A hyaena on stilts: comparison of the limb morphology of Ictitherium ebu (Mammalia: Hyaenidae) from the Late Miocene of Lothagam, Turkana Basin, Kenya with extant Canidae and Hyaenidae

Specimen collection

The remains used for the study of Ictitherium ebu include the manus, radius, ulna, humerus and femur of a cast of specimen KNM-LT 23145 from the Late Miocene of Lothagam, Kenya, housed in the National Museums of Kenya (NMK) (Fig. 2). A tibia is present as well, but it is broken at the diaphysis, which limits its relevance to this study. The collections of Naturhistoriska Riksmuseet (NRM), Museum für Naturkunde (ZMB), Senckenberg Naturmuseum (SMF), Alexander Koenig Zoological Research Museum (ZFMK), Royal Museum for Central Africa (RMCA), Naturalis Biodiversity Center (RMNH) and La Specola (MZUF) were visited to collect photographs and measurements of 79 specimens of extant species for comparison with KNM-LT 23145 (Table 1). Adult specimens of both sexes were chosen, with a preference for wild-caught individuals.

Only three of the nine spotted hyaena specimens used in this study had completely fused epiphyses (ZMB MAM 7784, ZMB MAM 13295 and ZMB MAM 47515), which is normally the main indicator of an adult animal. However, according to Egeland, Egeland & Bunn (2008), sub-adulthood is characterised by unfused or partially fused epiphyses with solid bone surfaces. Sponginess only occurs near-epiphyses. Adults have epiphyses that are mostly or completely fused, with the entire surface of the bone being solid. Specimens ZMB MAM 14818, ZMB MAM 16575, and ZMB MAM 82415 have fused epiphyses with entirely solid bone surfaces; only thin grooves show the epiphyses to not be entirely fused. They could thus be interpreted as adult specimens. ZMB MAM 82413, ZMB MAM 82471 and ZMB MAM 82516 do not have fused epiphyses, but do not show much spongy bone around the epiphyses. These specimens can then be interpreted as older sub-adults. The striped hyaena ZMB MAM 82363 also had a humerus that was not fully fused, which would indicate a subadult if interpreted in the same way as the spotted hyaena specimens (Egeland, Egeland & Bunn, 2008). However, it does not have any measurements that are below the range of the other specimens. Only the radial diameter is smaller than the others, but three specimens have equally small radial diameters. Truss distances all fall within the range of other specimens.

MZUF 13354 (wolf) illustrates the difference between zoo animals and wild animals quite accurately. In the species dataset, MZUF 13354 has the most extreme value for every single index, with eight outliers. The short bones of this animal are the likely cause of these values.

Linear morphometrics

Specimens were measured according to the protocol of Samuels & Van Valkenburgh (2008) and Samuels, Meachen & Sakai (2013), with an added measurement for the midshaft mediolateral diameter of the radius (RD) (Fig. 3, Table 2). The measurement for FGT was carried out differently from the one in Samuels & Van Valkenburgh (2008) and Samuels, Meachen & Sakai (2013), where the height of the greater trochanter of the femur was measured vertically instead of diagonally.

The measurements were taken with vernier callipers for measurements up to 15 cm. Measuring tape was used for measurements above 15 cm in all museums except for the Museum für Naturkunde in Berlin, where the measurements were carried out with larger callipers. Measurements were recorded to the nearest 0.1 cm (Table S1).

Index calculations

The measurements were converted into indices (Samuels & Van Valkenburgh, 2008; Samuels, Meachen & Sakai, 2013) in Excel 16.0.15330.20260 (Table 3, Dataset S1). The manus proportions index was excluded from further analysis due to a lack of measurements. The radial robustness index, metacarpal/radial index, humeral/femoral index and metacarpal/humeral index were included so as to take into account the metacarpal III measurements and the relationship between the humerus and the femur.

For initial interpretation, preliminary boxplots of the indices plotted against species were created. These boxplots revealed some outliers in the measurements, among which some are measurement errors. These values were often far too extreme to be viewed as simple extremes in the data. For example, the humeral epicondylar breadth of ZMB MAM 82516 was measured to be half that of the other specimens, while the length is within the range of the other specimens. With the use of ImageJ 1.53n to approximate what measurement would have been obtained on-site, it was determined that 10 outliers needed to be removed from the dataset (Table S2).

Aside from these outliers, there were some missing values in the dataset. For these missing values, means were interpolated in Excel 16.0.15330.20260 for species with missing indices by using the mean of the same index for the other specimens of the same species. Boxplots were recreated for the final analysis, with all of these changes incorporated.

Image collection

The distal humerus of each specimen was photographed in anterior view. The camera was set to have an ISO of 200/250, an aperture of F8–F10, after which the shutter speed was adjusted for brightness. Photographing was carried out with flash. The scale bar was held in place using two alligator clips on 4-way swivels, at the height of the specimen. Some photographs appeared overexposed after data collection. These images were edited using GIMP 2.10.30 with lowering of highlights and lowering the point at which highlights turn to white. The specimens edited in this manner were SMF 97379, SMF 97380 and ZMB MAM 89495. Some photographs were removed from the analysis due to misalignment of the anterior view or deformation of the humerus, see Table S3.

Truss creation

Using TpsUtil64 1.81 a TPS file was created from the photographs acquired during data collection. The TPS file was imported into TpsDig264 2.32, labelled and scaled according to the scalebar in each image. Six landmarks were assigned to each specimen following Andersson (2004) (Fig. 4). The coordinates of the landmarks for each specimen were exported to Excel 16.0.15330.20260 using MorphoJ 1.07a (Table S4), after which the distances of the truss were calculated (Dataset S1). These truss distances combine into a schematic representation of the size and shape of the elbow joint.

Software and graphics

All tests were carried out in R version 4.2.2 (R Core Team, 2022). The code and complementary files can be found in Dataset S1. Boxplots were created of the 13 indices of the linear morphometrics. Colours for plots were selected using the webpage “Coloring for colorblindness” (Nichols, 2023). Plots were created using ggplot2, ggtext, ggpubr and tidyverse packages (Wickham, 2016; Wickham et al., 2019; Wilke & Wiernik, 2022; Kassambara, 2023). Silhouettes of the different extant species were acquired through PhyloPic (Keesey, 2023).

Permutational multivariate analysis of variance

Permutational multivariate analysis of variance (PERMANOVA) (Anderson, 2001) was used for analysis of the linear morphometrics. It tests if the centroids of a group of objects are the same. The test is a non-parametric alternative to multiple analysis of variance (MANOVA). It is used here because Shapiro–Wilks normality test results were found to be largely non-parametric for the linear and truss data (Table S5). The assumption for PERMANOVA is that the observations are exchangeable under the null hypothesis. Therefore, objects must be independent and have similar multivariate dispersion. A PERMANOVA can be affected by heterogeneous dispersions in an unbalanced design (Anderson & Walsh, 2013; Anderson et al., 2017). Therefore, groups of N = 1 and N = 2 were removed from the analysis and assumption testing (Brown hyaena, Coyote and Ictitherium ebu).

Assumption tests for similar multivariate dispersion were carried out using the vegan package (Oksanen et al., 2022). A multivariate analogue of Levene’s test for homogeneity of variances (betadisper) (Anderson, 2006) was used with Euclidean distances, after which an analysis of variance (ANOVA) (Fisher, 1921), as well as a permutation test of multivariate homogeneity of group dispersions (permutest) (Legendre, Oksanen & ter Braak, 2011) were used to validate the assumption of similar multivariate dispersion.

For species, the ANOVA yielded a p of 0.01, while the permutest yielded a p of 0.01, both of which are significant (α = 0.05). For family both results were significant as well, with the ANOVA and permutest yielding a p of 0.004 and 0.005 respectively. Thus, multivariate dispersion was not similar, which can cause the test to be too conservative when there is large dispersion in groups of large numbers of samples and too liberal when there is large dispersion in groups of small numbers of samples (Anderson & Walsh, 2013).

A two-way PERMANOVA with Euclidean distances and 999 permutations was carried out to compare the indices between family and species, using the function adonis2 from the vegan package (Oksanen et al., 2022). Species is nested in family. Therefore the performed PERMANOVA is nested as well. Post-hoc Holm-corrected pairwise PERMANOVAs were carried out to identify differences in multivariate means between pairs of species using the function pairwise.adonis from the wrapper function PairwiseAdonis (Holm, 1979); Martinez (Arbizu, 2020). The Holm method of post-hoc correction is a more powerful sequentially rejective Bonferroni correction.

Non-metric multidimensional scaling

Non-metric multidimensional scaling (NMDS) (Kruskal, 1964) with a maximum of 999 random restarts, two dimensions and Euclidean distances was carried out using the function metaMDS from the vegan package (Oksanen et al., 2022) for both the indices as well as the Truss distances. Its goal is to plot dissimilar objects far apart from each other and similar objects close together in ordination space (Legendre & Legendre, 1998). First, a distance matrix is constructed using Euclidean distances, as the data are non-ecological. A number of dimensions is chosen, in our case k = 2 for ease of interpretability. An initial configuration is chosen; in the case of metaMDS, this is done with metric scaling (Oksanen et al., 2022). This initial configuration is important, as the solution to the algorithm that is used depends partly on this configuration (Legendre & Legendre, 1998).

A matrix of fitted distances is calculated, then compared to the initial distances using monotone regression (which is non-metric) fitted by least-squares (Legendre & Legendre, 1998). Goodness of fit (stress) is used to evaluate the regression, which measures how far the new configuration is from being monotonic to the original distances. It is a relative measure, as it only measures the decrease in lack-of-fit between iterations in this procedure. The configuration is then moved slightly in the direction in which stress decreases the most rapidly (Kruskal, 1964; Legendre & Legendre, 1998). The matrix is then recalculated, and steps are repeated until a minimum lack-of-fit is reached and no more progress can be made or until a tolerated lack-of-fit is reached (Legendre & Legendre, 1998). These then become the coordinates of our two-dimensional ordination. In our case, the programme is allowed to carry out up to 999 random restarts before the process is halted (Oksanen et al., 2022). Data are then centred, as well as rotated so that the first principal component will be on the first axis. The variable scores of the NMDS were used to explain the ordination. The NMDS results were validated using a Shepard diagram and goodness of fit of individual points (Dexter, Rollwagen-Bollens & Bollens, 2018), using the stressplot and goodness functions of the vegan package (Oksanen et al., 2022). These are available in the supplementary material (Fig. S1, Fig. S2, Table S6).

Stepwise flexible discriminant analysis

Stepwise variable selection was carried out for species, family, locomotion and habitat on both datasets. Using a greedy Wilk’s lambda F-test (Mardia, Kent & Bibby, 1979) from the klaR package (Weihs et al., 2005) the indices were selected based on an F-test decision of 0.05 (Table S7). The variable selection works by defining a start variable that separates the group most, then selects additional variables (Weihs et al., 2005). It makes this selection based on the Wilk’s lambda criterion, which means it selects the one which minimises Wilk’s lambda of the model, adding more variables if the p-value still shows statistical significance (p = 0.05).

The data were divided into 30% training data and 70% validation data using the caret and tidyverse packages (Wickham et al., 2019; Kuhn, 2008). The training data were then used to carry out flexible discriminant analysis (FDA) (Hastie, Tibshirani & Buja, 1994) for each variable using the function fda from the mda package (Hastie et al., 2022). Ictitherium ebu, brown hyaena and coyote were removed before dividing the data. The brown hyaena and coyote have small sample sizes (N = 2 and N = 1 respectively) and were used to validate the predictions of the data. The values for I. ebu were predicted to assess its similarities to other taxa and their ecologies.

Flexible discriminant analysis works as an adaptation of linear discriminant analysis (LDA) (Hastie, Tibshirani & Buja, 1994). Linear discriminant analysis finds a reduced number of discriminate coordinate functions to be able to optimally separate groups. This number is always the number of groups minus one. The non-parametric multiresponse regression technique BRUTO (Hastie, 1989) was used for FDA. It generates a very large base set automatically, then achieves parsimony by shrinking coefficients in a sensible, structured manner (Hastie, Tibshirani & Buja, 1994). The function “bruto” in the mda package functions by fitting a model by adaptive backfitting using smoothing splines (Hastie et al., 2022). The number of adaptive models is equal to the number of response variables in the model, but for each variable the same amount of smoothing is used. The variable can either be omitted, linear, or fitted by a smoothing spline. During each step of the backfitting procedure, model selection is based on an approximation of the generalised cross-validation criterion (Hastie, Tibshirani & Buja, 1994; Hastie et al., 2022). Once selection has finished, the model is backfitted using the chosen amount of smoothing (Hastie et al., 2022). This analysis was preferred over regular LDA, as LDA assumes normal distributions in the data. Shapiro–Wilks normality test results were found to be largely non-parametric for the linear and truss data (Table S5).

This model is then validated by testing how accurately the model predicts the validation data using the “predict” function from base-R (R Core Team, 2022). The variable most similar to I. ebu, coyote and brown hyaena was predicted in each model using the same function.

Linear morphometrics

Non-metric multidimensional scaling

A convergent solution was found after 20 tries by the model. Stress for the non-metric multidimensional scaling is 0.13. The Shepards plot shows an R² of the non-metric fit of 0.98 (Fig. S1). In the goodness of fit table (Table S6), none of the values of goodness of fit are above 0.05.

Axis 1 of the NMDS is primarily controlled by SMI and HEI (Tables 3 and 5). SMI represents the muscles acting across the shoulder joint while HEI represents the relative area for the origins of the forearm flexors, pronators, and supinators. Canids plot more on the left side and hyaenids more on the right side Fig. 5.

Table 5:

NMDS1 and NMDS2 loadings of the non-metric multidimensional scaling of the morphometric indices.

	NMDS1	NMDS2
SMI	−0.16	0.08
BI	−0.01	−0.06
HRI	−0.05	0.03
OLI	0.03	0.17
URI	−0.02	0.17
FRI	−0.03	0.04
GI	−0.03	0.05
FEI	−0.02	0.06
HEI	−0.12	0.04
RRI	0.01	0.12
MCRI	0.06	−0.0004
HFI	0.08	−0.01
MCHUM	−0.01	−0.05

DOI: 10.7717/peerj.17405/table-5

Axis 2 is primarily controlled by OLI, URI and RRI (Tables 3 and 5). OLI relates to the muscles used in elbow extension, while RRI is the indicator for radial robustness and stress resistance. URI represents the overall robustness and resistance to stress of the ulna. All these indices relate to the robustness of the forearm and extension of the elbow. The maned wolf, aardwolf and striped hyaena plot low on this axis, while the red fox, side-striped jackal and spotted hyaena plot high. While the aardwolf overlaps all Canidae except the red fox, the other Hyaenidae do not show overlap with the Canidae. The aardwolf and maned wolf, which are the two non-cursorial species, plot close together. I. ebu plots in the polygon formed by the maned wolf as well as the aardwolf.

Flexible discriminant analysis

Stepwise flexible discriminant analysis of the different species was performed on the 13 morphometric indices, of which 10 were selected using a greedy Wilk’s lambda F-test. The indices FEI, BI, HFI, MCRI, GI, OLI, MCHUM, URI, HEI and FRI (Table 3) were selected based on an F-test decision of 0.05 (Table 3, Table S7). The resulting plot of the first two axes mainly separates the fox and maned wolf from the other data (Fig. 6).

Together, the first two dimensions account for 71.63% of the variance of the data. The first dimension accounts for 44.56% of the variance, while the second dimension accounts for 27.06% of the variance. The test data were predicted by the model with an accuracy of 0.9.

The red fox is separated from the rest of the data and clusters to the left (Fig. 6). The more cursorial Hyaenidae plot together on the right. In the centre are the other Canidae and the aardwolf, with the maned wolf plotting lower on Canonical variate 2 (CV2). The coyote is predicted to be a wolf, while the brown hyaena data is predicted to be from a striped hyaena. When I. ebu is added to the model, it is predicted to be a maned wolf specimen.

Three other models were generated based on family, locomotion and habitat. As FDA reduces the dimensions of the tested groups by one, these were one-dimensional models (Fig. S4). The family model based on the training data reclassified the test data with an accuracy of 0.95, based on the indices HEI, MCHUM, HFI, GI, SMI, FRI and URI (Table 3, Table S7). The extant specimens were correctly classified. I. ebu scores a low value on CV1 and was predicted as Canidae, which is incorrect. The locomotion model reclassified the test data with an accuracy of 0.86, based on OLI, MCRI and HFI (Table 3, Table S7). All extant specimens are predicted to be cursorial, which is correct. I. ebu was predicted as cursorial and has an intermediate value on CV1. Finally, the habitat model reclassified the test data with an accuracy of 0.95, based on BI, HFI, FEI, MCRI and MCHUM (Table 3, Table S7). The coyote is correctly predicted to be a mixed habitat species, while one of the brown hyaenas is incorrectly also assigned to being mixed habitat.I. ebu was predicted as an open habitat species and has a value slightly above intermediate for the plot.

Truss analysis

Non-metric multidimensional scaling

A repeat of the best solution was reached after 26 tries. Stress is 0.03. The Shepards plot shows an R² of the non-metric fit of 0.99 (Fig. S2) and no goodness of fit value is higher than 0.01 for the individual points (Table S6). There is a lot of overlap between species, mainly on the left side of the plot (Fig. 7).

The first axis is mostly controlled by distance 5–6, the distal width of the trochlea, and distance 4–5, the distal width of the capitulum (Table 6) and reflects overall size. The second axis is primarily controlled by distance 5–6 and distance 2–5 and relates to the distal extension of the trochlea. It separates species like the wolf, with a less extended trochlea and squarer anterior distal humerus, from species like the striped hyaena, with a more extended trochlea. I. ebu is separated from the other species, indicating an intermediate size and trochlear extension.

Table 6:

NMDS1 and NMDS2 loadings of the non-metric multidimensional scaling for the truss analysis.

Distances	NMDS1	NMDS2
2-6	0.10	0.01
2-5	−1.53	−0.45
3-5	−0.03	−0.02
1-5	−0.02	−0.18
2-4	0.36	−0.18
5-6	3.43	0.71
4-5	3.12	0.07
3-4	−1.04	−0.11
2-3	2.15	0.37
1-2	−0.65	−0.02
1-6	0.44	−0.06

DOI: 10.7717/peerj.17405/table-6

Flexible discriminant analysis

Flexible discriminant analysis of the different species was carried out for the different truss distances. The model accounts for 90.19% of the variance in its first two dimensions. CV1 accounts for 51.91% of the variance. It reflects the shape of the capitulum and with it, the overall anterior distal humerus from a squarer (wolf) to a more rectangular (brown hyaena) shape. CV2 accounts for 38.28% of the variance. It reflects the overall size of the specimens (Fig. 8). Model accuracy is 0.83. It separates species in a manner similar to the NMDS. The coyote is predicted to be an aardwolf, while the brown hyaena data is closest to the striped hyaena. When I. ebu is added to the model, it is predicted as being a specimen of the aardwolf.

Three other models were generated based on family, locomotion and habitat. As FDA reduces dimensions of the tested groups by one, these were one-dimensional models (Fig. S5). The family model reclassified the test data with an accuracy of 0.8. I. ebu was predicted as Hyaenidae and has a relatively high value on the plot. The extant species are correctly assigned to their own families. The locomotion model reclassified the test data with an accuracy of 0.75. I. ebu has a very high value and was predicted as having a generalist locomotor mode. The coyote was also assigned a generalist category, while the brown hyaena was once predicted to be cursorial and once a generalist. Finally, the habitat model reclassified the test data with an accuracy of 0.47 and I. ebu was predicted as an open habitat species with a value close to the mean. The extant species were also predicted to be open habitat species, which is incorrect for the coyote, but correct for the brown hyaena.

Flexible discriminant analysis validation

The flexible discriminant analyses of species have high reclassification rates and sensibly interpret the brown hyaena and coyote. The brown hyaena is classified as the striped hyaena in both the linear and truss flexible discriminant analysis. This classification makes sense, as they overlap in the NMDS of both methods and are simply put, large species of hyaena (Figs. 5 and 7). The coyote is predicted by its overall morphology to be a wolf, another member of its genus. The truss analysis would instead predict an aardwolf, as it is small like I. ebu.

The family models appear quite accurate as well, with no classification errors and high reclassification rates in both the linear morphometric and truss analysis (0.95 and 0.8 respectively). The incorrect predictions of the coyote and brown hyaena show that the locomotor and habitat truss models are less accurate (Fig. S5). In the habitat model of the truss analysis the brown hyaena even appears to be a far outlier, indicating that its morphology is not captured well by the model. This model also has the lowest reclassification score, at 0.47.

Permutational analysis of variance

Indices HEI, BI and MCHUM (Table 3) have the most significant differences in the pairwise PERMANOVAs (Fig. S3), showing that forelimb length ratios and epicondylar breadth are important characteristics when separating species. HEI and BI are also important characteristics for separating cursorial carnivores from other groups (Samuels, Meachen & Sakai, 2013), indicating that these indices detect similar differences in morphology in this smaller dataset. The similar proportions of I. ebu and the maned wolf would suggest similar adaptations.

Size differences

Size explains a large amount of dissimilarity in all of the species plots (Figs. 5, 6, 7 and 8). This pattern is particularly clear in the truss plots (Figs. 7 and 8), where the first axis of the NMDS and the second axis of the Truss seem to show mostly size differences. A similar size pattern was observed by Andersson (2004), Fig. 6), where small species are quite uniform in the main axis that reflects shape. At larger sizes, the differences between families increase with size. These dissimilarities may relate to a shift in diet from small to large prey (Carbone et al., 1999; Andersson, 2004).

Shape differences

The linear morphometrics NMDS results show that forearm robustness and the relative size of the olecranon present important differences in shape in these species (Fig. 5). The maned wolf, aardwolf and striped hyaena all plot lower on NMDS2. They all have a shorter olecranon, which reflects the fact that these predators do not need to apply much strength with their forelimbs during hunting and handling of prey (Martín-Serra, Figueirido & Palmqvist, 2016). Furthermore, they have less robust forearms. The striped hyaena is predominantly a scavenger (Rieger, 1981) and thus plots lower than the spotted hyaena, which more commonly hunts large prey (Hayssen & Noonan, 2021). The maned wolf is a generalist species that feeds on vegetal items as well as small to medium-sized prey (Rodrigues et al., 2007). The aardwolf uses its long tongue to consume termites, but never uses its front legs to dig (Kruuk & Sands, 1972). Both species have a generalist locomotor mode (Samuels, Meachen & Sakai, 2013). I. ebu plots within the range of the maned wolf, not far from the aardwolf, indicating some overlap with these species in forearm function. Overall, it appears that NMDS2 reflects shape that is impacted by hunting strategy and cursoriality.

The truss analysis NMDS shows that for the elbow joint the extension of the medial trochlear flanges is an important discriminator (Fig. 7). In this analysis, Hyaenidae have more extended medial trochlear flanges. This extension reflects an increase in lateral stability during humeral articulation, which is an indicator of an animal that grapples with its prey (Andersson, 2004). Andersson (2004) placed both Hyaenidae and Canidae in a group with the cheetah (Acinonyx jubatus) to indicate non-grapplers. Only the maned wolf is interpreted as a grappler but has the elbow joint morphology of a non-grappler (Andersson, 2004). However, as is the case for Fig. 6 of Andersson (2004), the canids plot higher than the hyaenids.

Phylogenetic signal

The linear morphometrics NMDS and FDA species plots both show a strong separation between families (Figs. 5 and 6). These differences may be due to size, as most hyaenids are larger than canids, or due to a phylogenetic signal in the differences in morphology, such as the SMI and HEI (Table 3) of the NMDS (Fig. 5).

The presence of a phylogenetic signal within the shape variables cannot be eliminated because extant canids have a cursorial ancestry (Andersson, 2004). This can be seen in the maned wolf, a generalist species that has traits of a more cursorial animal (Andersson, 2004; Samuels, Meachen & Sakai, 2013). A phylogenetic signal might even be more widely present in the data. However, skeletons not only reflect relatedness, but their form also reflects their function. Therefore, while phylogenetic corrections could have been implemented, they would be of limited use in this relatively simple, two-family case.

The ecomorphology of I. ebu

While a quick look at the variance showed a closer similarity of I. ebu to the maned wolf in the overall morphology of the limb bones (Fig. S3), both the maned wolf and aardwolf showed similar values in the NMDS (Figs. 5 and 7). The linear species FDA (Fig. 6) predicted I. ebu to be similar to the maned wolf, far from the other species, in accordance with the interpretation of the boxplots. The truss analysis of the distal humerus showed a closer affinity to the aardwolf in both the NMDS and species FDA (Fig. 7, Fig. 8), which is likely due to its size, which appears to correspond with a more uniform elbow joint shape (Andersson, 2004).

While the family of I. ebu was predicted as Canidae in the linear analysis, the truss analysis interpreted the family as Hyaenidae (Fig. S4, Fig. S5). It may be that the morphology of I. ebu is overall more similar to Canidae, but the elbow is more similar to Hyaenidae. While the locomotor model predicts cursorial locomotion for the overall morphology, the elbow predicts generalist locomotion. However, both species predictions would suggest a generalist locomotor mode. The models interpret I. ebu as an open habitat animal, in agreement with Werdelin (2003). The Lower Nawata represents a relatively mixed habitat, as pure grassland was likely not long-lived but could have been present (Wynn, 2003). The maned wolf, while mainly occurring in grasslands, can also be found in savannah woodlands (Queirolo et al., 2011). Given its similar morphology to the maned wolf, I. ebu could also have been present in these types of environments.

Overall, the long, gracile limbs of I. ebu were not an adaptation for cursoriality, but for being able to look over the tall grasses of its environment and pounce on prey, similar to the maned wolf (Hildebrand, 1954; Janis & Wilhelm, 1993) and possibly the serval (Leptailurus serval), a felid not analysed here that removes prey from crevices (Janis & Wilhelm, 1993; Ewer, 1998). The longer legs of I. ebu also contribute to its walking efficiency, as locomotion efficiency increases with longer legs at all gaits (Pennycuick, 1975; Janis & Wilhelm, 1993).