Introduction
Paleoecological studies frequently make use of relationships between morphology and ecology discovered in living organisms to infer ecology in fossil taxa. While this long-standing practice is often the best method for obtaining ecological information about past ecosystems, recent applications of phylogenetic comparative methods (e.g., Maestri et al. Reference Maestri, Luza, de Barros, Hartz, Ferrari, de Freitas and Duarte2016; Verde Arregoitia et al. Reference Verde Arregoitia, Fisher and Schweizer2017; Vidal-García and Scott Keogh Reference Vidal-García and Scott Keogh2017; Tseng and Flynn Reference Tseng and Flynn2018; Tarquini et al. Reference Tarquini, Chemisquy and Prevosti2020) offer robust tests of one of the fundamental assumptions of this method, namely, that ecology constrains the evolution of morphology. While ecology shapes morphology via natural selection, there is also often substantial phylogenetic constraint on the morphology of a given organism (Vidal-García and Scott Keogh Reference Vidal-García and Scott Keogh2017; Jones et al. Reference Jones, Benitez, Angielczyk and Pierce2018; Wong et al. Reference Wong, Chan and Chan2018; Tarquini et al. Reference Tarquini, Chemisquy and Prevosti2020). This problem becomes particularly important when the evolutionary histories of the extant organisms used to develop the ecological proxy differ in some important way from the fossils under study. In particular, changes in the phylogenetic affinities of organisms with a particular ecology (e.g., if the burrowing mammal guild in North America were to be taken over by marsupials) can lead to spurious ecological correlations with morphology that may reflect historical artifact, rather than the expected natural selection constraint on morphology.
Although evolutionary constraint on ecological proxies may seem inconsequential if one is only concerned with the match between ecology and morphology and not how it evolved, the issue should still not be ignored. A significant correlation between ecology and morphology does not demonstrate causation; therefore the danger of finding spurious correlations due to phylogenetic covariation is high (e.g., Lauder Reference Lauder1990; Wainwright and Reilly Reference Wainwright and Reilly1994). If the relationship observed in modern organisms reflects a limited number of natural selective events within phylogenetically restricted groups, and it is applied to fossil groups where the evolutionary potential and ecological context differs, the ecology inferred will be uncertain. Thus, in the absence of demonstrable performance consequences for a change in morphology on resource use, only a morphological adaptation that evolves consistently in response to a given natural selective pressure within multiple independent phylogenetic lineages offers a truly robust ecological proxy.
The evolution of mammalian diet is an area of active research, as the features that make mammals unique are, in many ways, a reflection of their highly specialized dietary adaptations. In carnivoran mammals, the dental morphology, especially of the carnassial teeth, is commonly used to determine the composition of the diet (Popowics Reference Popowics2003; Van Valkenburgh Reference Van Valkenburgh2007), although there has also been work showing substantial phylogenetic constraint in carnassial morphology (Meloro and Raia Reference Meloro and Raia2010). Various proxies proposed for diet may differ in the degree to which evolutionary constraints confound the effort to reconstruct paleoecology, so in this study, we examine several commonly applied proxies for diet in fossil carnivorans in a phylogenetic comparative framework and use non-phylogenetic statistics to determine how effective these proxies are in predicting diet of carnivorans and to what degree that effectiveness is structured by the distribution of morphologies and ecologies in living Carnivora.
Methods
For this study we compared both linear and geometric morphometric measurements of morphology to dietary data to assess the ability of dental morphology to predict diet. We collected morphological data from images of museum specimens, and we obtained dietary data from the primary literature. We used both phylogenetic and non-phylogenetic statistics to make these predictions in order to assess the role evolutionary constraint plays in modulating the relationship between diet and the morphology of living species.
Dietary Classification
Following Van Valkenburgh (Reference Van Valkenburgh1988), we classified carnivoran diet into three categories based on the amount of vertebrate tissue in the diet. Hypercarnivores have greater than 70% vertebrate tissue in their diet, mesocarnivores have between 50% and 70%, and hypocarnivores have less than 50%. When quantitative data on the amounts of different food types consumed by a species from stomach content or scat analysis were available, we used this information to classify the species. However, for several species, quantitative values were unavailable. In these cases, we inferred the relative amounts of food items consumed from qualitative descriptions in the literature. If the qualitative description pointed to a food item being dominant in the diet, using terms like “mainly” or “mostly,” we considered that food item to be greater than 50% of the diet. If the description considered something commonly eaten but not a primary food item, we classified that food item as 30%–50% of the diet. If the literature pointed to food items being “rare” or “occasionally consumed,” we considered that food item to be less than 30% of the diet. If nonmeat items collectively composed more than 30% of the diet, the species was considered a mesocarnivore, or if more than 50% of the diet was nonmeat, a hypocarnivore. If no qualitative description was present, the relative amounts consumed of the food items listed were considered equal.
To investigate the role that our dietary classification might play in determining our results, we also explored a second classification (following Pineda-Munoz et al. [Reference Pineda-Munoz, Lazagabaster, Alroy and Evans2016b] in considering omnivory in more nuanced ways) that maintained the existing mesocarnivore category but broke hypocarnivores into omnivores and insectivores, and hypercarnivores into meat and bone specialists. Unfortunately, this scheme caused methodological issues resulting from very few species being assigned to the hypocarnivorous insectivore and hypercarnivorous bone specialist categories. As a consequence, we did not use this classification in our discriminant analyses, as this dataset would violate the assumptions of the method.
Dental Linear and Areal Measurements
Standardized photographs of the occlusal and lateral view of the upper and lower dentition were taken for carnivoran specimens at the University of California Berkeley Museum of Zoology. A total of 232 adult specimens were used from 125 extant terrestrial carnivoran species; a list of included specimens can be found in the Supplementary Material. Specimens with visible pathologies, substantial tooth wear, or any other visible dental abnormalities were not photographed. For each species, we used a single individual to represent the species. While we did have multiple individuals available, simply including them in the analysis as separate data points creates a problem with pseudoreplication; two members of the same species are not just as independent of one another as are two specimens drawn from different species. It is also quite difficult to meaningfully include multiple individuals of the same species in a phylogenetic comparative framework, as it is impossible to know how deep the divergence is between those individuals unless you have molecular sequence data or pedigrees for the actual individuals included in the analysis. It is possible that an atypical individual could confound our analysis, which is why we were careful to select specimens that had no obvious pathologies or abnormalities for inclusion in the analysis. This is also why we are interpreting the aggregate patterns of the analysis, rather than the species-by-species differences. Using a sample of fewer than five specimens to represent a species in a phylogenetically broad study like this one has been shown by past morphometric studies to be methodologically valid and statistically unproblematic for the inferences we draw here (Marcus et al. Reference Marcus, Hingst-Zaher and Zaher2000; Morgan Reference Morgan2008; Prevosti et al. Reference Prevosti, Turazzini, Ercoli and Hingst-Zaher2012; Tarquini et al. Reference Tarquini, Chemisquy and Prevosti2020).
On each image, we measured 18 dimensions commonly used to infer the degree of carnivory, listed in Table 1 and illustrated in Figure 1. These measurements were drawn from Van Valkenburgh (Reference Van Valkenburgh1988) and Popowics (Reference Popowics2003). Linear and areal measurements were taken in ImageJ 1.50b (Schneider et al. Reference Schneider, Rasband and Eliceiri2012), scaled using a millimeter ruler included in each photograph. All measurements were measured from each view in which they could be seen. All but canine length (CL) were measured in occlusal view, unless a dimension could not be reliably measured in this view for a specific specimen, in which case the measurement was taken from the lateral view. All species measurements are from a single individual; where we had measurements from multiple individuals, we selected the most completely measured individual for analysis. Where more than one specimen was equally complete, we selected an exemplar specimen randomly.
Dental characters and landmarks measured as proxies for diet in carnivorans on a skull of Lutra canadensis, MVZ 128092 (Museum of Vertebrate Zoology, Berkeley, Calif.). 1, canine length (CL); 2, canine width (CW); 3, upper P4 length (P4LB); 4, upper P4 width (P4W); 5, upper P4 shear length, paracone to metacone (P4PM); 6, upper P4 sulcus length (PRBL); 7, upper M1 buccal length (M1BL); 8, upper M1 lingual length (M1LL); 9, upper M1 buccolingual width (M1W); 10, total grinding area of lower molars (TGA); 11, largest lower premolar width (PMW); 12, largest lower premolar length (PML); 13, blade length of lower carnassial (BL); 14, lower m1 length (M1L); 15, postcarnassial molar length (PCML); 16, lower m1 width (N1W); 17, lower m1 shear length (N1PP); 18, lower m1 shear width (N1MP). Ps, parastyle; Pr, protocone; Pa, paracone; Mt, metacone; Pad, paraconid; Mtd, metaconid; Ptd, protoconid; Hld, hypoconulid.
Linear morphometric diet proxy measurements, their abbreviations, and their assignments to character sets.
We grouped the linear and areal measurements into multivariate clusters of associated measurements (see Table 1) to determine the ability of different sets of measurements to discriminate dietary categories. The clusters were: all characters, upper dentition, lower dentition, lower carnassial, upper carnassial, upper postcanine dentition, and all postcanine (which includes upper and lower tooth dimensions).
Geometric Morphometric Quantification of Carnassial Shape
The same photographs of specimens were used for geometric morphometric analyses as were used for linear and areal measurements. We quantified shape by placing landmarks on four homologous points of the occlusal view of the upper and lower carnassial teeth of each specimen using the software program tpsDig2 (Rohlf Reference Rohlf2009). We placed landmarks on the lower carnassial in the order of protoconid, paraconid, metaconid, and hypoconulid and on the upper carnassial in the order of protocone, paracone, metacone, and parastyle. We measured the scale of each photograph by calibrating against a ruler that had been photographed along with each specimen. We used lateral photographs of the carnassial as references to see the location of the tip of each cusp and provide better accuracy when placing the landmark on the occlusal view of the tooth. Lateral views were not used in the analysis, as wear can significantly alter the shape of the tooth, adding error to any shape analysis. When placing landmarks on carnassials with cusps that had been worn down significantly, we placed the landmarks in the middle of the area of wear. Landmarks were aligned using generalized Procrustes analysis, and the mean for each species with more than one specimen was calculated in the geomorph package in R (Adams and Otarola-Castillo Reference Adams and Otarola-Castillo2013).
Predicting Diet from Tooth Morphology
We used flexible discriminant functional analysis (fDFA) to quantify how well tooth morphology (either using linear/areal measurements or geometric morphometrics) predicts diet across carnivorans. Non-phylogenetic analyses were implemented in the R package mda v. 0.4-4 (Hastie and Tibshirani Reference Hastie and Tibshirani2013), and phylogenetic analyses (pfDFA) were implemented using the R code provided by Motani and Schmitz (Reference Motani and Schmitz2011). The prior proportions for each diet were set to equal the proportion of species in each diet in the dataset. For the pfDFA, we used the time-calibrated carnivoran phylogeny of Nyakatura and Bininda-Emonds (Reference Nyakatura and Bininda-Emonds2012) with polytomies randomly resolved using a constant rates birth–death model (Kuhn et al. Reference Kuhn, Mooers and Thomas2011). We sampled 1000 randomly resolved phylogenies, which were used in all subsequent phylogenetic analyses. Motani and Schmitz (Reference Motani and Schmitz2011) recommend incorporating the estimate of Pagel's λ (Pagel Reference Pagel1997), which maximizes the relationship between the ecological trait and the morphological measurements into the pfDFA. λ is a measure of phylogenetic signal; when λ = 1, the phenotypic covariance among species is perfectly described by the branch lengths of the phylogeny, and when λ = 0, there is no phylogenetic patterning. If both morphology and ecology have a strong phylogenetic signal, the estimate of λ that maximizes the relationship between them will be 0, so a non-phylogenetic fDFA will be run. Therefore, to investigate the influence of phylogeny upon the ability of tooth morphology to estimate diet we fixed λ = 1.
Phylogenetic Signal
For each of the linear and areal measurements as well as each set of characters defined for the analysis, we assessed the amount of phylogenetic signal in the character using Blomberg's K (Blomberg et al. Reference Blomberg, Garland and Ives2003) in the R package phytools (Revell Reference Revell2012). Phylogenetic signal within the geometric morphometric landmark data was estimated using a multivariate version of Blomberg's K (Adams Reference Adams2014) implemented in the R package geomorph (Adams and Otarola-Castillo Reference Adams and Otarola-Castillo2013). K is calculated as the mean square error among species over the mean square error estimated when species are randomly permuted across tips; this ratio is then scaled by the Brownian motion expectation, so that under Brownian motion, K = 1. A K < 1 indicates that variance is partitioned within clades and thus has a lower phylogenetic signal than expected under Brownian motion, while K >1 indicates variance is partitioned between clades and thus has a stronger phylogenetic signal than expected under Brownian motion. Analyses were run for each character or character set individually over the 1000 randomly resolved trees and the results of the runs summarized graphically.
Mass Dependence
We used actual measurements of linear and areal morphometrics in our fDFAs, rather than normalizing the morphological parameters to size. Size normalization is commonly performed by dividing the individual measures by some index of overall size, such as first molar length. However, because so much of the dentition of carnivoran mammals is evolutionarily shaped by adaptation to different diets, such a size index is difficult to choose without turning the index into a complex reflection of the evolutionary changes in two different characters, the measurement of interest and the size proxy. In addition, ratio data are usually nonnormally distributed (Geary Reference Geary1930) and hence violate the assumptions of parametric statistics (Zar Reference Zar2010), so we have chosen not to size-correct our morphological parameters by generating simple indices. We did test for the confounding influence of size, as explained below.
The geometric morphometric analyses were rescaled during the generalized Procrustes analysis, thus removing absolute size. However, because analyses of linear and areal proxies were run using the absolute magnitudes of measurements, it is possible that fDFA results could partially reflect the role of overall size in determining diet. Certainly, mass plays a role in shaping the diets of mammals (Carbone et al. Reference Carbone, Mace, Roberts and Macdonald1999; Price and Hopkins Reference Price and Hopkins2015; Pineda-Munoz et al. Reference Pineda-Munoz, Evans and Alroy2016a), so we must consider the possibility that our multivariate DFAs would partition species on the basis of overall size alone, and not the particulars of ecomorphology. Similarly, our assessment of phylogenetic signal could be detecting the phylogenetic signal in size (known to be substantial; Alroy Reference Alroy1998; Ashton Reference Ashton2004; Kamilar and Cooper Reference Kamilar and Cooper2013). To determine whether or not our results reflected overall size rather than ecomorphology, we ran a regression of each proxy measure against species average mass and subjected the residuals from this regression to the same DFAs and phylogenetic signal analyses as the uncorrected data. Any patterns that hold in both analyses are expected to be reliable, and if there are substantial differences in results, those could be explained by the confounding effects of overall mass evolution.
Results
Predicting Diet from Tooth Morphology
DFA of dental morphology as a predictor of diet finds relatively high rates of misclassification, even though almost all the DFAs are significantly better than expected by chance. Results are relatively consistent across different character sets and between the different methods compared here. The most significant and consistent effect explored was that of phylogeny. The non-phylogenetic DFA shows that dental morphology discriminates diet successfully most of the time, but with a significant degree of error. In contrast, the phylogenetic DFA shows much higher rates of error in predicting diet, roughly twice that of the methods that do not account for evolutionary relatedness.
Non-phylogenetic DFAs generally classify diet incorrectly for 20%–30% of included species (see Table 2), from a minimum of 12.5% for the combined set of all linear and areal proxy measures to a maximum of 37% for the linear measures of the upper carnassial. All measures tend to underpredict the degree of carnivory about two to four times as often as they overpredict it. Most of these misclassification errors (between 78% and 100%) are off by only one category, but a notable proportion (from 2% to 6%) of almost every analysis includes prediction of hypocarnivores as hypercarnivores or vice versa. The only analysis without such errors is the combined set of all linear and areal measures. Among the smaller sets of dental characters, the set of data from the upper dentition (22.7% error rate) slightly outperforms the lower dentition (25.8% error rate), but the five lower carnassial measurements (27.3% error) give a more accurate prediction than the four upper carnassial measures (36.9% error), a finding that does not seem to be simply a result of the number of measurements, as the geometric morphometric (GM) data on carnassials using equivalent numbers of landmarks shows the same pattern.
Confusion matrices for discriminant function analyses (DFAs). Reported values from phylogenetic DFAs using proportional prior; results are similar for equal prior. Error rate for phylogenetic DFAs is reported as median and interquartile range (in parentheses) over the 1000 randomly resolved trees. UPR, underprediction rate, or the rate at which the degree of carnivory is underestimated; OPR, overprediction rate, or the rate at which the degree of carnivory is overestimated; H/HCR, hypocarnivore/hypercarnivore confusion rate, or the frequency with which hypocarnivores are predicted to be hypercarnivores, or vice versa; GM, geometric morphometric.
Phylogenetic DFAs are much less accurate in predicting diet. Analyses using linear and areal dimensions yield phylogenetic misclassification rates between 36% and 55%, while changes in the primary cusp placement as quantified by the geometric morphometric data result in 46%–61% misclassification. While these rates are significantly better than random (given that diet has three categories in these analyses), the classification is wrong more often than it is right in the upper carnassial geometric morphometric dataset, while it is correct just barely more than half the time based on the linear dimensions of the upper carnassial (Table 2). Phylogenetic DFAs also overpredict carnivory approximately twice as often as they underpredict it, a switch relative to non-phylogenetic analyses, and the dramatic misclassifications (hypocarnivores for hypercarnivores and vice versa) are substantially more frequent, between 12% and 26% of cases.
While it is difficult to plot multivariate data in a way that simply illustrates the source of the classification errors, Figure 2 offers some insight into the reasons for these classification errors. When morphology (in this case a single morphological proxy, relative grinding area, equal to the square root of lower molar grinding area [TGA] divided by blade length [BL]) is plotted on the phylogeny and compared with dietary category, it is evident that the morphology does tend to align reasonably well with diet, but that the morphology in an absolute sense is not the same between different groups of carnivorans with the same diet and that the changes in diet between closely related taxa are not consistently associated with morphological changes in the dentition. At the same time, over the broad scale, the patterns of association between morphology and diet are consistent, as expected, given that the majority of classifications are correct, even with limited data. On the fine scale, changes in diet and morphology also show the mismatch seen in phylogenetic DFAs, where increases in the degree of carnivory are sometimes associated with the expected morphological change (e.g., Gulo gulo and Martes pennanti within the Mustelidae) and sometimes are not (e.g., Vulpes macrotis within the Canidae).
Relationship between dental morphology and dietary category among the studied carnivoran taxa. Phylogeny shown here is one topology randomly sampled from the 1000 randomly resolved trees. Colored branches represent maximum-likelihood estimate of relative grinding area assuming Brownian motion. Pie charts on nodes represent posterior probability of diet estimated from 500 stochastic character maps using the best-fitting Q matrix allowing all transition rates to vary (model = “ARD”) and empirical estimates at the root. Analyses and visualizations were implemented using the R package phytools (Revell Reference Revell2012). Note that we have shown hypocarnivorous insectivores and omnivores with subtly different colors to show where these categories differ, even though there were not enough species sampled to allow splitting the categories in this way for analysis.
Phylogenetic Signal
The results of DFAs make it clear that the role of phylogeny in determining the relationship between morphology and diet is significant, which begs the question of how strong the phylogenetic signal is in dental morphology. Our analysis of Blomberg's K for the linear measures shows that phylogenetic signal is significant for all characters, but relatively low (Table 3). Median K values are between 0.2 and 0.35, leaving plenty of non-phylogenetic variation that could correlate with diet if these characters were strongly constrained by diet.
Phylogenetic signal measured by Blomberg's K for individual characters. Mean (K mean) and standard deviation (SD K) over 1000 randomly resolved trees. The last column reports the difference between mass-corrected and uncorrected K values (K diff), showing that although mass correction reduces the phylogenetic signal in most characters, many characters retain substantial degrees of phylogenetic signal. See Table 1 for diet proxy measurements abbreviations.
When looking at phylogenetic signal in the analyzed sets of characters (Table 4), it is evident that this variation in signal between characters averages out, and the sets of linear and areal measures yield K values between 0.26 and 0.38, other than the character set intentionally chosen to have low phylogenetic signal. Geometric morphometric approaches show notably higher levels of phylogenetic signal, with K values between 0.4 and 0.43. The complete dataset of linear and areal measures has a relatively high phylogenetic signal value of 0.38, which is unsurprising, as including this many measures in a dataset begins to approach the level of geometric integration captured by a geometric morphometric analysis.
Phylogenetic signal in character sets, measured by Blomberg's K value, with median value and confidence limits over 1000 trees. This value varies between 0 and 1, with high values indicating larger amounts of phylogenetic structure in the data. We also report the number of insignificant p-values among the 1000 trees for significance test of the phylogenetic signal on each tree.
Mass Dependence
Mass is a potential confounding factor in this analysis, as it clearly carries a strong phylogenetic signal (K mean = 0.19; Table 3). When running tests of Blomberg's K on univariate measures, the phylogenetic signal in overall body size is indubitably part of the signal this test detects. Our test of mass-corrected dental dimensions found that the proportion of the phylogenetic signal in the characters accounted for by size varies widely by character (Table 3). Most K values drop by about 20%–60%. Notable exceptions include TGA and postcarnassial molar length (PCML), which actually increase substantially with mass correction, and shear lengths of the upper P4 (P4PM) and lower M1 (N1PP), both of which see only miniscule drops in mass-corrected K values. All the mass-uncorrected character sets have significant levels of phylogenetic signal over 99% or more of the trees analyzed. The mass-corrected character sets again see substantial drops in the levels of phylogenetic signal, and for both character sets based on the lower dentition alone (all low and low carn) and the well-represented characters (well rep) less than half of the trees yielded significant levels of phylogenetic signal.
While phylogenetic signal is relatively low whether or not we correct our data for mass, we were concerned that mass might also affect the success of DFAs. Given that mass varies significantly with diet (Price and Hopkins Reference Price and Hopkins2015), size is likely to be part of the success of our dietary discrimination. Because past analyses have taken measures to remove size from diet proxies (Van Valkenburgh Reference Van Valkenburgh1988; Popowics Reference Popowics2003; Samuels Reference Samuels2009), we also need to account for the influence of size in our predictions from linear and areal measures of the dentition. Our mass-corrected DFAs performed similarly to the uncorrected analyses; error rates varied between 15% and 33% for nonphylogenetic DFAs and between 38% and 56% for phylogenetic DFAs (Table 5). All of the nonphylogenetic DFAs of mass-corrected measures were within approximately 10% of the error rates their non–mass corrected counterparts, evenly split between higher (all up, up carn, up postcanine, and well rep) and lower (all low, low carn, and all) error rates. Patterns of under- and overprediction were congruent in the mass-corrected analyses to their uncorrected equivalents.
Confusion matrices for mass-corrected discriminant function analyses (DFAs). Reported values from phylogenetic DFA using proportional prior; results are similar for equal prior. Error rate for phylogenetic DFAs is reported as median and interquartile range (in parentheses) over the 1000 randomly resolved trees. UPR, underprediction rate, or the rate at which the degree of carnivory is underestimated; OPR, overprediction rate, or the rate at which the degree of carnivory is overestimated; H/HCR, hypocarnivore/hypercarnivore confusion rate, or the frequency with which hypocarnivores are predicted to be hypercarnivores, or vice versa.
Discussion
The biases in the mammalian fossil record, which preserves mostly durable remains such as teeth, make it highly desirable to estimate paleoecology of extinct species from dental morphology. The results of our studies offer a warning about some of the limitations of these approaches. The linear and geometric dental proxies explored here predict dietary ecology, but the error rate of discriminant functions is substantial, classifying 20%–30% of the species incorrectly in most analyses. Even the best of these analyses misclassifies the diets of species 12.5% of the time, and this rate is achieved using a set of dental measures not obtainable for the vast majority of fossil mammals. Because DFAs that account for evolutionary history perform even more poorly than do those without phylogenetic information, our results suggest that diet and tooth morphology are linked via phylogeny. This phylogenetic signal combined with the observation that diet is evolutionarily labile (Feranec et al. Reference Feranec, Bryant, Froelich, Showers, Genna, Cerling, Harris, Leakey, Deniro, Epstein, Dompierre, Churcher, Ehleringer, Monson, Sage, Flanagan, Pearcy, Helliker, MacFadden, Fraser, Friedli, Lötscher, Oeschger, Siegenthaler, Stauffer, Heckathorn, McNaughton, Coleman, Hofmann, Stewart, Hutchinson, Janis, Ehrhardt, Janis, Damuth, Theodor, Koch, Tuross, Fogel, Hoppe, Webb, Koford, Macfadden, MacFadden, Marino, McElroy, Salawitch, Spaulding, McNaughton, Tarrants, McNaughton, Davis, Migongo-Bake, Hansen, Puig, Videla, Monge, Roig, Cona, Quade, Cerlinga, Barry, Morgan, Pilbeam, Chivas, Lee-Thorp, van der Merwe, Wedin, Li, Schoeninger, DeNiro, Solounias, Moelleken, Plavcan, Stowe, Teeri, Strömberg, Teeri, Stowe, Livingstone and Wang2003; Price et al. Reference Price, Hopkins, Smith and Roth2012) suggests that these misclassifications may be a result of species that have rapidly changed their diets but have seen only minor changes in dental morphology, particularly if the ancestral morphology was very characteristic of the previous diet. Such evolutionary lags in the response of dental morphology to dietary change may yield a mismatch between diet and tooth shape and weak evolutionary correlations between the two on the fine scale in spite of the large-scale phylogenetic signal in dental morphology.
It is worth noting that dental morphology performs quite well at identifying diet on the broadest scale; only very few species are dramatically misclassified (i.e., a hypercarnivore classified as a hypocarnivore or vice versa), and hypercarnivores are quite well differentiated from hypocarnivores. In distinguishing dramatically different diets, dental morphology is clearly more than adequate. However, the greatest difficulties arise in differentiating mesocarnivores from the other two groups, suggesting that the resolution of morphological diet proxies may be the limiting factor, rather than the actual relationship of morphology to diet. The error rate of this analysis suggests that inference along the dietary spectrum may be challenging using dental morphology. Hence, it may be difficult to use tooth morphology alone to understand the histories of understudied groups such as mesocarnivores and omnivores.
Morphological diet proxies do not seem to perform differently among the three dietary categories, indicating that their accuracy is fairly uniform across the dietary spectrum. This is true in spite of the fact that hypercarnivores in particular experience a biased morphological evolution toward greater meat specialization (Holliday Reference Holliday, Goswami and Friscia2010) and greater morphological variance than less-carnivorous taxa (Holliday and Steppan Reference Holliday and Steppan2004). The absence of greater error in diet prediction in hypercarnivores supports the suggestion (Holliday Reference Holliday, Goswami and Friscia2010) that an evolutionary ratchet is driving greater morphological specialization among these taxa, rather than a simple developmental constraint. Furthermore, the fact that morphological misclassification of hypercarnivorous taxa seems to happen in cases (see Fig. 2) where hypercarnivory has appeared relatively recently is consistent with our suggestion that it is the mismatch between the evolutionary rates of diet and of morphology that leads to classification errors.
Our results do offer some hope that, within the limitations of dietary resolution obtainable here, the data available from fossil carnivorans offer a reasonable estimate of diet. While dietary categorization is most accurate with the greatest amount of information, a fairly accurate classification (18% error in the non-phylogenetic DFA) can be achieved using geometric morphometric analyses of the upper and lower carnassials alone, and isolated upper or lower dentition can predict dietary categories with only about a 25% error rate (from our non-phylogenetic DFA), and less than 5% chance of a really dramatic misclassification. This indicates that, while individual taxa may be misclassified, on the broad scale, these measures offer a good correlation with dietary patterns over a majority of species in the sample.
Past analyses have developed carnivory indices corrected for body size (Van Valkenburgh Reference Van Valkenburgh1988; Popowics Reference Popowics2003; Van Valkenburgh et al. Reference Van Valkenburgh, Wang and Damuth2004; Friscia et al. Reference Friscia, Van Valkenburgh and Biknevicius2007). Given that body size does vary in a nonrandom fashion with diet (Price and Hopkins Reference Price and Hopkins2015; Pineda-Munoz et al. Reference Pineda-Munoz, Evans and Alroy2016a), these efforts might be expected to actually decrease the capacity of dental measurements to predict diet. Meloro and Raia (Reference Meloro and Raia2010) found that, while carnassial size has a relatively small component of phylogenetic signal, carnassial angle, a size-independent measure of carnassial shape roughly equivalent to our mass-corrected regression of lower carnassial dimensions, has a strong phylogenetic signal. Meloro and Raia's finding might suggest that size-corrected data should have greater phylogenetic signal, while other studies (Van Valkenburgh Reference Van Valkenburgh1999; Van Valkenburgh et al. Reference Van Valkenburgh, Wang and Damuth2004) suggest we should find strong dietary signal in the mass-corrected data. In fact, we find that mass-corrected data are not consistently better or worse than uncorrected data at predicting diet and actually have less phylogenetic signal than uncorrected values. Certainly, our findings cannot be explained by the confounding influence of body-size evolution, because prediction accuracy is not substantially better for mass-corrected data. This does not mean that individual linear measures should not be size-corrected when used to predict diet, but rather that the multivariate tools available to us are capable of accounting for the covariance of measures as a result of overall size in detecting the influence of diet on morphological data.
Dietary categorization schemes can make a significant difference to our ability to detect macroevolutionary patterns in ecological data (Pineda-Munoz et al. Reference Pineda-Munoz, Lazagabaster, Alroy and Evans2016b), a problem that is of real importance, given the relative rarity of robust, quantitative dietary data for even the best-studied clades of mammals (Price et al. Reference Price, Hopkins, Smith and Roth2012). It is fair to ask whether a change in classification scheme would help with the ability of dental morphology to discriminate diet; unfortunately, we were unable to get a large enough sample of some diets (bone specialists and insectivores) to run a robust DFA, so we cannot determine whether subdividing these categories would help with discrimination; however, because they are such a small part of our sample, they clearly cannot be entirely responsible for the error in classification. Furthermore, our finding that the categories most commonly misclassified are those that differ less dramatically suggests that a finer categorization is more likely to lead to more errors, not fewer. In the end, while dental morphology does vary among carnivorans with differing diets, using that information to place taxa in a dietary category has a certain unavoidable degree of error, generally between 15% and 30% for most methods, and those errors are not always a result of the arbitrary boundaries imposed by any categorization scheme. Some of them are real and substantial; a significant number of taxa will simply be incorrectly assigned by these approaches.
More troubling still is the poor performance of phylogenetically informed DFAs. These results indicate one of two possibilities: (1) the direction of morphological evolution does not seem to consistently track the direction of dietary evolution; or (2) both diet and morphology of teeth are tightly structured by phylogeny over the large scale with little variation in both on the finer scales, such that removing evolutionary information erases the real relationship between the two. It would be difficult to distinguish these two possibilities with the dataset presented here, although the work of Meloro and Raia (Reference Meloro and Raia2010) suggests that the second explanation is possible, given their finding of little variation in carnassial morphology within carnivoran families. However, studies of diet (Feranec et al. Reference Feranec, Bryant, Froelich, Showers, Genna, Cerling, Harris, Leakey, Deniro, Epstein, Dompierre, Churcher, Ehleringer, Monson, Sage, Flanagan, Pearcy, Helliker, MacFadden, Fraser, Friedli, Lötscher, Oeschger, Siegenthaler, Stauffer, Heckathorn, McNaughton, Coleman, Hofmann, Stewart, Hutchinson, Janis, Ehrhardt, Janis, Damuth, Theodor, Koch, Tuross, Fogel, Hoppe, Webb, Koford, Macfadden, MacFadden, Marino, McElroy, Salawitch, Spaulding, McNaughton, Tarrants, McNaughton, Davis, Migongo-Bake, Hansen, Puig, Videla, Monge, Roig, Cona, Quade, Cerlinga, Barry, Morgan, Pilbeam, Chivas, Lee-Thorp, van der Merwe, Wedin, Li, Schoeninger, DeNiro, Solounias, Moelleken, Plavcan, Stowe, Teeri, Strömberg, Teeri, Stowe, Livingstone and Wang2003; Dunlop et al. Reference Dunlop, Rayner and Doherty2017; Spencer et al. Reference Spencer, Newsome and Dickman2017) have frequently suggested that diet is quite labile even within species, as well as within clades, particularly among the Carnivora, suggesting that the tight structuring of diet required by this second possibility may be unlikely. Both possibilities have similar implications for the practice of inferring the diets of individual species from dental morphology. On the fine scale of species by species, differences in dental morphology are not reliably predictive of diet with more than 70%–80% certainty. Furthermore, much of that prediction depends on the phylogenetic distribution of morphology and of diets observed in living species, and it is difficult to know whether that distribution is an artifact of some confounding factor or a real and predictive relationship. Without knowing for sure what supports the correlation between diet and tooth shape, some caution should be exercised in using dental morphology to predict diets in fossil carnivorous mammals, especially those evolutionarily distant from most living species on which the predictions are based.
If our explanation is correct, it is the evolutionary lag between dietary change and dental morphological evolution that causes the unexpected loss of accuracy in the phylogenetic DFAs relative to the non-phylogenetic analyses. This hypothesis would be testable by the inclusion of extensive fossil data but would require a robust and broadly applicable non-morphological diet proxy for fossil carnivores, which is sadly lacking. Our finding that controlling for phylogenetic information impairs the accuracy of predictions is troubling, because it indicates that dietary change is not consistently predicted by the direction and magnitude of dental evolution in carnivorans. Hence, predictions of dietary differences based on the morphological differences between closely related species must be regarded with the same uncertainty as those inferred for distantly related groups. In the absence of a complete knowledge of the evolutionary trajectory, it is hard to be sure whether the morphological difference between any two related species reflects a simple, directional change in diet or a more complex path that may make dental morphology misleading. Importantly, the substantial phylogenetic signal in dental morphology also means that the hypotheses of diet for taxa well outside the phylogenetic bracket of taxa used in constructing a proxy, as, for example, the Creodonta would be for a proxy developed using extant Carnivora, are less accurate than they would be if obtained from a part of the evolutionary tree where the dental morphology and diet are better known. While dental morphology can be a guide to diet, interpretations about fossil taxa with no close living relatives must be made with acknowledgment of the uncertainties introduced by long evolutionary distances.
The reasons for the comparative inertia in dental morphology suggested here are likely complex. As with the masticatory apparatus of ruminants (Fraser and Rybczynski Reference Fraser and Rybczynski2014), there is an interaction among genetic constraints on development, the functional constraints on the dentition based on position within the jaw, and temporal variation in the direction and magnitude of the selective forces diet imposes on dental morphology as diets vary with environmental change. In general, we expect that dental morphology, which is largely shaped by genetic factors early in development, will change more slowly than diet, which is determined in part by behavior and environment and can vary between individuals within a population or over the lifetime of a single organism.
We suggest that the most likely explanation for our results is that diet is relatively labile among carnivorans and that lags in the evolution of morphology may yield a relationship to diet that is not entirely consistent and predictable. While dental morphology offers a good prediction of diet over the broad range of carnivorans, dental adaptations to particular diets vary. Lineages can accomplish functional adaptations in anatomically distinct ways when presented with the same selective environment (many-to-one mapping of form to function; Wainwright et al. Reference Wainwright, Alfaro, Bolnick and Hulsey2005) and conversely can exploit different resources with the same morphological structures. This answer is disappointing, because it implies that dietary reconstructions from the dental morphology of carnivorans contain a certain unavoidable level of error. However, it is perhaps not surprising, as similar results have been shown for dental morphology of ungulates (Feranec et al. Reference Feranec, Bryant, Froelich, Showers, Genna, Cerling, Harris, Leakey, Deniro, Epstein, Dompierre, Churcher, Ehleringer, Monson, Sage, Flanagan, Pearcy, Helliker, MacFadden, Fraser, Friedli, Lötscher, Oeschger, Siegenthaler, Stauffer, Heckathorn, McNaughton, Coleman, Hofmann, Stewart, Hutchinson, Janis, Ehrhardt, Janis, Damuth, Theodor, Koch, Tuross, Fogel, Hoppe, Webb, Koford, Macfadden, MacFadden, Marino, McElroy, Salawitch, Spaulding, McNaughton, Tarrants, McNaughton, Davis, Migongo-Bake, Hansen, Puig, Videla, Monge, Roig, Cona, Quade, Cerlinga, Barry, Morgan, Pilbeam, Chivas, Lee-Thorp, van der Merwe, Wedin, Li, Schoeninger, DeNiro, Solounias, Moelleken, Plavcan, Stowe, Teeri, Strömberg, Teeri, Stowe, Livingstone and Wang2003; Damuth and Janis Reference Damuth and Janis2011; Kaiser et al. Reference Kaiser, Mueller, Fortelius, Schulz, Codron and Clauss2013), as well as other supposedly evolutionarily independent diet proxies (Mihlbachler et al. Reference Mihlbachler, Campbell, Ayoub, Chen and Ghani2016; Fraser et al. Reference Fraser, Haupt and Barr2018), although such findings have created some controversy (DeSantis et al. Reference DeSantis, Fortelius, Grine, Janis, Kaiser, Merceron, Purnell, Schulz-Kornas, Saarinen, Teaford, Ungar and Žliobaitė2018; Clauss Reference Clauss2019). In the end, while dental morphology in mammals (and other taxa; Das and Coe Reference Das and Coe1994; Foffa et al. Reference Foffa, Young, Stubbs, Dexter and Brusatte2018) certainly responds evolutionarily to diet, ecology is invariably more labile than morphology, and that flexibility makes these predictions less than perfect. While we can gain some idea of diet from dental morphology, we must be aware of the potential sources of error in these approaches and, until more reliable proxies emerge, interpret them cautiously, especially when applied to more distantly related lineages such as stem carnivorans or creodonts. Other available diet proxies do not perform significantly better; while stable isotopic methods are quite effective at predicting diets in herbivores (Macfadden and Cerling Reference Macfadden and Cerling1994; Koch Reference Koch1998), there is not an equivalent method of discriminating the diets of carnivores and omnivores except where preserved collagen allows use of nitrogen isotope proxies (Koch Reference Koch1998; Hedges and Reynard Reference Hedges and Reynard2007). Such preservation rarely occurs in fossils older than late Pleistocene. Dental microwear has been applied to carnivores but has a rate of error similar to or greater than that reconstructed from dental morphology (Goillot et al. Reference Goillot, Blondel and Peigne2009). It therefore seems that our current estimates of the diets of extinct carnivorans are doomed to a certain level of inaccuracy, a problem that appears to result from the mismatch between the tempo and mode of ecological and morphological evolution.
Acknowledgments
The authors would like to thank C. Conroy at the Museum of Vertebrate Zoology for collections access. L. Schmitz kindly provided code for data analysis, as well as some helpful insights on phylogenetic discriminant function analysis. E. Davis provided suggestions on data analysis. This work was funded by National Science Foundation grant DEB-1256897 to S.S.B.H. and S.A.P.
Data Availability Statement
Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.q573n5tjq. (1) Raw measurement data on which the analyses are based. (2) Diet data used in the analyses.
Appendix. Confusion matrices for discriminant function analyses (DFAs). Columns in bold are correct identifications. Values in each cell represent the number classified by the DFA into that category. In the phylogenetic DFAs, the number in parentheses is the interquartile range of that value over the 1000 trees.
