1In bioarchaeology and forensic anthropology, the fundamental task of the anthropologist when confronted with human skeletal remains is to provide the biological profile of the individual concerned. Biological sex is a crucial component of this profile, because the estimation of other variables such as stature and age (Trotter and Gleser, 1958; Brooks and Suchey, 1990) will depend on the sex assessment, which therefore has an important role in the inferences that can be drawn about the deceased’s identity. Estimating the sex of individuals that are buried in the same funerary assemblage helps to characterise the palaeodemographic patterns of the past population under study, but also to evaluate their health and living conditions and to identify the community’s funerary practices (Duday, 2009; Nilsson Stutz and Tarlow, 2013; Boucherie, 2020). The objective when investigating human remains in a forensic case is quite different: sex estimation in this context facilitates the investigations by narrowing down the possibilities regarding the victim’s identity (Christensen et al., 2014). This has significant implications, as the use of a method liable to produce an erroneous sex estimation will clearly have knock-on effects for the anthropological results (Bruzek and Murail, 2006; Klales, 2020a).
2Estimating the sex of adult skeletal remains can be done routinely thanks to easy-to-use, reproducible and reliable methods based on the os coxae. The classification accuracy of these methods, which rely either on morphoscopic features (Bruzek, 2002; Klales et al., 2012; Santos et al., 2019) or on measurements (Murail et al., 2005; Bruzek et al., 2017; Santos et al., 2020), is above 95%. However, not every situation in archaeological or forensic contexts presents the ideal scenario in which the os coxae is available and fully preserved. This is particularly the case in forensic contexts where human remains frequently undergo taphonomic alterations (Komar and Grivas, 2008; Christensen et al., 2014) that can jeopardise the reliability of such methods. For instance, using the probabilistic sexual diagnosis method (Diagnose Sexuelle Probabiliste - DSP) requires collecting at least four measurements from the os coxae (Murail et al., 2005; Bruzek et al., 2017), but the final accuracy varies depends on which measurements are available (Quatrehomme et al., 2017; Chapman et al., 2020).
3If the os coxae is not usable, anthropologists can also perform a primary sexual diagnosis based on other skeletal elements like the skull or long bones (Jantz and Ousley, 2005; Spradley and Jantz, 2011; Curate et al., 2017) or, in the case of a funerary assemblage, they can proceed to a secondary sexual diagnosis (Murail et al., 1999; Santos, 2021). However, both approaches have limitations: while the second is not suited to isolated human bones, as frequently seen in forensic contexts, the first requires access to optimally preserved bones.
4Given these limitations, genetic analyses, and more recently proteomics, are therefore commonly seen as the only option to identify the sex of severely damaged skeletons (Boberova et al., 2012; Stewart et al., 2017). However, budget and technical constraints mean that they cannot be implemented systematically, especially on the scale of a funerary ensemble. The results can also be impaired by contamination and degradation phenomena (Llamas et al., 2016; Ottoni et al., 2017).
5To increase the number of adult individuals that can be sexed, it thus appears necessary to design an alternative sexing method suited to fragmentary remains, by using a skeletal element that is at once sexually dimorphic and presents a lower risk of taphonomic alteration. The cranial base – defined by the occipital and temporal bones – is promising in this respect: this cranial area not only presents visual differences between males and females, although these need to be adequately quantified (Ascádi and Nemeskéri, 1970; Walker, 2008), but also tends to have a high rate of taphonomic survival (Holland, 1986; Bello et al., 2002) due to its highly resistant component elements such as the basilar part, the occipital condyles and the petrous part of the temporal bone. This paper synthesises the main results obtained by a study that did explore the sexual dimorphism of the human cranial base from a large assemblage of identified western European adults. We present the predictive sexual diagnosis models that were built up from this sample and explain their use, which is now facilitated by the implementation of a free R package.
6A large skeletal assemblage composed of individuals of known sex was compiled with a view to designing a new sexual diagnosis tool suited to fragmented human remains. Seven osteological collections curated in four different western European countries (Belgium, France, Switzerland and Portugal) were used. Altogether, 537 adults of more than 20 years of age, including 261 females (x̄=52.2 years, SD=18.3) and 276 males (x̄=50 years, SD=17.9) were examined (table 1). Pathological individuals were excluded. These identified osteological assemblages were collected either by museums or universities after the partial or complete dismantling of 19th-century cemeteries (e.g., the Simon, Schoten, Châtelet, Lisbon and Coimbra collections) or by medical institutions after past autopsies or under current body donation programmes (Olivier and Nice collections). Most of the individuals included in the study lived during the 19th and 20th centuries (n=445) and 92 of them were alive during the 21st century. Sex is known for each individual and the exact age is known for 84% of the sample.
Table 1
Information on the skeletal assemblages used in this study |
Informations sur les assemblages osseux utilisés dans cette étude
7To reduce subjectivity induced by morphological rating, an exclusively morphometric approach was carried out. After reviewing the literature, 73 measurements were selected to characterise the morphometry of the occipital and temporal bones and their structural elements (i.e., the basilar part, the foramen magnum, the occipital condyles and the mastoid process) (figure 1). These measurements and their definitions are listed in detail in table 2. The lack of standardisation in the terminology of measurements, especially for the mastoid process, has been highlighted (Petaros et al., 2015), and all available definitions of the length and width of this element were therefore included in the study. Measurements were registered in millimetres using a sliding calliper or a metric tape. To assess repeatability, variables were collected a second or third time by the author, at intervals of two weeks to one month, from 238 individuals (44% of the total corpus). Additionally, three other observers collected all the measurements for 62 individuals belonging to the Châtelet, Schoten and Nice collections (11% of the total corpus). To evaluate the repeatability and reproducibility of the protocol, an intra-class correlation coefficient (ICC) was calculated for each variable. When the ICC was below 0.70, the measurements were immediately excluded from the set. The remaining 54 reproducible measurements were used to construct the predictive models.
Figure 1
Example of measurements used to investigate the sexual dimorphism of the cranial base: A. length of the foramen magnum; B. maximal bicondylar breadth; C. length 2 of the left mastoid process (direct distance between porion and mastoidale) |
Exemple de variables métriques utilisées pour étudier le dimorphisme sexuel de la base du crâne : A. longueur du foramen magnum ; B. largeur bicondylaire maximale ; C. longueur 2 du processus mastoïde gauche (distance directe entre le porion et le mastoidale)
Table 2
List of the 73 measurements taken on the cranial base to investigate its sexual dimorphism |
Liste des 73 variables métriques enregistrées sur la base du crâne pour en examiner le dimorphisme sexuel
8This study is not the first to examine the sexual dimorphism of the occipital and temporal bones in adult individuals. However, some issues in previous studies were noted:
9- the sexual dimorphism of the occipital and the temporal bones was investigated separately: very few predictive models combined the measurements of both bones (Holland, 1986; Gapert, 2009; Macaluso, 2011; Say-Liang-Fat, 2014);
- the statistical treatment of the metric data used to construct the predictive models was not systematically performed in a strict and clear manner (e.g., absence of intra- and inter-observer error tests, lack of consistency in univariate or multivariate analyses and in the use of cross-validation and independent validation samples, infrequent use of a decision threshold other than 0.50) (Holland, 1986; Kemkes and Göbel, 2006; Bernard and Moore-Jansen, 2009; Gapert, 2009; Gapert et al., 2009a-b; Macaluso, 2011, Jaja et al., 2013; Say-Liang-Fat, 2014; Chovalopoulou and Bertsatos, 2017).
10In this study, the statistical analyses were therefore carried out with caution and according to a codified and rigorous protocol to ensure the highest reliability of the results.
- 1 Individuals with more than 30% of missing data for each anatomical zone (occipital and temporal bon (...)
11After managing missing data with the R package missMDA1 (Josse and Husson, 2016), checking data normality and variance equality, Welch’s t-tests were used to highlight significant differences between male and female measurements, with a p-value significance threshold of 0.05. Any measurement that was not found to be significantly dimorphic was excluded when building up the predictive models. The directional symmetry of the 17 bilateral measurements was also assessed with paired t-tests. As 11 of these measurements showed significant directional asymmetry (table S1 – p-value<0.05), all of them, both left and right, were included in the subsequent predictive statistical analyses.
12Predictive sexual diagnosis models were then developed in R v3.4 using logistic regressions and a stepwise approach for variable selection. The equations were developed according to the following procedure:
13- the predictive models were developed and cross-validated within a training sample comprising 437 adults (table 1);
- the models were then validated on an entirely independent sample (the validation sample) comprising 100 individuals from the Coimbra collection (table 1).
14Multivariate models were developed by either combining occipital and temporal measurements, or considering only occipital variables or temporal variables, or again with different combinations of measurements (e.g., measurements of the right part of the occipital bone with measurements of the left temporal bone). The main objective was to build up models that could be adapted to different patterns of preservation of the cranial base (complete or fragmentary), while keeping to a reasonable number of measurements for each model (<15) to ensure ease of use.
- 2 The accuracy percentage, or correct classification percentage, corresponds to the number of individ (...)
- 3 The error rate is a percentage used to evaluate the proportion of individuals in the training or th (...)
- 4 The sex bias is a percentage used to show the difference between the proportion of females and the (...)
- 5 In this case, a percentage of indeterminacy is calculated alongside the accuracy to show the propor (...)
15For each model, accuracy percentages2, error rates3 and sex bias4 were evaluated in the training sample, after leave-one-out cross validation (LOOCV), and in the validation sample by applying four different thresholds on the posterior probabilities: 0.50 – as is common in anthropology; 0.70; 0.80 and 0.95 – as used for DSP (Murail et al., 2005; Bruzek et al., 2017). The latter three thresholds allow an ‘indeterminate’ category to be included in the sex estimation results (Santos et al., 2019; Galeta and Bruzek, 2020), so that if the posterior probability of the sex assessment for an individual is below the threshold (0.70, 0.80 or 0.95), the individual’s sex is classified as indeterminate5.
16Finally, the applicability and reliability of the predictive models were assessed in an application sample (n=29) comprising complete or fragmented crania from two Belgian archaeological collections (Coxyde site - 12th-15th century; Koekelberg site - 19th century) and a Belgian forensic context (Bois du Cazier identification mission, see Polet et al., 2024), in which sex was initially known either by DSP or genetic analyses (table 1).
17Among adults, each measurement, except the foramen magnum index, was found to be significantly different between the sexes, males having larger values than females (p<0.05) (table S2). Similar observations have previously been made in other European populations (Gapert et al., 2009a-b; Kemkes and Göbel; 2006; Macaluso, 2011; Say-Liang-Fat, 2014; Amores-Ampuero, 2017; Chovalopoulou and Bertsatos, 2017).
18Thirteen predictive sexual diagnosis models were built up, using a total of 30 different measurements (table 3). AD1 is a model suited to a complete cranial base (where both the occipital and temporal bones are preserved), while AD2 to AD4 are exclusively composed of occipital variables, AD5 to AD7 of temporal variables, and models AD8 to AD13 are suited to different patterns of preservation of the cranial base. The 4 best performing models (AD1, AD9, AD10 and AD13) reached 85.1 to 86.8% of correct classification after cross-validation in the training sample, and 85.1 to 92.5% accuracy in the validation sample, with a threshold of 0.70 (table 4). This threshold was found to be the best compromise to reach the highest percentage of accuracy, the lowest error rate (<15%) and the lowest proportion of indeterminacy (<33%). The models developed with temporal measurements only (AD5, AD6 and AD7) had higher accuracies than those set up from the occipital bone (AD2, AD3 and AD4) (table 4).
Table 3
Predictive models built up from the adult cranial base with details of the measurements used |
Modèles prédictifs établis sur la base du crâne adulte avec précisions des variables métriques utilisées pour chacun d’entre eux
Table 4
Correct classification, sex bias and percentages of indeterminacy and error found for each predictive model built up from the adult cranial base in the training sample after LOOCV and in the validation sample, with a 0.70 threshold |
Pourcentages de classification correcte, de biais de sexe, d’indétermination et d’erreur obtenus dans l’échantillon d’apprentissage après validation croisée et dans l’échantillon de validation, avec un seuil de décision fixé à 0,70
19With a 0.95 threshold, accuracies were high (>90%), but the number of indeterminate individuals was overwhelming (>85%) (table S3). This latter point was also observed with a 0.80 threshold (table S4).
20The accuracies obtained in this study, even with a 0.50 threshold, outperformed those obtained in previous research that combined occipital and temporal measurements (Say-Liang-Fat, 2014) or considered occipital (Gapert et al., 2009a-b; Macaluso, 2011; Chovalopoulou and Bertsatos, 2017) and temporal bones separately (Kemkes and Göbel, 2006; Bernard and Moore-Jansen, 2009; Jaja et al., 2013).
21When applied to either complete or highly fragmented archaeological and forensic specimens, the best performing predictive models produced good correct classification results (>80%), with the 0.70 threshold (table S5).
22To facilitate their use on adult target individuals, the predictive sex estimation models presented above were incorporated into an R package called BASE (BAsicranial Sex Estimation). This offers a graphic user interface through an R-shiny application, so that no specific knowledge of the R language is needed.
23This R package is available on GitLab (https://archive.softwareheritage.org/browse/directory/5e3e3753e1f9c360b0ddbebaa2d2b42e5d7686d9).
24Installation prerequisites and other technical instructions can be found in the README.
25The R package comprises two tabs. The "Instructions" tab gives a description of each measurement included in the models (figure 2). In this tab, the user can retrieve the measurement abbreviation for each variable, its full name, its definition with the initial reference and the precision of the tool required to record it. The two illustrations below aim to give the user a clear view of how to measure the variable correctly, with a schematic description of the measurement and a photograph of a real cranium from which the measurement was taken.
Figure 2
The "Instructions" tab of the BASE R package with the description of the Ast_po_L/Ast_po_R measurement corresponding to the distance between the asterion and the porion on the temporal bone. The definition, measuring tool, initial reference and two illustrations are provided to help the user in taking measurements |
Présentation de l’onglet "Instructions" du package R BASE avec la description de la variable métrique Ast_po_L/Ast_po_R correspondant à la distance entre l’astérion et le porion sur l’os temporal. Les précisions concernant la définition, l’outil utilisé pour faire la mesure, la référence bibliographique et deux illustrations sont partagées avec l’utilisateur pour faciliter l’étape d’acquisition des mesures
26The first "Analysis" tab is dedicated to the sexual diagnosis of a target individual (figure 3). Here, users can choose which predictive model they want to apply from the list in the top left corner. Below the name of the model is a reminder of the types of measurement included in it: from the occipital bone only, from the temporal bone, or from both bones. These details are useful for the anthropologist to decide which models are applicable to the target individual, depending on the state of preservation of the cranial base under study.
Figure 3
The "Analysis" tab of the BASE R package. Example of the application of predictive model AD2, on the fragmented individual KOEK 35 of our application sample (Koekelberg site - 19th century), detailing the sex estimation result obtained with a 0.70 threshold. The sex finally attributed from the cranial base (F) measurements is consistent with the female sex established by DNA for this archaeological individual |
Présentation de l’onglet "Analyse" du package R BASE. Exemple de l’utilisation du modèle prédictif AD2 sur l’individu fragmenté KOEK 35 de notre échantillon d’application (site du Koekelberg – XIXe siècle). Les résultats obtenus pour l’estimation du sexe, avec un seuil de décision de 0,70, sont détaillés. Le sexe attribué à partir de la morphométrie de la base du crâne (F) est cohérent avec le sexe féminin établi par analyses génétiques sur cet individu archéologique
27In the top right corner, users can choose from 0.50 to 1 for the posterior probability threshold they want to apply. The threshold is set by default to 0.70 since this is recommended as the best compromise, as investigated in our study, to obtain the highest percentage of accuracy, the lowest error rate and the smallest proportion of indeterminacy. Setting a threshold of 0.50 implies that the target individual will be categorised as either female or male in any case (Galeta and Bruzek, 2020). When choosing a threshold above 0.50, the target individual will be categorised as either female or male only if the posterior probability obtained reaches the specified threshold. If the posterior probability is below the threshold, the individual’s sex remains indeterminate.
28In the "Data input" section, the user can enter the measurements collected from the target individual according to the model chosen. The ‘Perform sex estimation’ button can then be selected. Any outlier measurement, i.e., completely outside the range of the reference sample measurements, will be underlined in red. This security procedure was included to flag up possible input or measurement errors.
29The results obtained for the target individual can be retrieved in the lower part of the tab. In the left-hand section, the sex attributed to the target individual is given by detailing the posterior probability obtained according to the threshold used. This posterior probability should be provided when discussing the result of the sexual diagnosis in the anthropological report. As a complement, in the right-hand section, the accuracy compared to the initial training sample, after leave-one-out cross-validation, is displayed by a confusion matrix.
30These recommendations should be considered when applying the predictive models built up from the cranial base to a target individual.
31These sex estimation models were developed from a reference sample composed of western European adults. Consequently, we recommend applying them only to adults whose skeletal maturation was complete and who are assumed to be of western European origin. Since secular trends have affected the morphology and morphometry of the cranium (Jantz and Meadows Jantz, 2016; Weinsensee and Jantz, 2011; Langley and Jantz, 2020), these models are suited to archaeological samples dating from medieval times and more recently, as in the case of our application sample. However, the potential for obtaining reliable results for archaeological individuals from more ancient populations should not be excluded, even though this has not yet been attempted.
32Regarding the interpretation of the results obtained with this sexual estimation method, some rules that need to be considered are detailed below. Firstly, as already mentioned, the 0.70 threshold is recommended to obtain the best results. Interpretations of the results obtained should then observe the following:
33- when the target individual presents a cranial base that is optimally preserved, all thirteen predictive models can be used, although we recommend using the four models that perform best (AD1, AD9, AD10 and AD13). To conclude as to the final attributed biological sex, the majority principle must be applied: if the great majority of the models gives the sex as female, then the anthropologist can conclude that the individual is a female according to the morphometry of the cranial base, with a posterior probability varying from x to x (e.g.: from 0.75 to 0.92). In cases where most of the models attribute the sex as female or male but the highest posterior probability is reached by another model that contradicts this result, the sex finally attributed must be transparently justified by the anthropologist;
- in the case of a fragmented individual, any applicable predictive model may be used. However, the majority principle must similarly be applied to conclude as to the sex finally attributed. In cases where an equal number of models gives the individual’s sex as female, male and indeterminate (e.g.: scenario 1 with 2 models pointing to female, 2 to male and 2 to indeterminacy; scenario 2 with 3 models pointing to female and 3 to male), we advise concluding as to the final sex by taking the result reached with the highest posterior probability. Thus, in the latter case, if model AD6 points to male with the highest accuracy of 0.82, then the anthropologist should conclude that the final attributed sex is male.
34This study unravels the modes of expression of the sexual dimorphism of the cranial base in adult individuals. On the occipital bone, a lower degree of sexual dimorphism was found, with lower accuracies obtained with models AD2, AD3 and AD4 (from 77.4 to 80.6%). This has been found in previous studies on European individuals using a threshold of 0.50 (Gapert, 2009; Macaluso, 2011; Amores-Ampuero, 2017; Chovalopoulou and Bertsatos, 2017), which systematically explained this result by referring to the early maturation of the occipital bone. This occurs at around 8 years of age in both females and males, for two main reasons: to ensure completion of the neural network and the effectiveness of its functional role as an architectural interface between the neurocranium and the face (Humphrey, 1998; Liebermann et al. 2000; Bulygina et al., 2006).
35Conversely, the predictive models AD5, AD6 and AD7 show higher accuracy percentages (83.7 to 84.7%), thus confirming that the temporal bones express sexual dimorphism more strongly. This echoes similar results found in European assemblages (Guyomarc’h and Bruzek, 2010; Say-Liang-Fat, 2014; Milella et al., 2021). These sexual variations can be explained by the later maturation of this specific bone, especially around the mastoid processes, during the pubertal growth phase influenced by strong hormonal activity (Eby and Nadol, 1986; Dahm et al., 1993; Cinamon, 2009). Moreover, mastoid processes are superstructures to which several muscles involved in maintaining head flexion, extension, rotation and balance are attached (Franklin et al., 2006; Guyomarc’h and Bruzek, 2010). Due to differences in body composition (e.g. fat and muscle mass distribution) and energy requirements, the impacts of biomechanical forces on muscle attachment sites such as mastoid processes are likely to differ between males and females (Rosas and Bastir, 2002; Franklin et al., 2006). These findings have been confirmed by a geometric morphometric study developed with a sub-sample of this western European metapopulation (Boucherie et al., 2022).
36The aim of this study was to set up an alternative sexing method suited to fragmentary adult remains. The design process raised questions about what needs to be considered to ensure that a sexing method is reproducible, reliable and robust. Different requirements in terms of osteological assemblages, protocols and statistics were carefully observed and highlighted:
37- the sample needs to be exclusively composed of identified adult individuals to avoid any circular reasoning. It has to be broad enough for statistical robusticity and comprise different age classes for optimum representativeness;
- to guarantee reproducibility and fast implementation, an exclusively metric approach was applied. Measurements were clearly defined, checked for reproducibility, and collected using a non-sophisticated tool;
- to ensure reliability, sample specificity was avoided by including an independent validation sample;
- to avoid dichotomous results (either female or male), the biological reality of indeterminacy was taken into account by using decision thresholds other than 0.50, to prevent any underestimation of possible overlaps between measurements for males and females;
- transparency in the final sex attribution was facilitated by providing the exact posterior probability obtained, which is directly accessible when using logistic regression formulae.
38These requirements are fundamental and need to be systematically observed when designing a new sexing method, in order to guarantee that anthropologists can be confident in the result obtained. Such efforts are especially important in forensics, where rigorous protocols are required and interpretations of results and the limitations of methods are expected to be transparently explained (Daubert versus Meerrell, 1993; Grivas and Komar, 2008; De Boer et al., 2018).
39In the literature about sex estimation, the last few years have seen increasingly frequent discussions about the decision-making process in relation to the threshold used in statistics (Jerković et al., 2018; Santos et al., 2019; Bartholdy et al., 2020; Avent et al., 2021). It appears that a new paradigm is needed when setting up a new sexing method: it is crucial to consider indeterminacy as it represents a biological reality – since different degrees of skeletal sexualisation do exist among individuals (Bruzek and Murail, 2006; Garofalo and Garvin, 2020) – and therefore cannot be ignored or underestimated. Consequently, as for this study, the use of a decision threshold other than 0.50 in statistics protocols is compelling (Galeta and Bruzek, 2020).
40This study underlines the need to set up the rigorous and standardised analytical framework that would be expected for any future development of a sexual diagnosis method. This point is becoming critical in a context where worldwide efforts to implement accreditation and standards in forensics are under way.
41BASE is a new sexual diagnosis tool that has been specifically designed for fragmentary adult remains. The thirteen predictive models (AD1-AD13) developed are fast and easy to use and available free of charge for sexing adult individuals in archaeological or forensic contexts, without requiring access to expensive or sophisticated equipment. To help with recording measurements, users can refer to the ‘Instructions’ tab that provides clear definitions and helpful figures. Additionally, a registration form that can be completed either directly in the field or in the lab while taking measurements may be found in the Supplementary Information (table S6).
42The main advantage of these predictive models lies in their suitability for fragmentary skeletons that present a complete or even a damaged cranial base. Thanks to the various combinations of measurements used, the models can be applied according to the state of preservation of the occipital and temporal bones, whether completely or partially preserved. This is a considerable advantage compared to other software such as FORDISC 3.1, which require several measurements taken from whole preserved crania to obtain a sex identification (Manthey and Jantz, 2020).
43These predictive models can diagnose the biological sex of an individual with an accuracy of 77.4% to 86.8%, using a 0.70 threshold, after cross-validation and with error rate and sex bias below 22% and 16%. These accuracies certainly cannot compete with those obtained with DSP (95% of reliability with a threshold of 0.95, error rate and sex bias below 5% and 0.8% – Murail et al., 2005; Bruzek et al., 2017; Santos et al., 2020) or MorphoPASSE (Klales, 2020b) from the os coxae, as this latter is the most dimorphic bone of the skeleton. Nevertheless, this classification performance appears to be appropriately satisfactory for an alternative primary sexual diagnosis tool designed for situations where the os coxae is unusable. Moreover, the list of reproducible measurements of the cranial base that was developed in this study can now also be used with confidence by anthropologists who are willing to apply a secondary sexual diagnosis to a funerary assemblage (Murail et al., 1999; Santos, 2021).
44The BASE package is dedicated to fragmentary adult remains recovered from archaeological or forensic contexts. Its prospects are promising in both bioarchaeology and forensic anthropology. The authors would welcome any feedback, suggestions or comments on the technical aspects of the package to ensure its continuous improvement.
45It should be pointed out that the applicability of the predictive models is limited to some extent by the fact that they were built up with measurements with specified laterality. Paired t-tests performed beforehand (table S1) showed significant asymmetry for most of the measurements, so that all of them (left and right) were included in the design of the predictive models. For now, we recommend using the predictive models with the measurements that have been specifically defined. In some situations, this approach might prevent users from applying a model to a target individual due to the absence of the specified measurement. The equivalence of the right and left measurements, and their impact on the final sex attribution, need to be tested as a priority in the future to ensure that anthropologists can use these predictive models with confidence with the counterpart measurement.
46Also to be noted is that this package was developed with a large identified western European assemblage. Further investigations are needed to assess its validity for non-European populations. Different patterns of expression of the sexual dimorphism of the cranial base are expected to be found among other populations, which will necessarily imply modulations of the predictive models discussed in this paper.
47For now, the current package facilitates the use of predictive models that were initially built up through logistic regression. Further work is needed to determine whether higher accuracies could be reached with other statistical tools or machine-learning algorithms (Santos et al., 2019; Klales, 2020a).
48This paper presents a new sexing estimation tool – the BASE R package – that was designed according to a rigorous statistical protocol to be fully suited to sexing fragmentary adult remains from their cranial base measurements. Combining occipital and temporal measurements for the first time, this tool offers new opportunities for anthropologists, working in either bioarcheological or forensic contexts, to assess the biological sex of a skeleton whose cranial base is preserved but whose os coxae is unusable. The thirteen predictive models (AD1-AD13) that were built up can achieve up to 85.1-86.8% accuracy. Following the instructions provided in the BASE R package, these predictive models can be easily and rapidly applied to a target individual by any anthropologist, using only a sliding calliper. To obtain reliable results, these models are recommended to be used only for western European individuals, with a decision threshold of 0.70 and following the majority principle detailed in this paper. This new and reliable tool will enable anthropologists to undertake sexual diagnoses for larger numbers of individuals, even when their skeletons are fragmentary and damaged. As more biological data is collected, a better understanding of living conditions, gender and the funerary practices of past populations can be achieved (Boucherie, 2020). In an international context where the number of missing persons is growing due to armed conflicts and migration crises, this new sexing tool, which is fast, easy to use and available free of charge could make a valuable contribution to expanding the range of methods available for forensic investigations. Any future feedback from practitioners will be most welcome to improve the BASE R package and adjust it to the reality in the field.
Table S1
Results of the paired t-tests tests carried out on the bilateral measurements of the cranial base in the adult training sample (n=437) (SD=standard deviation, redundant measurement in italics). P-values<0.05 are in bold. Measurements are in mm.
Table S2
Results of the Welch tests carried out by sex on the cranial base measurements in the adult training sample (n=437) (SD=standard deviation, redundant measurement in italics). P-values<0.05 are in bold. Measurements are in mm or mm2 (*)
Table S3
Correct classification, sex bias, indeterminacy and error percentages found for each predictive model set up from the adult cranial base in the training sample after LOOCV and in the validation sample, with a 0.95 threshold
Table S4
Correct classification, sex bias and percentages of indeterminacy and error found for each predictive model set up from the adult cranial base in the training sample after LOOCV and in the validation sample, with a 0.80 threshold
Table S5
Results obtained when applying the thirteen predictive models set up from the adult cranial base in the application sample made up of two archaeological samples (Coxyde site - 12th-15th century; Koekelberg site - 19th century) and a forensic sample (Bois du Cazier identification mission, see Polet et al., 2024). NA refers to the individuals for whom measurements were not available, so that the predictive models are not applicable
Table S6
Acknowledgements: We would like to thank the different curators who gave us access to the identified osteological assemblages under their care: Dr. P. Semal (Royal Belgian Institute of Natural Sciences - Châtelet, Schoten & Coxyde collections); Pr. M. Besse and Dr. J. Desideri (University of Geneva - Simon collection); Dr. S. Wasterlain (University of Coimbra - Coimbra collection); Dr. S. Garcia (MUNHAC - Lisbon collection); Dr. M. Friess (Muséum National d’Histoire Naturelle, Paris - Olivier collection); Pr. G. Quatrehomme (Côte d’Azur University - Nice collection); Dr. Ann Degraeve and Dr. Katrien Van de Vijver (Urban Brussels - Koekelberg collection); Mrs Colette Ista (Bois du Cazier). Recommendations on statistical procedures were kindly provided by Dr. Ph. Collart (BIOPS, ULB). This paper presents the main results of a doctoral project conducted at the Université libre de Bruxelles and funded by a teaching and research assistant contract within the CReA-Patrimoine (ULB) and by short mission grants provided by the Fonds de la Recherche Scientifique (FNRS) and the Fédération Wallonie-Bruxelles (FWB). We thank the anonymous reviewers for their comments that contributed to the improvement of the paper.