. PMID:20945772. 0006187. PMID:19587790.
Olive, P.J.W., Pinnegar, J.K., Polunin, N.V.C., Richards, G., and Welch, R. 2003. Semmens, B.X., Ward, E.J., Parnell, A.C., Phillips, D.L., Bearhop, S., Inger, R.,
Isotrophic trophic-step fractionation: a dynamic equilibrium. J. Anim. Ecol. Jackson, A.L., and Moore, J.W. 2013. Statistical basis and outputs of stable
Can. J. Zool. Downloaded from www.nrcresearchpress.com by NORTH CAROLINA STATE on 11/20/14

72: 608–617. doi:10.1046/j.1365-2656.2003.00730.x. isotope mixing models: comment on Fry (2013). Mar. Ecol. Prog. Ser. 490:
Pacella, S.R., Lebreton, B., Richard, P., Phillips, D., DeWitt, T.H., and Niquil, N. 285–289. doi:10.3354/meps10535.
2013. Incorporation of diet information derived from Bayesian stable isotope Sinisalo, T., Valtonen, E.T., Helle, E., and Jones, R.I. 2006. Combining stable
mixing models into mass-balanced marine ecosystem models: a case study isotope and intestinal parasite information to evaluate dietary differences
from the Marennes-Oléron Estuary, France. Ecol. Model. 267: 127–137. doi:10. between individual ringed seals (Phoca hispida botnica). Can. J. Zool. 84(6):
1016/j.ecolmodel.2013.07.018. 823–831. doi:10.1139/z06-067.
Painter, M.L., Chambers, C.L., Siders, M., Doucett, R.R., Whitaker, J.O., Jr., and Smith, J.A., Mazumder, D., Suthers, I.M., and Taylor, M.D. 2013. To fit or not to fit:
Phillips, D.L. 2009. Diet of spotted bats (Euderma maculatum) in Arizona as evaluating stable isotope mixing models using simulated mixing polygons.
indicated by fecal analysis and stable isotopes. Can. J. Zool. 87(10): 865–875. Methods Ecol. Evol. 4(7): 612–618. doi:10.1111/2041-210X.12048.
doi:10.1139/Z09-075. Spence, K.O., and Rosenheim, J.A. 2005. Isotopic enrichment in herbivorous
Parnell, A.C., Inger, R., Bearhop, S., and Jackson, A.L. 2010. Source partitioning insects: a comparative field-based study of variation. Oecologia, 146: 89–97.
using stable isotopes: coping with too much variation. PLoS ONE, 5(3): e9672. doi:10.1007/s00442-005-0170-9. PMID:16012818.
doi:10.1371/journal.pone.0009672. PMID:20300637. Stock, B.C., and Semmens, B.X. 2013. MixSIAR GUI user manual: version 1.0.
Parnell, A.C., Phillips, D.L., Bearhop, S., Semmens, B.X., Ward, E.J., Moore, J.W., Available from http://conserver.iugo-cafe.org/user/brice.semmens/MixSIAR.
Jackson, A.L., Grey, J., Kelly, D., and Inger, R. 2013. Bayesian stable isotope Tieszen, L.L., Boutton, T.W., Tesdahl, K.G., and Slade, N.A. 1983. Fractionation
mixing models. Environmetrics, 24: 387–399. doi:10.1002/env.2221. and turnover of stable carbon isotopes in animal tissues: implications for
Perga, M.-E., and Grey, J. 2010. Laboratory measures of isotope discrimination ␦13C analysis of diet. Oecologia, 57: 32–37. doi:10.1007/BF00379558.
factors: comments on Caut, Angulo and Courchamp (2008, 2009). J. Appl. Van den Meersche, K., Van Rijswijk, P., Soetaert, K., and Middelburg, J.J. 2009.
Ecol. 47: 942–947. doi:10.1111/j.1365-2664.2009.01730.x. Autochthonous and allochthonous contributions to mesozooplankton diet
Peterson, B.J., and Fry, B. 1987. Stable isotopes in ecosystem studies. Annu. Rev. in a tidal river and estuary: integrating carbon isotope and fatty acid con-
Ecol. Syst. 18: 293–320. doi:10.1146/annurev.es.18.110187.001453. straints. Limnol. Oceanogr. 54: 62–74. doi:10.4319/lo.2009.54.1.0062.
For personal use only.

Phillips, D.L. 2012. Converting isotope values to diet composition: the use of Vanderklift, M.A., and Ponsard, S. 2003. Sources of variation in consumer–diet
mixing models. J. Mammal. 93(2): 342–352. doi:10.1644/11-MAMM-S-158.1. ␦15N enrichment: a meta-analysis. Oecologia, 136: 169–182. doi:10.1007/s00442-
Phillips, D.L., and Eldridge, P.M. 2006. Estimating the timing of diet shifts using 003-1270-z. PMID:12802678.
stable isotopes. Oecologia, 147(2): 195–203. doi:10.1007/s00442-005-0292-0. Vaslet, A., France, C., Phillips, D.L., Feller, I.C., and Baldwin, C.C. 2011. Stable-
PMID:16341714. isotope analyses reveal the importance of seagrass beds as feeding areas for
Phillips, D.L., and Gregg, J.W. 2001a. Uncertainty in source partitioning using juveniles of the speckled worm eel Myrophis punctatus (Teleostei: Ophichthi-
stable isotopes. Oecologia, 127: 171–179. doi:10.1007/s004420000578. PMID: dae) in Florida. J. Fish Biol. 79: 692–706. doi:10.1111/j.1095-8649.2011.03052.x.
24577646. PMID:21884107.
Phillips, D.L., and Gregg, J.W. 2001b. Uncertainty in source partitioning using Vezina, A., and Platt, T. 1988. Food web dynamics in the ocean. 1. Best-estimates
stable isotopes (erratum). Oecologia, 128: 304. doi:10.1007/s004420100723. of flow networks using inverse methods. Mar. Ecol. Prog. Ser. 42: 269–287.
Phillips, D.L., and Gregg, J.W. 2003. Source partitioning using stable isotopes: Votier, S.C., Bearhop, S., MacCormick, A., Ratcliffe, N., and Furness, R.W. 2003.
coping with too many sources. Oecologia, 136: 261–269. doi:10.1007/s00442- Assessing the diet of great skuas, Catharacta skua, using five different tech-
003-1218-3. PMID:12759813. niques. Polar Biol. 26: 20–26. doi:10.1007/s00300-002-0446-z.
Phillips, D.L., and Koch, P.L. 2002. Incorporating concentration dependence Ward, E.J., Semmens, B.X., and Schindler, D.E. 2010. Including source uncer-
in stable isotope mixing models. Oecologia, 130: 114–125. doi:10.1007/ tainty and prior information in the analysis of stable isotope mixing models.
s004420100786. Environ. Sci. Technol. 44: 4645–4650. doi:10.1021/es100053v. PMID:20496928.
Phillips, D.L., Newsome, S.D., and Gregg, J.W. 2005. Combining sources in stable Ward, E.J., Semmens, B.X., Phillips, D.L., Moore, J.W., and Bouwes, N. 2011. A
isotope mixing models: alternative methods. Oecologia, 144: 520–527. doi: quantitative approach to combine sources in stable isotope mixing models.
10.1007/s00442-004-1816-8. PMID:15711995. Ecosphere, 2(2): art19. doi:10.1890/es10-00190.1.
Phillips, D.L., Schuur, E.A.G., Brooks, J.R., Ben-David, M., and Fry, B. 2006. When Webb, S.C., Hedges, R.E.M., and Simpson, S.J. 1998. Diet quality influences the
isotopes aren’t enough: using additional information to constrain mixing ␦13C and ␦15N of locusts and their biochemical components. J. Exp. Biol. 201:
problems. In Conference Abstracts and Programme of the 5th International 2903–2911. PMID:9739072.
Conference on Applications of Stable Isotope Techniques to Ecological Stud- Willson, J.D., Winne, C.T., Pilgrim, M.A., Romanek, C.S., and Gibbons, J.W. 2010a.
ies, Belfast, Northern Ireland, 13–18 August 2006. Queens University Belfast, Seasonal variation in terrestrial resource subsidies influences trophic niche
Belfast, Northern Ireland. p. 59. width and overlap in two aquatic snake species: a stable isotope approach.
Radtke, R.L., Showers, W., Moksness, E., and Lenz, P. 1996. Environmental infor- Oikos, 119: 1161–1171. doi:10.1111/j.1600-0706.2009.17939.x.
mation stored in otoliths: insights from stable isotopes. Mar. Biol. (Berl.), 127: Wilson, R.M., Chanton, J., Lewis, F.G., and Nowacek, D. 2010b. Concentration-
161–170. doi:10.1007/BF00993656. dependent stable isotope analysis of consumers in the upper reaches of a
Ramírez-Hernández, G., and Herrera, M., L.G. 2010. Nutritional importance of freshwater-dominated estuary: Apalachicola Bay, FL, USA. Estuaries and
seeds and arthropods to painted spiny pocket mice (Lyomis pictus): the effects Coasts, 33: 1406–1419.
of season and forest degradation. Can. J. Zool. 88(12): 1226–1234. doi:10.1139/ Wolf, N., Carleton, S.A., and Martínez del Rio, C. 2009. Ten years of experimental
Z10-087. animal isotopic ecology. Funct. Ecol. 23: 17–26. doi:10.1111/j.1365-2435.2009.
Reid, D.J., Quinn, G.P., Lake, P.S., and Reich, P. 2008. Terrestrial detritus supports 01529.x.
the food webs in lowland intermittent streams of south-eastern Australia: a Yeakel, J.D., Novak, M., Guimarães, P.R., Dominy, N.J., Koch, P.L., Ward, E.J.,
stable isotope study. Freshw. Biol. 53: 2036–2050. doi:10.1111/j.1365-2427. Moore, J.W., and Semmens, B.X. 2011. Merging resource availability with
2008.02025.x. isotope mixing models: the role of neutral interaction assumptions. PLoS
Robbins, C.T., Felicetti, L.A., and Sponheimer, M. 2005. The effect of dietary ONE, 6(7): e22015. doi:10.1371/journal.pone.0022015. PMID:21760944.

Published by NRC Research Press