## Introduction

This study implements a large scale and methodologically homogenous critical assessment of the question using a statistical approach in modern “ungulates” (Perissodactyla, Artiodactyla, and Paenungulata). The aim is to document the actual, neutral head posture in life of modern species (using field observations) in order to compare this to the head posture inferred from LSC orientation in a dry skull (using CT scanning). This will enable us to address if the orientation of the plane of the LSC is a good proxy to reconstruct the head posture of extinct species, in order to ultimately make future paleobiological reconstructions more reliable. We will also test if some variables such as diet, body size, habitat, and head-butting are significantly correlated to head posture and/or LSC orientation as is usually believed24,30,45,47.

## Materials and methods

### Sampling

As the inclusion of a statistically-significant number of taxa was essential to this study, we chose to focus primarily on ungulate-grade mammals (i.e. Paenungulata, Artiodactyla, Perissodactyla, and Tubulidentata). Ungulates are more abundant than carnivores or primates in zoos, easily identifiable, and well represented in institutional dry skull collections. They display a wide array of body sizes, a greater variety of documented head postures47, a wider range of expected inner ear orientations (as hypothesized from the inclination of the snout compared to that of the brain-case48), and more varied degrees of adaptation to head-butting34 than any other mammalian group. Moreover, they are the ideal target group to address if diet (browsing v. grazing) plays a significant role in the orientation of the LSC, as previously suggested in the literature20,24. Finally, they usually display an elongated snout, which makes it easier to compare the orientation of the head in live animals to that of the corresponding dry skulls.

### Head posture in live animals

Head posture was documented by taking pictures of zoo animals in lateral view using a camera equipped with a spirit level (Fig. 1) to ensure that pictures were taken as close to the horizontal plane as possible. The animals were photographed in 2018 and 2019 at the National Zoological Garden, Pretoria (South Africa), Johannesburg Zoo (South Africa), Montecasino Bird Garden, Fourways (South Africa), Lory Park Animal and Owl Sanctuary, Midrand (South Africa), Ménagerie du Jardin des Plantes, Paris (France), Parc Zoologique de Paris (France), Prague Zoo (Czech Republic), Chester Zoo (United Kingdom), Zoologischer Garten Berlin (Germany), Tierpark Berlin (Germany), and Zooparc of Beauval (France). The saiga antelope pictures were kindly provided by K.H. Vogel. The dataset represents about 10,000 pictures documenting the head posture of 129 species and is available here: https://osf.io/4vpnj/?view_only=3dc987012fcd44a6a64ad7d8949ec01f (https://doi.org/10.17605/OSF.IO/4VPNJ). The pictures were taken from outside the enclosures to avoid interaction with the animals. It was essential for this study that the animals remain calm and act naturally, so their environment was not disturbed, and the animals were not put on leash or isolated. As such, individual identification was not possible. Representatives of both sexes are mixed in the dataset as sexes could not always be determined. The typical photography set up is illustrated in Fig. 1.

The orientation of the head compared to the horizontal plane was measured by J.B. using ImageJ as the angle between the horizontal border of the picture (horizontality of which was ensured by the use of a spirit level on the camera, Fig. 1) and the main axis of the head (traced as the axis running from just above the upper lip to the middle of the occiput on the back of the head) in strict lateral view (Fig. 2a). An average neutral head posture was then calculated for each species (Table 1). The intraspecific standard deviation (measurement error) for neutral head posture is ± 1.6°.

The bony labyrinth is one of the first organs to completely ossify in mammals as its adult size and shape are reached at mid-gestation50,51. However, the orientation of the LSC seems to show age-related variations in some tetrapod species, including humans, which may impact their head posture26,52,53. As such, juveniles were excluded from the dataset.

### Head posture from dry skulls

A total of 285 medical quality CT-scans and micro-CT scan representing 118 species mostly from the American Museum of Natural History (AMNH), Ditsong Museum (AZ and TM), Evolutionary Studies Institute of the University of the Witwatersrand (BP), Wits Life Science Museum (WLSM), School of Anatomical Science of the University of the Witwatersrand (MS and ZA), Natural History Museum of Basel (NMB), Yale Peabody Museum of Natural History (YPM), and Zoological Museum of the University of Zurich (ZM) (see the Supplementary Table S1 for details), were used.

The bony labyrinths of each skull were segmented manually and reconstructed in 3D using the software AVIZO 9 (FEI VSG, Hillsboro OR, USA) at the virtual imaging labs of the Evolutionary Studies Institute and the Natural History Museum of Basel. The skull was reconstructed using either the Isosurface or threshold functions under the same software. The angle between the plane of the LSC and the main axis of the skull was then measured in lateral view in 2D (Fig. 2b). The plane of the LSC was determined visually in lateral view, following most previous authors19,20,22,24,25,26. The main axis of the skull was traced as the axis running from just above the premaxilla (approximately at the level of the centre of the nasal opening) to the middle of the occiput (Fig. 2b) in order to maximize the homology with the measurements taken on living animals. This angle represents the anterior tilting of the head if the LSC is considered horizontal. This angle is hereafter referred to as “the reconstructed cranial orientation” or “reconstructed head posture”. Measurements were taken bilaterally when both bony labyrinths were available and then averaged for each species (Table 1). For consistency, all measurements were taken by the same author (J.B.). None of the samples expressed strong lateral tilting of the LSC or an undulating morphology that could impede taking this measurement or affect its accuracy. The intraspecific standard deviation (measurement error) for reconstructed head posture is ± 2.1°.

The complete dataset of reconstructed head postures is available in the Supplementary Table S1. This dataset was complemented by measurements made on the published pictures from Girard and Schellhorn5,30 (see Supplementary Table S1).

As for the picture dataset, only the individuals showing reasonable signs of maturity (e.g. cranial bone fusion, erupted molars) were considered.

### Data processing

The dataset was analyzed using phylogenetic comparative methods to control for the non-independence of observations54,55,56. We used the time-calibrated phylogenetic tree of mammals of Bininda-Emonds et al.57 because it encompasses all the species in our dataset and fossils can be easily added to it in future analyses. The tree was pruned to match the species in our dataset using function ‘drop.tip’ in R package ape58. All subsequent analyses were performed in R v3.6.3 (R Core Team, 2020). The phylogenetic signal of individual variables was estimated using Pagel’s lambda59 for continuous features (reconstructed cranial orientation, neutral head posture, body mass) using function ‘phylosig’ in package phytools60. Lambda was chosen over the other commonly used estimator K61 because of the latter’s poor performance for trees with small sample sizes and polytomies62,63, both of which can be found in our dataset. For binary traits (head-butting; see below), phylogenetic signal was estimated with the D-statistic64 using function ‘phylo.d’ in package caper65. To test whether the plane of the LSC can be used as a reliable proxy to reconstruct the neutral head posture, we regressed the neutral head posture of living animals on the reconstructed cranial orientation using data in Table 1, and phylogenetic generalized least squares (PGLS) regressions66,67. PGLS were compiled using the ‘gls’ function in package nlme68, with correlation structures for each evolutionary model specified in ape58. A model selection procedure based on the corrected Akaike information criterion (AICc) was applied to the regressions using the package AICcmodavg69. Five evolutionary models were considered for this selection procedure (see70): Brownian Motion, Pagel’s Lambda, Ornstein–Uhlenbeck, Early Burst, and White Noise – i.e. non-phylogenetic, ordinary least squares (OLS) regression. All regressions were performed using raw and log-transformed data (natural logarithm). Both variables in the models are in the same unit and order of magnitude, and the models built with raw data showed a higher significance and met parametric assumptions better than models built with log-transformed data. For this reason, we used the former to assess the relationship between the two variables.

Because of the high degree of body mass allometry in neuroanatomical features71,72,73, body mass measurements for all species in the sample were taken from the literature (Supplementary Table S1) and included as a co-predictor to be tested against models built with only the reconstructed and neutral head postures as predictors in the AICc-based model selection procedures. The coefficient of determination and p-value for generalized least squares regressions cannot be compiled straightforwardly due to the autocorrelated structure of the residuals67. Following Paradis55, we compiled a pseudo-R-squared and p-value based on McFadden’s formula74, based on a likelihood ratio test between our model and a null model. Normality and homoscedasticity of the residuals were assessed using a Shapiro–Wilk test and a Q-Q plot, and graphically using residuals v. fit plots, respectively75.

Finally, phylogenetic one-way Analyses of Variance (phylANOVA)76 with False Discovery Rate posthoc corrections77 were used to test for a difference between groups in three separate factors, for both reconstructed cranial orientation and neutral head posture (Table 1). The first factor is diet, for which species were categorized as “browser”, “grazer”, “mixed” (for a mixed diet between browsing and grazing), or “other” (for omnivorous and myrmecophagous species). The second predictor is whether a species practices head-to-head combat. The head-butting category includes wrestling and ramming species (hereafter referred to simply as head-butting species) but excludes flank-butting species (e.g. giraffes). The last predictor is the habitat, which was scored between open (savannah or steppes), closed (forest or jungle), mixed (mix of open and closed habitats), rocky (for species living on steep, rocky slopes), or semi-aquatic. The scoring of all three predictors was done using the literature (see the list in Table 1). PhylANOVAs were performed using function ‘phylANOVA’ in phytools60.

### Ethics declarations

As the animals were not approached or armed, no ethical clearance was necessary for this study.

## Results

All the data in the dataset for which a phylogenetic signal could be measured (neutral head posture, reconstructed cranial orientation, body mass, and head-butting) carry a strong phylogenetic signal (lambda > 0.8 for the first three variables; D = − 0.2841056 for head-butting) (See Supplementary Table S1).

Species with body mass under 100 kg have an average neutral head posture of 30° and reconstructed cranial orientation of 39°, whereas species larger than 100 kg have a neutral head posture averaging 37° and average cranial orientation of 40°. This suggests an effect of body mass on head posture as was hypothesized by Köhler47, but not on the orientation of the LSC (Fig. 3). This is consistent with statistical analyses, which identify a very weak effect of body mass on neutral head posture (R2 = 0.040; p-value = 0.014), and none on reconstructed cranial orientation (R2 = 0.023; p-value = 0.054) using OLS. However, once corrected for phylogeny using PGLS, the effect of body mass on head posture (R2 = 0.030, p-value = 0.06025) and cranial orientation (R2 = 0.012, p-value = 0.9903) is no longer significant.

Phylogenetic regressions (Fig. 4a) identify a statistically significant (p-value = 4.519e−07) but relatively low correlation (R2 = 0.261) between neutral head posture and the reconstructed cranial orientation. This supports that the orientation of the LSC in life is correlated to the neutral head posture in “ungulates”. The equation of the linear model is:

$${\text{ Neutral head posture }}\left(^\circ \right)\, = \,0.{384}\, \times \,{\text{reconstructed cranial orientation }}\left(^\circ \right)\, + \,{13}.{468}.$$
(1)

The 95% confidence interval for the slope (0.242–0.526) is significantly different from 1, which means that this model cannot be approximated to an isometric relationship (which would be expected if the LSC was held horizontally).

The model including body mass, neutral, and reconstructed head postures, and the interaction term of the three as co-predictors was selected by AICc as fitting our data best (Fig. 4b). This model shows results very similar to those of the simple regression model, being significant with a slightly stronger correlation (R2 = 0.325; p-value = 3.235e−07). The equation of the resulting model is written:

$${\text{ Neutral head posture }}\left(^\circ \right)\, = \,{11}.{677}\, + \,0.{328}\, \times \,{\text{reconstructed cranial orientation }}\left(^\circ \right) - 0.00{1}\, \times \,{\text{body mass }}\left( {{\text{kg}}} \right) - 0.000{2}\, \times \,{\text{reconstructed cranial orientation }}\left(^\circ \right) \, \times {\text{ body mass }}({\text{kg}}).$$
(2)

The slope confidence interval is slightly lower than that of the previous model (for reconstructed cranial orientation: 0.148–0.508), which removes it even further from an isometric relationship (Fig. 4b). The very low coefficients for body mass and reconstructed orientation × body mass are not significantly different from zero (see Supplementary Table S1), and are reported here to ensure full transparency of our results.

Surprisingly, for both simple and multiple regression models, the evolutionary model selected by AICc was the Early Burst (EB) model61, representing a rapid adaptive radiation followed by stasis. EB models are known to be rarely selected as the best evolutionary model in such selection procedures78.

On average, browsers tend to hold their heads less tilted anteriorly (26°) than mixed feeders (32°), and grazers (36°) in neutral posture (Fig. 5). This seems to reflect on the reconstructed cranial orientation as browsers have a higher reconstructed head posture (33°), than mixed feeders (40°) and grazers (44°) (Fig. 5). Phylogenetic ANOVAs indicate a significant difference between browsers and grazers for the reconstructed orientation of the skull (F = 7.723; p-value = 0.046), but not for the neutral head posture (F = 2.663; p-value = 0.516). Mixed feeders are statistically indiscernible from both browsers and grazers in any case.

Variations of reconstructed cranial orientation do not show any trend with habitat preference (Fig. 6), but a slight trend toward more downwardly tilted head postures seems to occur in species living in a more open habitat (Fig. 6); however, this trend is not significant (F = 1.343; p-value = 0.792). Semi-aquatic species seem to have a more posteriorly tilted LSC resulting in a higher reconstructed cranial orientation (23°) than fully terrestrial species, but this does not reflect on the neutral head posture (Fig. 6). Unfortunately, the sample size for this category was too low to effectively test if this difference was significant or simply the result of the scarcity of semi-aquatic species in the dataset.

Head-butting is found to have a highly significant statistical effect on reconstructed head posture (F = 39.467; p-value = 0.002), with a difference of 13° between head-butting and non-head-butting species on average (Fig. 7); however, with only 1° difference on average (Fig. 7), the same is not true for the neutral head posture (F = 3.126; p-value = 0.591).

## Discussion

### LSC orientation is correlated with head posture, but is not horizontal

The assumption that the plane of the LSC can be used as a reliable indicator of the horizontal plane on dry skulls and thus can serve as a proxy to reconstruct the “habitual” or “alert” head posture in extinct species has been a long-held17, yet insufficiently tested hypothesis in mammals. Attempts to test this hypothesis have highlighted that the plane of the LSC is often tilted upward compared to the horizontal in most mammalian species5,6,11,30,45. A famously baffling example is that of humans, in which aligning the plane of the LSC to the horizontal plane results in a “habitual” posture of the head inclined 30° down anteriorly1,2,3,4,5,6,8,11,12,18,46. In the current study, the steenbuck (Raphicerus campestris) is the only species in which the average neutral and reconstructed head postures are the same (Table 1), which means that, on average, the LSC is parallel to the horizontal plane when the steenbuck’s head posture is neutral. Among the species for which the neutral head posture and reconstructed cranial orientation could be compared, only half of them show a difference between the averages of the two that is below 10°. As such, even though the plane of the LSC should be horizontal on theoretical grounds13, this is not the rule in “ungulates”.

The two phylogenetic regressions provided here (Fig. 4) are the first large sample size attempts to address the existence and nature of a correlation between the orientation of the plane of the LSC and the neutral head posture in mammals across a large taxonomic sampling. We find that whether corrected or not for body mass, the correlation between the reconstructed cranial orientation and the neutral head posture is significant (p-value < 0.0001); however, if the plane of the LSC was held horizontally in neutral head posture, the regression line should not differ significantly from an isometric line. Instead, both regression lines have slopes that significantly differ from 1 (Fig. 4), which means that they cannot be approximated by isometric lines. For this reason, though there is a significant correlation between the orientation of the LSC and that of the head in “ungulates”, the plane of the LSC should not be considered horizontal when reconstructing ancient head posture.

According to the phylogenetic regressions, the equation that describes the relationship between the reconstructed cranial orientation and the neutral head posture is given in Eq. (1), and that between cranial orientation, head posture, and body mass is given in Eq. (2). As estimating body mass in extinct species is always contentious79,80,81,82, the first of these equations may seem more practical to estimate the actual head posture of a given extinct ungulate species. In both cases, the variance of the residuals is high, which might indicate a low predictive power of these models (R2 equals 0.26 and 0.33, respectively).

Another hypothesis is that an overall misalignment of all three semicircular canals would enable all semicircular canals to record a component of horizontal and vertical accelerations43.

### The effect of phylogeny

Both the neutral and reconstructed head postures carry an important phylogenetic signal (Lambda equals 0.97 and 0.84, respectively). For paleontologists, this strong phylogenetic signal implies that the best way to predict the head posture of an extinct “ungulate” is to look at the neutral head posture of its modern relatives. In comparison, once the data are corrected for phylogeny, diet is found to have only a weak correlation with the reconstructed cranial orientation (F = 7.723; p-value = 0.046), and no significant effect on the neutral head posture (F = 2.663; p-value = 0.516). This reflects well in the dataset.

The Tylopoda is the group in which the head is the most consistently tilted upward, with an average neutral head posture of 10°. This reflects on their reconstructed cranial orientation which averages 17°. Tylopods nevertheless include grazers (genus Vicugna), mixed feeders (Camelus bactrianus), and browsers (Camelus dromedarius) that all keep their heads relatively high (Figs. 4, 8a). On the other end of the spectrum, pigs display remarkable consistency at holding their head low (average neutral head posture for Suoidea = 46°), with for example the grazing Phacochoerus and the browsing Catagonus both keeping their head 46° below the horizontal plane on average (Table 1, Fig. 4).

A more significant effect of body mass might nevertheless be found while including very small-bodied species (e.g. rodents, shrews) because their head posture would be more constrained by its proximity to the substrate. Similarly, species with a sprawling posture would also have to keep their head higher, as already observed in many reptiles12,41,42,43.

The relationship between LSC orientation and phylogeny has been empirically anticipated as the way the LSC enters the vestibule (either directly above the posterior ampulla or at different levels within the ampulla) is distinctive between different clades of ruminant50,51,86,87,88.

### Diet

As grazers have to keep their head low while foraging on grass, whereas browsers have to catch leaves higher in bushes and trees, and because herbivores spend most of their time acquiring low-energetic food49,92,93, it is expected on an evolutionary scale that the skull, neck musculature, and vestibular apparatus of herbivores would adapt to these different feeding strategies and that it would reflect in their head posture, even at rest20,30,45. As such, a gradually more anteriorly tilted neutral head posture is expected as moving from browsing species to mixed feeders, and finally grazers20,30. Schellhorn30 was the first to compare the reconstructed cranial orientation of Rhinocerotidae to their actual head posture (using an open-access database of photographs) and found results consistent with such a gradient.

The situation in rhinocerotids illustrates that for the whole dataset. Diet is found to have no statistical effect on neutral head posture (p-value = 0.516), and its relationship to the reconstructed cranial orientation is barely significant (p-value = 0.046). Average values for neutral and reconstructed head postures show a visible increase in tilting with a more grass-rich diet, but there is a strong overlap between dietary groups for both variables (Fig. 5). This suggests that even though diet could potentially be reconstructed in extinct species using the orientation of the LSC, caution should be taken as i) browsers and grazers could be statistically discriminated, but mixed feeders could not; ii) the correlation between reconstructed cranial orientation and diet does not seem to reflect on the neutral head posture. The reason why remains unknown; and iii) the high p-value suggests that adding more data (particularly CT data) in the future may affect this correlation.

No significant result indicative of a correlation between habitat and head posture or reconstructed cranial orientation was found; however, semi-aquatic species show a noticeably high reconstructed posture of the skull on average (23°) compared to other species (Fig. 6). The low number of semi-aquatic species in the dataset (one rhinocerotid, three tapirids, and two hippopotamids, Table 1) likely prevents this trend to be identified as significant in our sample. A high head posture is not observed in semi-aquatic species (Fig. 6), even though it would be expected of species that have to keep breathing above water level most of the time while immersed27,30. Recently, a semi-aquatic habit for the Triassic archosaur Proterosuchus and the therapsid Lystrosaurus has been hypothesized as these two species would have their head tilted upward anteriorly when the plane of the LSC is horizontal27,94 (Fig. 11c). In Proterosuchus this upward tilting would be about 17°27, whereas in Lystrosaurus it would be between 19° and 23°94. However, an upward tilting of the neutral head posture is also habitually observed in many fully terrestrial species, such as Camelus dromedarius and was occasionally spotted in Capra ibex (Fig. 11a; see Supplementary Table S1). Among modern archosaurs, an upward tilting of the head is observed in some sea birds and the common starling Sturnus vulgaris8. An upward tilting of the beak when the dry skull is held with the plane of the LSC horizontal seems to be observed in the razorbill (Alca torda) and the Heron (Ardea cinerea) (as inferred from the figures and measurements in Duijm8, see Supplementary Table S1); however, the reconstructed cranial orientations in the razorbill and heron vary between 0° and -4° only (Fig. 11b).

Among fossil species, the likely terrestrial sauropod Ngwevu intloko (Fig. 11c) would also have had its reconstructed cranial orientation tilted 17° above the horizontal (identified as Massospondylus carinatus in Sereno et al.20; see Chapelle et al.95). The peculiar orientation of the LSC in Lystrosaurus, Proterosuchus, and Massospondylus is remarkable as such a downward tilting of the LSC superior to 15° has never been found in any modern species to date, particularly not in the semi-aquatic species studied here which all have a posteriorly tilted LSC as all other “ungulates” (average reconstructed cranial orientation of semi-aquatic species = 23°) (Fig. 6; Table 1). The heron and razorbill mentioned above would have an anterior tilting of less than 4° (Supplementary Table S1), and Duijm’s8 dataset includes mostly semi-aquatic species, which limits comparisons. Overall, a correlation between an upward tilting of the LSC and semi-aquatic lifestyle is not supported by current data.

Noteworthily, the most aquatic species of the dataset, Hippopotamus amphibius, stands out on the scatter-plots as an outlier (Fig. 4). This is due to the large difference between its strongly anteriorly tilted neutral head posture and its almost horizontal reconstructed cranial orientation (difference = 39°). Hippopotamus amphibius normally spends very little time on land49, and it can be hypothesized that its bony labyrinth morphology would be more adapted to life in water than on land96. This hypothesis is supported by the orientation of the head while swimming in H. amphibius, which is more consistent with its reconstructed cranial orientation (Fig. 12), and the fact that the difference between the neutral and reconstructed head postures in the more terrestrial Hexaprotodon liberiensis (23°) falls more within the range of variation of “ungulates” (Fig. 4). Further field observations of underwater hippopotamus head posture will be necessary to address this hypothesis.