Professional Documents
Culture Documents
Multivariate Carbon and Nitrogen Stable Isotope Model for the Reconstruction of Prehistoric Human Diet
A.W. Froehle,1* C.M. Kellner,2 and M.J. Schoeninger3,4
1 2
Department of Community Health, Wright State University, Dayton, OH 45435 Department of Anthropology, Northern Arizona University, Flagstaff, AZ 86011 3 Department of Anthropology, University of California, San Diego, La Jolla, CA 92093 4 Center for Academic Training and Research in Anthropogeny (CARTA), La Jolla, CA 92093 KEY WORDS cluster analysis; discriminant function analysis; apatite; collagen; bioarchaeology variance, providing a multivariate model for diet reconstruction. Both carbon variables dominate the rst function, while d15N most strongly inuences the second. Independent support for the functions ability to accurately classify individuals according to diet comes from a small sample of experimental rats, which cluster as expected from their diets. The new model also provides a statistical basis for distinguishing between food sources with similar isotopic signatures, as in a previously analyzed archaeological population from Saipan (see Ambrose et al.: AJPA 104(1997) 343-361). Our model suggests that the Saipan islanders 13C-enriched signal derives mainly from sugarcane, not seaweed. Further development and application of this model can similarly improve dietary reconstructions in archaeological, paleontological, and primatological contexts. Am J Phys Anthropol 147:352369, 2012. V 2011 Wiley Periodicals, Inc.
C
ABSTRACT Using a sample of published archaeological data, we expand on an earlier bivariate carbon model for diet reconstruction by adding bone collagen nitrogen stable isotope values (d15N), which provide information on trophic level and consumption of terrestrial vs. marine protein. The bivariate carbon model (d13Capatite vs. d13Ccollagen) provides detailed information on the isotopic signatures of whole diet and dietary protein, but is limited in its ability to distinguish between C4 and marine protein. Here, using cluster analysis and discriminant function analysis, we generate a multivariate diet reconstruction model that incorporates d13Capatite, d13Ccollagen, and d15N holistically. Inclusion of the d15N data proves useful in resolving protein-related limitations of the bivariate carbon model, and splits the sample into ve distinct dietary clusters. Two signicant discriminant functions account for 98.8% of the sample
Recent meta-analyses (Kellner and Schoeninger, 2007; Froehle et al., 2010) have developed a new, regressionbased model for prehistoric diet reconstruction using published experimental animal bone stable isotope data (Hare et al., 1991; Ambrose and Norr, 1993; Tieszen and Fagre, 1993; Howland et al., 2003; Jim et al., 2004; Warinner and Tuross, 2009). When carbon stable isotope ratios (13C:12C, represented as d13C) of bone collagen (d13Ccollagen) and bone apatite (d13Capatite) are plotted against one another, these experimental data sort into two discrete groups according to protein source (C3 vs. C4/marine). The resulting d13Ccollagen vs. d13Capatite regression lines for the two protein-specic groups differ in their vertical intercepts (on the d13Ccollagen axis), but not for slope (Froehle et al., 2010). Applied to diet reconstruction, the position of an individual relative to one or the other line along the d13Ccollagen axis provides isotopic information on sources of dietary protein, while placement along the d13Capatite axis indicates the ratio of C3 to C4 foods in whole diet (d13Cdiet), as Ambrose and Norr (1993) originally proposed. The regression model offers more detailed and accurate dietary estimates than either d13Ccollagen or d13Capatite does alone, and avoids much of the redundancy inherent in the apatite-collagen spacing model (the arithmetic difference between d13Capatite and d13Ccollagen values, denoted as D13Capatite-collagen: see Krueger and Sullivan, 1984; Lee-Thorp et al., 1989; Ambrose et al., 1997). Although the bivariate regression model has advantages over previous methods, a plot of archaeological human data against the experimentally derived regresC V 2011
sion lines reveals two limitations concerning protein sources (Fig. 1a). First, the model distinguishes poorly between C4 and marine sources of protein, confounding determination of protein sources in populations living in coastal areas that also harbor C4 wild vegetation or agricultural crops. Second, many individuals fall between the protein-specic lines, the meaning of which, with regard to protein sources, remains somewhat unclear. Previous work showed that swine eating an experimental diet with mixed protein ($80% C3, $20% C4) fell, as expected, slightly above the C3 protein regression line derived from rodents consuming 100% C3 protein (Froehle et al., 2010). Given the small rodent sample size and scatter within data, however, the swine were statistically indistinguishable from the 100% C3 protein rodent population. It is therefore difcult to say whether the position of the archaeological humans close to, but not on, one of the protein-specic lines reects normal variation in populations consuming monoisotopic protein, or whether it indicates mixed protein consumption.
*Correspondence to: Andrew Froehle, Department of Community Health, Wright State University, Dayton, OH. E-mail: andrew.froehle@wright.edu Received 23 June 2011; accepted 29 October 2011 DOI 10.1002/ajpa.21651 Published online 30 December 2011 in Wiley Online Library (wileyonlinelibrary.com).
353
Because we are aware of only one study that has reported experimental animal d15N values (Ambrose, 2000), the present analysis uses published data from archaeological human populations with reliable, nonisotopic evidence of diet. The experimental animal data (Ambrose and Norr, 1993; Ambrose, 2000) serve as a post hoc check of our interpretation of the human data. Finally, we provide an example application of the new model to an archaeological dataset, for which previous work left a dietary question unanswered (Ambrose et al., 1997).
Fig. 1. Means (with 95% condence intervals of the mean) for d13Ccollagen and d13Capatite in (a) archaeological human populations and (b) cluster centroids, plotted against experimental animal regression lines and their 95% prediction intervals (experimental data compiled in Froehle et al., 2010; data from Ambrose and Norr, 1993; Tieszen and Fagre, 1993; Howland et al., 2003; Jim et al., 2004; Warinner and Tuross, 2009). The long-dashed lines represent the 95% prediction interval of the mostly-C3-protein regression, while the short-dashed lines delimit the 95% prediction interval of the C4/marine-protein regression. Human data sources are in Table A1. The experimental data from which the regression lines derive have been adjusted to reect pre-industrial atmospheric carbon levels by adding 1.5% (Marino and McElroy, 1991). Cluster diets: (1) 100% C3 diet/protein; (2) 30:70 C3:C4 diet, >50% C4 protein; (3) 50:50 C3:C4 diet, marine protein; (4) 70:30 C3:C4 diet, !65% C3 protein; (5) 30:70 C3:C4 diet, !65% C3 protein.
In order to resolve these protein-related ambiguities, here we expand the model beyond the sole use of carbon variables by adding nitrogen stable isotope data (d15N). Values of d15N in bone collagen closely reect protein sources, varying with consumption of different proportions of plant vs. animal protein, as well as with terrestrial vs. marine protein sources (Schoeninger et al., 1983; Minagawa and Wada, 1984; Schoeninger and DeNiro, 1984). This study also introduces multivariate statistical techniques to the study of stable isotopes. We employ cluster analysis to investigate relationships between diet and the three isotope variables holistically, and develop a discriminant function model for the simultaneous use of the three variables in dietary reconstruction.
354
A.W. FROEHLE ET AL. using the average (UPGMA) and Wards method, two common measures in archaeological analyses (Baxter, 1994). K-means cluster analysis differs from hierarchical analysis in that it does not combine discrete groups into a unied, nested hierarchy (Everitt et al., 2001), but instead produces a user-specied number of centroids (geometric cluster centers, in this case numbering from one through six) around which group members aggregate. We used two main criteria to determine which number of clusters best reected natural groupings in the data. First, we sought the greatest level of consistency between hierarchical and k-means analyses in assignment of cases to clusters. Second, we evaluated the dietary validity of the k-means cluster centroids, which are multivariate-average cluster centers with cluster-specic average d13Ccollagen, d13Capatite, and d15N values. We only accepted centroids that we could interpret clearly in dietary terms, deriving our judgments in part from the placement of the centroids along the d13Ccollagen and d13Capatite axes relative to regression lines derived from experimental animal data (Froehle et al., 2010). We used previous evidence of diet for the populations in the dataset to check for consistency between cluster membership and cluster diet as implied by the centroids, and to further rene our centroid diet interpretations. After determining the most realistic number of clusters, we used linear discriminant function analysis to generate equations for predicting group membership from d13Ccollagen, d13Capatite and d15N. Discriminant function analysis uses known group membership from a training sample (in this case, our database) to produce linear equations (functions) describing how groups vary in terms of the independent variables (Everitt and Dunn, 2001). Our data met the assumptions of multivariate normality within groups and equality of covariance matrices between groups (Garrett, 1989; Everitt and Dunn, 2001). To estimate the resulting functions misclassication rate, we employed a post-hoc leave-one-out test, evaluating the results for consistency with classications dened by the cluster analysis. We further checked the functions validity against the experimental rat data of Ambrose and Norr (1993) and Ambrose (2000). To demonstrate how archaeological studies of diet might use our new model, we applied the discriminant functions to a population from the island of Saipan, part of the Marianas Islands in the northwestern Pacic. Archaeological and isotopic evidence for diet have remained somewhat inconclusive (Ambrose et al., 1997): though Saipans ecosystem is C3-dominated, stable isotope data indicated moderate consumption of 13C-enriched foods. Two candidates, C4 sugarcane and seaweed, were identied as potential sources of this signal, but the previous analysis offered no clear choice between the two (Ambrose et al., 1997). Here, we show how our discriminant function model can distinguish between the signals of the two foods in the bones of the Saipan islanders, a method that may be applicable to similar problems in other populations.
(Schoeninger, 1985; Phillips and Koch, 2002) and the degree of reliance on marine resources (Schoeninger et al., 1983). This latter aspect of nitrogen-isotope data is critical to the present analysis, since the relationship between d13Capatite and d13Ccollagenvalues in marine resources converges with that of C4 protein sources (Kellner and Schoeninger, 2007). By adding d15N values in this analysis, we expect to be capable of distinguishing between populations consuming C4 vs. marine protein.
355
Euclidian distance between centroids Cluster 1 13.47 13.29 4.17 10.75 Cluster 2 Cluster 3 Cluster 4
10.79 10.32
6.69
a Centroid carbon isotope values interpreted for percent C3 and C4 in whole diet using experimental animals analyzed in Froehle et al. (2010). For this comparison, experimental animal carbon data were adjusted to reect preindustrial atmospheric carbon levels by adding 1.5% (Marino and McElroy, 1991). Collagen carbon and nitrogen isotope values were interpreted for protein sources following DeNiro (1987) and Froehle et al. (2010).
their utility in precise dietary reconstruction, and were inconsistent with the substantial variability in diet within and between populations in the database. The six-group regime was thrown out because it produced a lower rate of overlap between hierarchical and k-means case assignments than did the four- and ve-group clustering regimes. Assuming ve groups, the UPGMA hierarchical and k-means methods assigned 151 of 158 (96%) cases to clusters in the same way. When we assumed four groups, UPGMA and k-means also classied 151 of 158 cases the same way, while Wards and k-means assigned a slightly higher number of cases, 154 of 158 (97%), to the same clusters. Among all such comparisons, the latter had the highest rate of consistency between clustering methods. We nevertheless chose to adopt the slightly less-consistent ve-group clustering regime because we found that it revealed more dietary discrimination than did the four-groups cluster assignments. To interpret diet for each group, we plotted the k-means cluster centroids isotope values (Table 1) against the experimental animal carbon-isotope regression lines (Fig. 1b): four of the centroids fell on or near one of the two protein-specic diet regression lines, while the fth cluster fell between the lines. It is important to note that the sample from which the C3 protein regression line is derived includes animals eating 95100% (N 5 12), 7585% (N 5 11), and 65% (N 5 1) C3 protein: as such, centroids falling on this line can be interpreted as consuming at least 65% C3 protein (although more likely !75% C3 protein). Dietary similarities and differences between clusters were also evaluated using Euclidian distances between centroids (Table 1) and the dendrograms splitting pattern (Fig. 2), both of which matched and conrmed that the two analyses organized the clusters similarly. The placement of centroids on the experimental animal plot indicates that Cluster 1 likely represents a fully C3 diet, and that Cluster 4 encompasses a 70% C3 diet with mostly C3 protein. The similarity of these two clusters relative to the rest of the sample is reected in their low-level split in the dendrogram, and their small Euclidian distance (4.17; smallest of any two centroids). The dendrograms highest-order split occurred between the high-C3 Clusters 1 and 4 on the one hand, and Clusters 2, 3, and 5 on the other. Given their placement along the d13Capatite axis on the experimental animal carbon plot, this is consistent with the latter three groups all having diets consisting of at least 50% C4 foods. Fitting with this split, the two greatest distances between centroids occurred between Clusters 1 and 2 (13.47), and between Clusters 1 and 3 (13.29). Of the three high-C4 groups, Cluster 5 is the easiest to interpret, consisting of a 70% C4 diet with mostly C3 protein. Clusters 2 and 3 defy interpretation using
Fig. 2. Simplied dendrogram showing the results of hierarchical cluster analysis. Average diet descriptions are based on centroid isotope values from the k-means cluster analysis and interpretations of those values using Froehle et al. (2010) and DeNiro (1987), as well as archaeological evidence. Proximity coefcients indicate the degree of similarity at which different groups split/merge. Higher coefcients indicate a higher tolerance for dissimilarity within groups/branches.
only carbon data because although they both clearly had overall diets high in C4 foods (70 and 50% of diet, respectively), protein sources remain indeterminate. It is clear that the vast bulk of the protein consumed by members of Cluster 3 came from 13C-enriched sources, but the carbon data do not indicate whether those were C4 or marine foods. Similarly, the position of Cluster 2 likely indicates a mix of protein from both C3 and 13C-enriched sources: although the centroid falls within the 95% prediction interval of the C4/marine protein line, that interval is quite wide due to the regression being derived from a small sample. Were the sample larger, akin to that of the C3 protein line, we would expect the Cluster 2 centroid to fall just outside the C4/marine prediction interval (see Fig. 1b) and thus between the two lines. We take this to indicate that at least some of the protein consumed by members of Cluster 2 came from C3 sources. As with Cluster 3, however, whether the 13C-enriched American Journal of Physical Anthropology
356
A.W. FROEHLE ET AL. between Clusters 2 and 3 due to their nitrogen values. On the basis of these comparisons, it seems reasonable to interpret the diets of these clusters in the following terms: Cluster 2 includes a 70% C4 diet with mixed C3 and C4 protein, while Cluster 3 encompasses a 50% C3 diet with mainly marine protein (interpretations that are also consistent with archaeological evidence for dietsee Discussion).
portion of the protein for Cluster 2 comes from C4 or marine foods is unclear from the carbon-isotope data alone. The inclusion of d15N data in the hierarchical and kmeans cluster analyses helps make sense of protein sources for Clusters 2 and 3. For example, the high-level split of Cluster 3 from Clusters 2 and 5 is suggestive of major dietary differences between these sub-branches of the dendrogram. Clusters 2 and 5 had very closely positioned k-means centroids with a Euclidian distance of just 4.28, similar to the distance between Clusters 1 and 4, and so close that the split was not visible in the fourgroups k-means clustering regime. Clusters 2 and 3 were almost twice as far from each other than were Clusters 1 and 4 (which split much lower in the dendrogram), despite smaller gaps between Clusters 2 and 3 in d13Capatite and d13Ccollagen (Cluster 1 vs. 4 differences: d13Capatite 5 3.8%, d13Ccollagen 5 1.7%; Cluster 2 vs. 3 differences: d13Capatite 5 2.8%, d13Ccollagen 5 0.8%). This pattern suggests that differences in nitrogen, not carbon, are largely responsible for the high-level split between Clusters 2 and 3, and indeed, their centroid d15N values are highly divergent: 10.1% vs. 17.7%, respectively (Clusters 1 and 4 differ by only 1.3% for d15N). Plotting the clusters in three-dimensional space (Fig. 3) emphasizes the difference
Fig. 3. Three-variable plot of archaeological sample according to k-means classications, including cluster centroids. Cluster diets: (1) 100% C3 diet/protein; (2) 30:70 C3:C4 diet, >50% C4 protein; (3) 50:50 C3:C4 diet, marine protein; (4) 70:30 C3:C4 diet, !65% C3 protein; (5) 30:70 C3:C4 diet, !65% C3 protein.
a Smaller values indicate greater effect on the discriminant function(s). All three variables contribute signicantly to the discrimination between groups.
TABLE 3. Function Eigenvalues and structure matrix Structure coefcientsb Function 1 2 3 Eigenvalue 13.402 6.939 0.243
a
13
a Eigenvalues indicate the amount of variance in the sample each function accounts for. The higher an Eigenvalue for a particular function, the higher the percentage of variance that function explains, which is also noted in this table. b Structure coefcients indicate the relative level of association between each variable and each function. Where these values are largest for a particular variable, the inuence of that variable on that function is greatest. The largest values for each variable are indicated by * above, and allow the functions to be characterized according to the key inuential variables. In this case, we can name Function 1 Carbon, and Function 2 Nitrogen. Because of its relatively low inuence on the samples variance, and because none of the variables were most highly correlated with it, we dropped Function 3.
357
(Fig. 4a,b; see Table A1 for function scores) that facilitate the application of the model to new cases.
Fig. 4. Archaeological sample discriminant function scores plotted by group, as assigned in the discriminant function analysis, along with function scores at the group centroids. Cluster diets: (1) 100% C3 diet/protein; (2) 30:70 C3:C4 diet, >50% C4 protein; (3) 50:50 C3:C4 diet, marine protein; (4) 70:30 C3:C4 diet, !65% C3 protein; (5) 30:70 C3:C4 diet, !65% C3 protein. (a) Individual data points from the archaeological sample. Using the leave-one-out technique, the discriminant function analysis classied only 4 of 158 cases (labeled on the plot) differently than did the k-means analysis. Three of these cases were statistical outliers in the original studies from which we drew the data. (b) Plot using the same data as above, but without individual data points. Error bars extend to 2r of the means of both sets of function scores (equivalent to functions at centroids) within each cluster. The limits of the boxes represent the extreme minimum and maximum individual case scores for each function within each cluster. This plot is intended to provide a clean working space for the evaluation of new cases (i.e., those not included in the present study) where diet remains indeterminate using other methods.
Nitrogen : F2 0:393 d13 Capatite 0:133 d13 Ccollagen 0:622 d15 N 8:703 We calculated function scores for the individual cases in the dataset and the cluster centroids, generating plots
358
A.W. FROEHLE ET AL. In the entire study sample, only seven individuals received conicting cluster assignments for the vegroups analysis, the UPGMA hierarchical method placing them in Cluster 2 while k-means relegated them to Cluster 5. All seven came from the broader Cahokia/ American Bottom/Illinois River regions, where the populations appear to have had rather homogeneous diets (at least relative to the overall sample in the present study), and where six of these seven individuals were statistical outliers within their respective populations (Cahokia 120: Ambrose et al., 2003; American Bottom ESLSQ 25, Range 14, IC 114, Illinois River 66a, 153: Hedman et al., 2002; see Table A1 for individual data). It makes statistical sense that different clustering methods might treat these six outliers differently, especially given rather ne isotopic and dietary distinctions between Clusters 2 and 5. The specic characteristics of these individuals, rather than general aws in the clustering regime, probably explain the classicatory discrepancies, and may point to intraregional migration. For example, Cahokia 120 appears to have eaten less C3 protein than other high status individuals at Mound 72, suggesting an origin in the Illinois River or Corbin Mounds (American Bottom) areas (Ambrose et al., 2003). The converse may be true of the three outliers from the American Bottom, all of whom appear to have eaten less C4 protein than their neighbors. These ndings may indicate an origin outside the region (Hedman et al., 2002), or possibly migration from Cahokia to the American Bottom. The Illinois River disagreements are more difcult to interpret due to small sample size and a high degree of variability at this site (Hedman et al., 2002). Finally, it remains unclear what underlies disagreement over the seventh individual, Illinois River 65 (Hedman et al., 2002), who fell into Cluster 2 per the hierarchical analysis, but Cluster 5 for the k-means analysis. Archaeological details of the burial provide no additional explanatory information (see Goldstein, 1980), but given the overall sample size, this one disagreement does not disqualify the clustering results. Three of the four misclassied individuals in the leave-one-out test of the discriminant function analysis also come from the American Bottom/Illinois River region, and are similarly explained as being statistical outliers. Two of these individuals were already noted above (ESLSQ 25, Illinois River 65) since they also received different classications from the hierarchical vs. the k-means clustering methods. A third case (IC 116) was also an outlier for the American Bottom population, although the hierarchical and k-means results did not disagree on this individuals assignment. Again, the statistical uniqueness of these individuals likely explains their misclassication by the discriminant functions. The fourth individual misclassied in the discriminant function analysis came from Tierra del Fuego (13), and is discussed below. Tierra del Fuego. No discrepancies between the hierarchical and k-means analyses occurred for any individuals from Tierra del Fuego. The sample split mainly between Cluster 3 (50% C4 diet; marine protein) and Cluster 1 (100% C3 diet), individuals in the former coming from southern and southeastern regions while those in the latter came mainly from the northern part of Tierra del Fuego (except one from Navarino Island in the south). This split agrees with the original studys suggestion that marine protein played an important dietary role in the south, while C3 guanaco meat served
individuals from Ontario, we accepted the classications as robust. Greater Cahokia region. Data from three areas within the greater Cahokia region (Pauketat, 1998) were included in this study: the American Bottom, the Illinois River Valley, and Cahokia itself. The American Bottom oodplain and upland sites together constitute the largest population in our analysis. As expected from the original study, which reported an extensive reliance on C4 foods in the overall diet (Hedman et al., 2002), 33 of 35 oodplain individuals and 13 of 17 upland individuals fell into Cluster 2. Among those assigned to Cluster 2, d15N values fell within the range of 7.910.7% (mean 5 9.3 6 0.09%), subsuming the intersite variability reported in the original study (Hedman et al., 2002), but consistent with a mixed protein base consisting of wild terrestrial and aquatic game supplemented with wild nuts and cultivated maize. The remaining two oodplain and four upland individuals split between Clusters 1 (100% C3), 4 (70% C3) and 5 (70% C4). While their d15N measurements fell well within the range of the larger population, their d13C values indicated less reliance upon C4 foods, in particular as a source of protein. Three of these individuals also caused disagreement between the hierarchical and k-means analyses, which we discuss below. The Illinois River specimens, from just outside the American Bottom and Cahokia regions proper, split mainly between the high-C4 Clusters 2 (70% C4; mixed protein) and 5 (70% C4) with a single individual falling into Cluster 4 (mostly C3). Despite their seeming randomness, these classications are consistent with archaeological evidence that some of these people incorporated a smaller proportion of C4 foods into their diet and displayed more dietary heterogeneity than their neighbors in the American Bottom (Hedman et al., 2002). On the other hand, the Illinois River population did not appear to differ from the American Bottom in terms of trophic level, demonstrated by their average d15N of 8.9 6 0.23% and the centroid d15N values for Clusters 2 and 5 (10.1% and 8.8%, respectively). We discuss the single Illinois River individual that fell into Cluster 4 below. Cahokia Mound individuals split between Clusters 4 and 5, as expected, along the social status lines reported in the original study (Ambrose et al., 2003). Low status individuals fell into Cluster 5, indicating a heavy reliance on maize as a dietary staple, but not as a source of protein. On the other hand, Cluster 4 incorporated the high status individuals, indicating that they obtained the bulk of their diet, including protein, from C3 sources, with a small overall maize component. Further, these cluster assignments are consistent with a slightly higher trophic level among the high status subgroup, whose mean d15N value was signicantly higher than the low status groups average (10.6 6 0.67% vs. 8.2 6 0.19% respectively; independent samples t-test, P \ 0.01). Interestingly, given the results from the nearly contemporaneous American Bottom (within 125225 years), the cluster assignments suggest that even low status individuals from Cahokia consumed less C4 food, especially as protein, than populations from the surrounding region. At the same time, however, low status Cahokia nitrogen-isotope values indicate a trophic position at the lower end of the range for the entire region, while the high status mean falls near the top of that range (see above). American Journal of Physical Anthropology
359
Function scores F1 F2 0.264 1.122 23.742 20.266 20.569 25.197 23.293 Cluster 1 none 2 2 4 5 5
Diet compositionb C3 non-protein/C3 protein (20%) C3 non-protein/C4 protein (5%) C4 non-protein/C3 protein (5%) C3 non-protein/C4 protein (70%) C4 non-protein/C3 protein (70%) C4 non-protein/C3 protein (20%) C4 non-protein/C3 protein (20%)
Carbon data from Ambrose and Norr (1993); nitrogen data from Ambrose (2000). Diet composition from Ambrose and Norr (1993). Protein as percent of total diet by weight in parentheses. Carbon data adjusted to reect preindustrial atmosphere by adding 1.5% (Marino and McElroy, 1991).
as the primary northern protein source (Yesner et al., 2003). Mean d15N values for the subgroups were also consistent with this interpretation, being signicantly higher in those assigned to Cluster 3 than those falling in with Cluster 1 (17.5 6 0.47% vs. 11.5 6 0.54% respectively; independent samples t-test, P \ 0.01). Our interpretation of whole diet, however, differs from the previous studys, where D13Capatite-collagen spacing values were taken to indicate a small C4 component for the northern subsample. Compared with the experimental animal regression lines (Kellner and Schoeninger, 2007; Froehle et al., 2010) instead of tissue spacing, the northern subset falls at the 100% C3 diet end of the C3 protein line. Therefore, it appears that people from northern Tierra del Fuego also differed from southern populations by eating a diet devoid of 13C-enriched foods, though we are unaware of any independent lines of evidence that support this interpretation. One individual, Tierra del Fuego 13, came from the south and was placed in Cluster 1 by the discriminant function analysis, in contrast to both clustering methods which assigned this individual to Cluster 4. The Cluster 1 assignment would associate this individual with the northern Tierra del Fuegan subgroup, reliant entirely upon C3 resources, while placement in Cluster 4 would indicate about 70% consumption of C3 foods. Tierra del Fuego 13 falls midway between the centroids of Clusters 1 and 4 for 13Capatite, but is enriched in d13Ccollagen and has higher d15N than both cluster centroids. This position suggests a diet midway between Clusters 1 and 4 in terms of C4 foods in overall diet, but also indicates the consumption of higher trophic level and more 13Cenriched protein than in either Cluster 1 or 4. This combination is consistent with a largely C3 diet similar to the northern Tierra del Fuego group, but with some incorporation of marine and perhaps C4 resources in the general fashion of the southern regions (although not sufcient to fall into Cluster 3). Thus, Tierra del Fuego 13 appears to have consumed a diet intermediate to the northern and southern patterns, which in its pairing of mainly C3 resources with some marine foods, is unique in the overall sample studied in the present meta-analysis. This individuals uniqueness may underlie the discrepancy between the discriminant function analysis and the clustering methods classications. San Nicolas Island. Finally, the entire subsample from San Nicolas Island fell into Cluster 3, the marine protein cluster. In fact, all three isotope values for the Cluster 3 centroid match almost exactly the corresponding means for San Nicolas islanders as a group. There were no dis-
agreements between hierarchical, k-means, or discriminant function analyses as to the classication of any individual from San Nicolas Island. This categorization agrees with the emphasis on marine foods described in the original study (Harrison and Katzenberg, 2003). It disagrees, however, in that it suggests a roughly 50% C4 contribution to diet, whereas the previous authors used D13Capatite-collagen values to conclude that the San Nicolas islanders consumed no C4 foods. Adjusted for the industrial effect (Marino and McElroy, 1991), mean d13Capatite and d13Ccollagen values for this sample (29.0% and 211.7%, respectively) are very similar to experimental values for rats consuming a controlled diet of 50:50 C3:C4 nonprotein with tuna sh as the protein source (Jim et al., 2004: d13Capatite 5 28.6%; d13Ccollagen 5 212.2%). Rats from the same experimental study, eating a diet of 100% C3 nonprotein with tuna sh as the protein source, have considerably less-enriched carbon-isotope values, especially for d13Capatite (213.4%). Together, our three-variable assessment and the experimental rat data strongly suggest that the San Nicolas Island population indeed exploited both marine and C4 resources. The source of the C4 signal is unknown, but others have suggested that mortars and pestles common to San Nicolas sites may have been used to process the seeds of native grasses (Meighan and Eberhart, 1953), some of which use the C4 photosynthetic pathway (Thomas, 1995). This hypothesis merits further exploration.
360
A.W. FROEHLE ET AL. 70% C3 protein by weight, fell accordingly into Cluster 4 where the centroid indicates an overall 30% C4 diet with !65% of protein coming from C3 sources. In Cluster 5, diet 12/13G fell very near the centroid while diet 6F fell within 2 standard deviations of the centroid, indicating an overall $70% C4 diet with protein coming from !65% C3 sources. This positioning matches well with the compositions of these diets, both of which were 80% C4 nonprotein by weight, with the remainder consisting entirely of C3 protein. Interestingly, despite similar placement along the Function 1/Carbon axis, diet 6F fell lower on the Function 2/Nitrogen axis than did 12/ 13G. The difference corresponds to the lower d15N values observed in the rats on diet 6F vs. 12/13G, and ts given that 6F contained protein from soy and wheat, while 12/ 13Gs protein source was milk casein (Ambrose and Norr, 1993). This nding supports the idea that positioning along the Function 2/Nitrogen axis provides information about trophic level when other aspects of diet are the same. Positioning on the discriminant function plot of diets 2B and 3C was somewhat less straightforward than for the other diets. Diet 2B fell outside of the boundaries of all Clusters, which may be explained by the fact that its specic diet (95% C3 nonprotein and 5% C4 protein by weight) was not represented in our archaeological sample. We might have expected 2B to fall closer to Cluster 1, however, considering that the vast majority of the diet carried a C3 isotopic signature. Diet 3C consisted of 95% C4 nonprotein and 5% C3 protein, and fell just within the boundaries of Cluster 2. This was to be expected in a general sense on the basis of the high C4 content of diet 3C, but with its C3 protein content, we might have predicted it to fall further to the left along the Function 1/ Carbon axis, and nearer to Cluster 5.
For the most part, given the composition of their experimental diets, the rats fell as expected on our discriminant function plot (Fig. 5). For example, the 100% C3 diet 1A fell within the range of the 100% C3 Cluster 1. Diet 5E, which was 30% C4 nonprotein and
Fig. 5. Discriminant function values calculated from published bone isotope values for rats fed experimental diets (Ambrose and Norr, 1993; Ambrose, 2000), and plotted against archaeological human clusters from this study. Rat carbon isotope values were adjusted to match the preindustrial atmosphere (and thus the archaeological data) by adding 1.5% (Marino and McElroy, 1991). Information on the composition of the experimental diets is in Table 4. Cluster diets: (1) 100% C3 diet/protein; (2) 30:70 C3:C4 diet, >50% C4 protein; (3) 50:50 C3:C4 diet, marine protein; (4) 70:30 C3:C4 diet, !65% C3 protein; (5) 30:70 C3:C4 diet, !65% C3 protein.
TABLE 5. Saipan Archaeological and foodweb data (Ambrose et al., 1997) Archaeological remainsa Function Valuesb Site MacHomes MacHomes Nansay Duty Free Site Duty Free Site Duty Free Site Duty Free Site Mean: Burial 2 5 6 1 2 3 4 d Capatite % (PDB) 210.2 27.7 29.7 29.6 28.5 29.1 210.0 29.3
13
d Ccollagen % (PDB) 218.0 218.3 219.0 218.6 218.9 218.7 218.5 218.6
13
d N % (AIR) 8.4 8.4 7.9 7.1 6.1 7.5 9.2 7.8 d13C (%) PDB
15
As reported in Ambrose et al. (1997), reecting preindustrial atmospheric carbon levels. Calculated using the present studys discriminant functions. c Plant/Diet values adjusted to preindustrial atmospheric carbon levels by adding 1.5% to reported modern values (Marino and McElroy, 1991). d Values assume enrichment of d13C in apatite by 10.2% over diet (Froehle et al., 2010), and reect expected values where the food in question constitutes 100% of diet. e Values based on C3 protein linear equation for d13Ccollagen vs. d13Cdiet from experimental data (Froehle et al., 2010), assuming the food in question constitutes 100% of diet.
361
rats on diets 2B and 3C were very similar to one another, and intermediate to those of rats consuming either C3 or C4 protein in sufcient amounts (see Table 4). A similar pattern of divergence from population-based expectations in human individuals may indicate protein deciencies, and is worth exploring further. Comparing skeletal remains with evidence for protein malnutrition to apparently healthy individuals within the same population, where diet is otherwise known to be homogeneous, could test the validity of this particular application of the model.
Fig. 6. Collagen and apatite carbon stable isotope values for Saipan islanders from Ambrose et al. (1997) plotted against compiled experimental animal data (see Froehle et al., 2010). Long-dashed lines represent the 95% prediction interval of the mostly-C3-protein regression, while the short-dashed lines dene the 95% prediction interval of the C4/marine-protein regression. Human data sources are in Table A1. The experimental data from which the regression lines derive have been adjusted to reect pre-industrial atmospheric carbon levels by adding 1.5% (Marino and McElroy, 1991). We used the position of the Saipan data relative to the experimental animal regression lines to determine protein sources as well as percent C3 diet. The C4/marine protein line is derived from animals consuming either 100% C4 protein or 100% marine protein. The C3 protein line is derived from animals consuming !65% C3 protein (13 of 25 individual animals in this group ate 5-35% C4 marine protein; see Froehle et al., 2010). On the basis of this comparison, Saipan islanders appear to have consumed mostly C3 protein, up to 35% marine protein, and on average had a diet consisting of $65 to 70% C3 foods. This is consistent with the assessment of Ambrose et al. (1997), and falls closest to this studys Cluster 4, the centroid for which is plotted here along with error bars delimiting two sample standard deviations from the mean. Cluster diets: (1) 100% C3 diet/protein; (2) 30:70 C3:C4 diet, >50% C4 protein; (3) 50:50 C3:C4 diet, marine protein; (4) 70:30 C3:C4 diet, !65% C3 protein; (5) 30:70 C3:C4 diet, !65% C3 protein.
Low protein content (5% of diet by weight) was common to both 2B and 3C, and may explain their somewhat abnormal positioning relative to the other rat diets and this studys clusters. Ambrose (2000) reports that rats from litters eating diets 2B and 3C were proteinlimited, since they achieved adult weights much lower than those on normal-to-high-protein diets. It may therefore be the case that the rats on diets 2B and 3C synthesized amino acids endogenously from nonprotein dietary precursors in order to build bone collagen, rather than incorporating most or all of those amino acids intact from dietary protein (see Schwarcz, 2000). In both cases, the low protein rats fell approximately where expected along the Function 2/Nitrogen axis, but were shifted $4 to 6 units to the right of the centroids for their expected clusters along the Function 1/Carbon axis. This nding makes sense, in that (barring starvation) tissue nitrogen will be derived largely from dietary protein, no matter its percentage contribution to diet. On the other hand, collagen carbon could be skewed toward nonprotein dietary components when protein intake is very low, which may explain why d13Ccollagen values for
362
1.77 1.02 1.08 1.25 0.95 1.10 1.41 1.22f EDe
TABLE 6. Expected stable isotope and discriminant function values for Saipan hypothetical diets
d N% (AIR)d
Sugarcaneb
d Ccollagen % (PDB)
15
d Capatite % (PDB)
13
d Cdiet % (PDB)
13
F1
F2
EDe
We then modeled expected human bone isotope values reective of each hypothetical diet (Table 6), assuming an enrichment of 10.2% in apatite over d13Cdiet (see Froehle et al., 2010) and using the d13Ccollagen /d13Cdiet equation from Froehle et al. (2010: pooled C3 protein equation, Table 2, p. 2664) to account for enrichment of d13C in bone collagen. We did not adjust d15N values for differential contributions from seaweed vs. sugarcane since both foods are very low in protein (Ambrose et al., 1997) and thus unlikely to differentially affect tissue nitrogen-isotope ratios. Individual hypothetical bone carbon-isotope values and measured nitrogen-isotope values were used to derive corresponding discriminant function scores for each hypothetical diet (Table 6). Euclidian distances were then calculated between each individuals scores for actual bone isotope measurements and their respective hypothetical diet-specic scores for the corresponding seaweed or sugarcane diets. Euclidian distances of the bone data from their respective hypothetical diet points provided information on the proportions of sugarcane and seaweed responsible for the 13C-enriched isotopic signal on Saipan. On average, actual bone values fell signicantly closer (independent samples t-test, two-tailed P \ 0.01) to sugarcane (mean Euclidian distance 5 0.60) than to seaweed (mean Euclidian distance 5 1.22). As a population, the Saipan islanders actual discriminant function scores fell almost entirely within the range expected for a diet
F1
F2
Fig. 7. Plot of mean discriminant function values for actual bone vs. the hypothetical sugarcane and seaweed diets. Error bars indicate the 95% condence intervals of the means for each function. The mean Euclidian distance between actual bone and sugarcane of 0.60 was signicantly smaller than the distance of 1.22 between bone and seaweed (independent samples t-test, two-tailed P < 0.01). One-way ANOVA revealed a signicant difference between group mean Carbon/F1 function values (P < 0.01). Post-hoc analysis (Tukey HSD) showed that for Carbon/ F1, seaweed differed signicantly from bone (P 5 0.03) and sugarcane (P < 0.01), but the latter two did not differ from one another signicantly (P 5 0.86). The three groups of function scores did not differ for Nitrogen/F2 (one-way ANOVA; P 5 0.25), which follows from the manner in which we modeled the hypothetical diets (see text for a detailed description).
13
C3:C4 dieta
MacHomes 2 MacHomes 5 Nansay 6 Duty Free Site Duty Free Site Duty Free Site Duty Free Site Mean:
1 2 3 4
MULTIVARIATE STABLE ISOTOPE DIET RECONSTRUCTION consisting of roughly two-thirds C3 foods and one-third sugarcane, and fell completely outside the range of a two-thirds C3/one-third seaweed diet (Fig. 7). Mean Function 1/Carbondiscriminant function scores did not differ between actual bone and sugarcane, whereas both differed signicantly from the hypothetical seaweed diet (one-way ANOVA, P \ 0.01; post hoc Tukey HSD: seaweed vs. bone, P 5 0.03; seaweed vs. sugarcane, P \ 0.01; sugarcane vs. bone, P 5 0.86). In contrast, none of the diets Function 2/Nitrogen means differed (one-way ANOVA, P 5 0.25), which makes sense given that d15N values exert the greatest inuence on Function 2/Nitrogen values, and we did not model dietary differences in d15N values due to sugarcane or seaweed. If our assumptions are valid, then these results suggest that the Saipan population obtained the bulk of its non-C3 dietary component from sugarcane, and not from seaweed. Although no archaeological evidence of sugarcane harvesting exists, the relatively high prevalence of dental caries in some Saipan populations vs. other nearby islands (various sources, discussed in Ambrose et al., 1997:359) appears to corroborate our ndings. There is variation within the group, however, with two individuals (MacHomes 5 and Duty Free 2) falling midway between their hypothetical sugarcane and seaweed diet points. Some individuals may have thus consumed seaweed in addition to sugarcane, but overall our model points to sugarcane as the primary producer of the 13C-enriched signal in the bones of these Saipan islanders.
363
CONCLUSIONS
Stable isotope data reect biological processes subject to numerous inuences that result in widely varying outcomes. As such, biometric statistical techniques, which provide probabilistic statements about the relationships between animal tissues and the macronutrients in their diets (e.g., Sokal and Rohlf, 1997; Phillips and Koch, 2002; Bocherens, 2009), offer the best method for understanding, predicting, and interpreting this variation. Recent work applying these techniques to experimental data (e.g., linear regression analysis: Kellner and Schoeninger, 2007; Froehle et al., 2010) is expanded here by applying multivariate cluster analysis and discriminant function analysis to study three isotope variables simultaneously in their relationships to diet. This analysis claried two aspects of diet left ambiguous by previous d13Ccollagen vs. d13Capatite regression models (Kellner and Schoeninger, 2007; Froehle et al., 2010). As expected, the inclusion of d15N data discriminated between marine and C4 protein consumers where carbon-isotope values alone could not. Therefore, in the absence of strong supporting archaeological data, it is important to measure both d13C and d15N from collagen when reconstructing diet in populations from regions where both marine and C4 protein are available. With regard to determining protein sources from positioning relative to the C3- and C4/marine-protein regression lines in the carbon model, the present analysis suggests, at least at the population level, that placement between the lines does indeed indicate mixed protein consumption. This is evident from the placement of Cluster 2 (see Fig. 1b), which included
the Ontario Agricultural and American Bottom populations that clearly obtained both C3 and C4 protein, but is not necessarily true for individuals given considerable scatter around population means even in people with no access to mixed protein, such as Ontario preagriculturalists. The distinction between individuals and populations is an important one for the application of the discriminant function model. While the position of any individual relative to the clusters would provide at least some indication of diet, some overlap between the ranges of diet clusters and their members could weaken any conclusions about individuals. This approach is betterapplied to populations, although samples need not be largefor example, the model was quite capable of making clear dietary distinctions between the northern (N 5 5) and southern (N 5 6) samples from Tierra del Fuego, or between high (N 5 3) and low (N 5 5) status groups from Cahokia. In large populations, this model could reveal nuances of dietary differentiation between sexes, age classes, and social status groups, as well as between regions and over time within societies. Differences within societies, however, may be graded (rather than discreet) or too negrained for the present model to detect, given the wide range of diets included in this sample. In those cases, it would be more productive to apply the same cluster analysis and discriminant function techniques within those populations to look for patterns of distinction. Finally, we should emphasize that the type of negrained, analysis we conducted on the Saipan data requires thorough study of the food web within which a population likely lived (see also Crowley, in press), without which reliable hypothetical diets could not be generated. Contextual information is critical to drawing the strongest conclusions about diet from the discriminant function model, although in the absence of any other information, the model may still have utilityplacement of an individual or population relative to this studys clusters could be used to generate hypotheses, and to direct other, nonisotopic avenues of research. There is considerable room to expand upon and rene this model. The outcomes of cluster analysis depend heavily on the specic sample used, and while the archaeological populations we analyzed here cover a wide range of diets, the addition of data from people eating other diets would likely alter slightly, but strengthen the results. Also, the present model was derived from data on well-preserved bone from healthy adults, potentially limiting the models applicability to other samples, such as children or individuals with clear evidence of malnutrition. Finally, further controlled diet experiments measuring d13Capatite, d13Ccollagen and d15N, would also provide a good test of this studys results, and would allow for the development of a similar model under circumstances of known whole diet and protein d13C and d15N. Continued work in this direction should produce even more effective tools for testing competing hypotheses about prehistoric diets.
ACKNOWLEDGMENTS
The authors thank Alyssa Crittenden and Andrew Somerville for helpful comments on early drafts of this article. They also thank two anonymous reviewers for their thoughtful comments and suggestions. American Journal of Physical Anthropology
TABLE A1. Archaeological bone stable isotope database and cluster centroids
364
Function scores
Populationa
d13Capatite % (PDB)b
d13Ccollagen % (PDB)b
SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf SOf
213.6 29.5 211.9 212.3 215.2 215.4 215.1 214.8 215.1 29.1 214.4 214.9 214.2 214.9 214.8 212.5 212.1 211.4 211.3 210.5 212.9 211.0 214.4 26.6 24.4 25.8 24.6 25.8 23.9 24.7 24.1 25.0 25.1 25.3 24.3 25.4 24.5 24.8 25.2 25.4 25.4 25.5 24.9 25.9 25.4 25.1
TABLE A1. (Continued) Cluster assignment d13Ccollagen % (PDB)b Site name ID F1 Hc Kd DFe F2 Time period d13Ncollagen % (AIR)b Function scores
Populationa
d13Capatite % (PDB)b
SO ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg ABg 11.0 7.9 9.1 8.5 9.0 8.1 9.0 9.4 9.8 9.3 9.1 8.6 10.2 10.2 9.3 9.7 9.2 8.5 8.7 8.2 9.1 9.5 10.3 10.2 10.1 9.3 9.7 9.9 9.0 10.2 9.2 8.3 9.3 9.3 9.6 10.9 9.1 9.4 8.8 8.7 9.5 9.5 9.1 9.6 9.6 8.9 8.9 Ossossane Ossuary ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ ESLSQ Florence St Florence St Florence St Florence St Florence St Florence St Florence St Florence St Range Range Range Range Range Range Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds Corbin Mounds 6 9 17 25 27 39 42 56 77 4 5 45 80 misc 52 14 38 48 67 68 76 78 12 21 29 30 31 34 14 29/9 18 19 13 16 20 14 IC 114 IC 119 IC 116 IC 118 IC 117 IC 46 IC 47 IC 120 IC 123 IC 124 IC 122 1636 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 1275 CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE CE 2 2 2 2h 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2h 2h 5 2 2 2 2 2 2 2 2 2 2 2 2 5 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 5 5 5 2 2 2 2 2 2 2 2 2 2 2 2 2h 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 5 5 5 5h 2 2 2 2 2 2 2 2 2.446 0.471 1.412 20.427 2.338 1.121 1.878 1.262 1.321 1.398 1.058 1.942 1.987 0.335 1.442 3.095 4.000 1.239 2.343 2.184 2.530 2.921 0.354 1.683 2.155 0.448 2.792 1.532 1.892 1.704 2.627 1.464 1.953 2.610 1.661 20.302 21.331 22.084 0.603 1.182 1.181 0.390 0.824 1.266 1.920 2.039 2.750
24.1 26.1 25.8 26.1 24.9 25.1 23.8 25.3 25.9 25.3 24.6 25.6 25.0 26.5 25.8 24.8 23.7 25.9 25.1 24.8 24.1 24.3 25.4 25.7 25.5 26.0 24.2 25.5 24.7 25.4 23.9 25.1 24.5 24.0 25.1 25.1 25.5 28.3 22.9 23.3 24.3 27.0 25.6 25.0 23.4 23.9 23.5
211.0 211.9 211.1 213.3 210.2 211.5 211.3 211.6 211.4 211.4 212.1 210.3 211.0 212.6 211.1 29.4 28.5 211.1 210.0 210.2 210.3 29.8 213.1 211.1 210.5 212.4 210.1 211.3 210.9 211.2 210.3 211.1 211.0 210.3 211.2 214.3 215.0 214.9 213.4 212.4 212.2 212.1 212.0 211.8 211.6 211.0 210.2
21.713 22.968 22.233 22.812 22.517 23.202 23.139 22.347 21.780 22.345 22.875 22.535 21.838 21.480 22.146 22.046 22.651 22.592 22.630 23.085 22.826 22.420 21.898 21.595 21.674 22.197 22.337 21.399 22.695 21.732 22.812 23.000 22.588 22.729 22.248 21.833 22.883 21.551 23.878 23.644 22.727 21.634 22.426 22.342 22.963 23.129 23.179
365
366
Function scores
Populationa
d13Capatite % (PDB)b
d13Ccollagen % (PDB)b
AB ABg ABg ABg ABg ABg IRVi IRVi IRVi IRVi IRVi IRVi IRVi IRVi IRVi CM72j CM72j CM72j CM72j CM72j CM72j CM72j CM72j TDFk TDFk TDFk TDFk TDFk TDFk TDFk TDFk TDFk TDFk TDFk SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg
23.1 26.1 22.3 27.8 211.8 214.0 29.3 25.0 26.7 26.2 25.4 24.4 28.7 29.2 26.1 29.0 26.7 210.0 210.0 21.6 24.4 24.6 23.8 215.7 215.9 214.8 28.1 29.7 210.6 27.9 210.6 210.7 213.4 213.9 26.9 26.7 29.2 28.2 26.8 28.1 27.2 27.0 26.9 25.4 26.2 27.5 26.7
TABLE A1. (Continued) Cluster assignment d13Ccollagen % (PDB)b Site name ID F1 3.564 1.746 3.963 1.988 1.894 2.752 4.906 2.692 3.107 5.029 3.365 3.130 2.005 2.728 2.193 3.610 2.534 5.290 27.567 1.731 3.046 25.276 21.730 A B C: 256-91 256-93 256-98 256-99 256-101 256-104 256-105 25-107.1 25-107 28-F-152 ACC 5-A16 ACC16 2 3 221265 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1350 BCE 1650 CE 1650 CE 1650 CE 1650 CE 1650 CE 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 Hc Kd DFe SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI SNI 16 16 16 16 16 16 16 16 16 16 16 16 16 18 18 18 18 18 Time period 210.3 211.6 29.9 211.2 211.6 210.9 28.8 211.1 210.5 28.5 29.7 210.5 210.7 210.6 211.3 210.2 211.2 28.8 17.9 17.0 20.6 15.6 16.5 17.8 16.4 17.6 17.8 16.7 17.5 18.2 14.9 16.4 17.9 19.1 17.8 20.8 12.0 10.1 17.7 11.4 8.8 220.3 211.3 210.5 218.6 215.3 d13Ncollagen % (AIR)b Function scores F2 3.773 3.865 6.095 2.733 3.240 4.063 2.332 3.755 4.038 2.755 4.272 4.365 2.600 3.153 4.426 4.886 4.023 5.776 1.905 21.977 4.013 0.230 22.962
Populationa
d13Capatite % (PDB)b
SNI SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg SNIg Cluster centroidsl Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5
26.9 29.0 28.4 28.2 28.2 28.0 25.1 27.6 27.8 25.6 28.6 28.0 28.8 27.8 28.9 27.8 28.0 26.9
Original sources of data indicated below. Data as reported in original studies, reecting preindustrial atmospheric carbon levels (Marino and McElroy, 1991). From UPGMA agglomerative hierarchical cluster analysis, splitting into ve groups. d From ve-groups k-means cluster analysis. e From discriminant function analysis. f Indicates different classication than obtained using k-means analysis. g Harrison and Katzenberg (2003); SO, Southern Ontario; SNI, San Nicolas Island. h Hedman et al. (2002); AB, American Bottom. i Hedman et al. (2002), originally from Schober (1998); IRV, Illinois River Valley. j Ambrose et al. (2003); CM72, Cahokia Mound 72. k Yesner et al. (2003); TDF, Tierra del Fuego. l Data as determined from the present studys cluster analyses and discriminant function analysis. See Table 1 for descriptions of average diet.
367
368
LITERATURE CITED
Amarger NA, Mariotti A, Mariotti F, Durr JC, Bourguignon C, Lagacherie B. 1979. Estimate of symbiotically xed nitrogen in eld grown soybeans using variations in 15N natural abundance. Plant and Soil 52:269280. Ambrose SH. 2000. Controlled diet and climate experiments on nitrogen isotope ratios of rats. In: Ambrose SH, Katzenberg MA, editors. Biogeochemical approaches to paleodietary analysis. New York: Kluwer Academic. p 243259. Ambrose SH, Norr L. 1993. Experimental evidence for the relationship of the carbon stable isotope ratios of whole diet and dietary protein to those of bone collagen and carbonate. In: Lambert JB, Grupe G, editors. Prehistoric human bone: archaeology at the molecular level. Berlin: Springer-Verlag. p 138. Ambrose SH, Buikstra JE, Krueger HW. 2003. Gender and status differences in diet at Mound 72, Cahokia, revealed by isotopic analysis of bone. J Anthropol Archaeol 22:217226. Ambrose SH, Butler BM, Hanson DB, Hunter-Anderson RL, Krueger HW. 1997. Stable isotope analysis of human diet in the Marianas Archipelago, Western Pacic. Am J Phys Anthropol 104:343361. Baxter MJ. 1994. Exploratory multivariate analysis in archaeology. Edinburgh: Edinburgh University Press. Bocherens H. 2009. Neanderthal dietary habits: review of the isotopic evidence. In: Hublin J-J, Richards MP, editors. The evolution of hominin diets: integrating approaches to the study of Palaeolithic subsistence. Dordrecht, Netherlands: Springer. p 241250. Codron J, Codron D, Lee-Thorp JA, Sponheimer M, Bond WJ, de Ruiter D, Grant R. 2005. Taxonomic, anatomical, and spatio-temporal variations in the stable carbon and nitrogen isotopic compositions of plants from an African savanna. J Archaeol Sci 32:17571772. Crawford GW, Smith DG, Bowyer V. 1997. Dating the entry of corn (Zea mays) into the lower Great Lakes region. Am Antiq 62:112119. Crowley BE. In press. Isotopic applications to primate feeding ecology: techniques, pitfalls, and creative applications for primatologists. Int J Primatol. DeNiro MJ. 1987. Stable isotopy and archaeology. Am Sci 75:182191. DeNiro MJ, Epstein S. 1978. Inuence of diet on the distribution of carbon-isotopes in animals. Geochim Cosmochim Acta 42:495506. DeNiro MJ, Epstein S. 1981. Inuence of diet on the distribution of nitrogen isotopes in animals. Geochim Cosmochim Acta 45:341351. DeNiro MJ, Schoeninger MJ. 1983. Stable carbon and nitrogen isotope ratios of bone collagen: variations within individuals, between sexes, and within populations raised on monotonous diets. J Archaeol Sci 10:199203. Delwiche CC, Zinke PJ, Johnson CM, Virginia RA. 1979. Nitrogen isotope distribution as a presumptive indicator of nitrogen xation. Bot Gaz 140 (Suppl):6569. Everitt BS, Dunn G. 2001. Applied multivariate data analysis, 2nd ed. London: Arnold. Everitt BS, Landau S, Leese M. 2001. Cluster analysis, 4th ed. London: Arnold. Froehle, AW, Kellner CM, Schoeninger MJ. 2010. FOCUS: effect of diet and protein source on carbon stable isotope ratios in collagen: follow up to Warinner and Tuross (2009). J Archaeol Sci 37:26622670. Garrett RG. 1989. The chi-square plot: a tool for multivariate outlier recognition. J Geochem Exploration 32:319341. Goldstein LG. 1980. Mississippian mortuary practices. Evanston, IL: Northwestern University Archaeological Program, Scientic Papers No. 4. Hare PE, Fogel ML, Stafford TW, Mitchell AD, Hoering TC. 1991. The isotopic composition of carbon and nitrogen in individual amino acids isolated from modern and fossil proteins. J Archaeol Sci 18:277292. Harrison RG, Katzenberg MA. 2003. Paleodiet studies using stable carbon-isotopes from bone apatite and collagen: examples
from Southern Ontario and San Nicolas Island, California. J Anthropol Archaeol 22:227244. Hedman K, Hargrave EA, Ambrose SH. 2002. Late Mississippian diet in the American Bottom: stable isotope analyses of bone collagen and apatite. Midcont J Archaeol 27:237271. Howland MR, Corr LT, Young SMM, Jones V, Jim S, Van der Merwe NJ, Mitchell AD, Evershed RP. 2003. Expression of the dietary isotope signal in the compound-specic d13C values of pig bone lipids and amino acids. Int J Osteoarchaeol 13:5465. Jim S, Ambrose SH, Evershed RP. 2004. Stable carbon isotopic evidence for differences in the dietary origin of bone cholesterol, collagen and apatite: Implications for their use in palaeodietary reconstruction. Geochim Cosmochim Acta 68:6172. Kellner CM, Schoeninger MJ. 2007. A simple carbon-isotope model for reconstructing prehistoric human diet. Am J Phys Anthropol 133:11121127. Krueger HW, Sullivan CH. 1984. Models for carbon-isotope fractionation between diet and bone. In: Turnlund JR, Johnson PE, editors. Stable isotopes in nutrition. Washington, DC: American Chemical Society Symposium Series. p 205220. Lee-Thorp JA, Sealy JC, van der Merwe NJ. 1989. Stable carbonisotope ratio differences between bone collagen and bone apatite, and their relationship to diet. J Archaeol Sci 16:585599. Marino BD, McElroy MB. 1991. Isotopic composition of atmospheric CO2 inferred from carbon in C4 plant cellulose. Nature 349:127131. Meighan CW, Eberhart H. 1953. Archaeological resources of San Nicolas Island, California. Am Antiq 19:109125. Minagawa M, Wada E. 1984. Stepwise enrichment of 15N along food chains: further evidence and the relation between d15N and animal age. Geochim Cosmochim Acta 48:11351140. Muzuka A. 1999. Isotopic compositions of tropical East African ora and their potential as source indicators of organic matter in coastal marine sediments. J Afr Earth Sci 28:757766. Pauketat TR. 1998. Reguring the archaeology of Greater Cahokia. J Archaeol Res 6:4589. Phillips DL, Koch PL. 2002. Incorporating concentration dependence in stable isotope mixing models. Oecologia 130:114125. Rennie DA, Paul EA, Johns LE. 1976. Natural nitrogen-15 abundance of soil and plant samples. Can J Soil Sci 56:4350. Schober T. 1998. Reinvestigation of maize introduction in WestCentral Illinois. Paper-in-lieu-of-Masters thesis: University of Illinois, Urbana-Champaign. Schoeninger MJ. 1985. Trophic level effects on 15N/14N and 13 12 C/ C ratios in bone collagen and strontium levels in bone mineral. J Hum Evol 14:151525. Schoeninger MJ, DeNiro MJ. 1984. Nitrogen and carbon isotopic composition of bone collagen from marine and terrestrial animals. Geochim Cosmochim Acta 48:625639. Schoeninger MJ, DeNiro MJ, Tauber H. 1983. Stable nitrogen isotope ratios of bone collagen reect marine and terrestrial components of prehistoric human diet. Science 220:13811383. Schwarcz HP. 2000. Some biochemical aspects of carbon isotopic paleodiet studies. In: Ambrose SH, Katzenberg MA, editors. Biogeochemical approaches to paleodietary analysis. New York: Kluwer Academic. p 189210. Schwarcz HP, Melbye J, Katzenberg MA, Knyf M. 1985. Stable isotopes in human skeletons of Southern Ontario: reconstructing palaeodiet. J Archaeol Sci 12:187206. Schwarcz HP, Schoeninger MJ. 1991. Stable isotope analyses in human nutritional ecology. Yrbk Phys Anthropol 34:283321. Shearer G, Kohl DH. 1978. 15N abundance in N-xing and nonN-xing plants. In: Frigerio A, editor. Recent developments in mass spectrometry in biochemistry and medicine, Vol. 1. New York: Plenum Press. p 605622. Smith DG, Crawford GW. 1997. Recent developments in the archaeology of the Princess Point Complex in Southern Ontario. Can J Archaeol 21:932. Sokal RR, Rohlf FJ. 1997. Biometry. New York: W.H. Freeman & Co. Sullivan CH, Krueger HW. 1981. Carbon-isotope analysis of separate chemical phases in modern and fossil bone. Nature 292:333335.
369
Virginia RA, Delwiche CC. 1982. Natural N abundance of presumed N2-xing and non-N2-xing plants from selected ecosystems. Oecologia 54:317325. Wada E, Hattori A. 1976. Natural abundance of 15N in particulate organic matter in the North Pacic Ocean. Geochim Cosmochim Acta 40:249251. Warinner C, Tuross N. 2009. Alkaline cooking and stable isotope tissue-diet spacing in swine: archaeological implications. J Archaeol Sci 36:1690167. Yesner D, Figuerero Torres MJ, Guichon RA, Borrero LA. 2003. Stable isotope analysis of human bone and ethnohistoric subsistence patterns in Tierra del Fuego. J Anthropol Archaeol 22:279291.