Previous research has ascribed misclassification in sex estimation using the Walker [3] traits as the result of changes that occur with advancing age [9, 11,12,13,14,15,16,17]. The current study explored potential trends in the trait scores that may be attributable to age and assessed how age can affect the positive predictive performance of the traits when estimating sex.

Overall, very few statistically significant differences were observed in the trait scores among the age cohorts. More specifically, the nuchal crest was the only trait to demonstrate any significant differences in the pooled sample, where individuals younger than forty years of age were noted to differ from individuals older than forty years of age. The nuchal crest is a major attachment site for structures of the head and neck, including the nuchal ligament, and the trapezius, semispinalis capitis, and splenius capitis muscles. De La Paz and colleagues [37] analysed the muscle structures and attachment sites of the head and neck for sex, population, and age-related differences. The study focused on the anatomical relation of the muscle and subsequent attachment sites to the morphology of the nuchal crest [37]. Typically, nuchal crests that are considered to be more robust (i.e., that would be given a score of 5) include both a prominent, often hooked nuchal crest paired with rough nuchal lines. In their results, De La Paz and colleagues [37] did not find significant sex or age-related differences in the muscle attachment site itself; however, the authors did identify variation in the types of attachments. The results demonstrated the variable attachment of the nuchal ligament and trapezius muscle, as a greater percentage of individuals exhibit no proximal cranial attachment by the superior portion of the trapezius muscle. Thus, the appearance of the nuchal crest as employed by forensic anthropologists may be influenced by the type of muscle attachment so that the nuchal area appears more robust. The current study also identified a significant, but weak, correlation between edentulism and the nuchal crest. The significant correlation between antemortem tooth loss and the nuchal crest may have been a factor in the significant differences identified between the age cohorts for the trait. Edentulism has a direct resultant effect on the masticatory function of the mandible and maxilla and therefore the muscles and skeletal structures related to this function and other functions of the head and neck may be affected [12, 18]. Factors that affect the muscle may directly affect the attachment site of the muscle to the bone and thus it is imperative that further studies are conducted to identify the effect of the correlation between edentulism and the nuchal crest as it was beyond the scope of the current study.

This apparent robusticity is not necessarily the result of sexual dimorphism and does not appear to be the result of aging. Indeed, the fact that the nuchal crest was noted to differ significantly between the population groups need to be considered. With the current sample having a skewed distribution with a greater proportion of young black South Africans compared to a greater proportion of older white South Africans, the apparent robusticity in the older cohorts is likely population variation. As is consistent with previous South African studies, white South Africans (both male and female) are on average more robust than black and coloured South Africans [4, 7], and this observation could explain why the nuchal crest of the older cohorts appear more robust in the current study. One key difference between the current findings with those in previous South African studies, is that Krüger and colleagues [4, 7] observed significant population differences for all traits rather than just the nuchal crest.

When the sample is divided so that the sexes are assessed separately the results indicated that the males did not demonstrate any significant differences for the nuchal crest, and all significant differences were observed in the female sample. More specifically, younger females (between 20 and 39 years) differed from the older females (between 60 and 79 years), while females that fell within the middle-aged cohort did not differ significantly from either the younger or older cohorts. Thus, the significant differences noted for the nuchal crest in the pooled sample was most likely due to significant differences between the cohorts of female individuals, and thereby emphasises the importance of exploring the sexes separately to more effectively identify age-related trends in craniofacial changes as each sex appears to be undergoing changes differently. Changes in the nuchal crest are assumed to be the result of continued and increasing strain of the nuchal ligament and trapezius muscle throughout an individual’s lifespan. Additionally, any stress experienced by the vertebral column as the result of morphological changes, such as vertebral wedging that may occur with osteoporosis or general vertebral body compression that occurs with age, would also have an effect on the nuchal ligament as it is a superior and posterior extension of the supraspinous ligament [4, 7, 38, 39].

Sex differences in musculoskeletal markers are typically attributed to differences in activity patterns or sexual division of labour, with males frequently engaging in harder physical labour resulting in increased hypertrophy of muscles and muscle attachment sites [40, 41] noted that both sex and age play a role in the robusticity of muscle markers of the humerus, where females demonstrated slighter entheseal changes compared to their male counterparts. Yet the results from the current study did not follow this trend with the nuchal crest. It should be acknowledged that the majority of studies that have explored musculoskeletal activity markers and age have looked at postcranial skeletal elements, and changes on muscle markers of the cranium are incompletely described in the literature. However, further factors that should be considered to better understand age-related changes to the nuchal crest (in addition to sexual dimorphism) include body size, muscle size, and muscle attachment size [37, 40, 41] also argues that aggregate muscle marker analyses (i.e., from more than one muscle attachment site) be conducted to gain a better understanding of the effects that the above factors have on the muscle attachments sites of individuals. Additionally, it should be acknowledged that despite the mastoid process also being a major muscle attachment site, this region demonstrated no significant associations with age or population affinity in the current study.

In addition to the nuchal crest, the females also demonstrated some significant differences (between the 20-year-old and 80-year-old cohorts) for the supra-orbital margin when assessing the sexes separately. Once again, the males did not present with any differences, nor were the differences observed with the pooled sample. Many other studies have analysed age changes to the orbit; however, these studies mainly focused on metric changes. Kahn and Shaw [42] noted significant changes with age to the orbital width and area in both males and females, and suggested that dramatic changes occur to the entire bony orbit throughout an individual’s life. Özer and colleagues [43] found slight correlations between changes in the width and height of the orbital aperture and increasing age in females but not in males. Additionally, Ugradar and Lambros [44] identified a relationship between increasing orbital volume and increasing age in females. However, as the current study only identifies significant differences between two of the cohorts there is a distinct possibility that it is an artefact from the sample. Therefore, it is imperative that further studies be conducted to better explore sex differences in the aging of the orbit, especially with regards to morphoscopic methods.

Another aspect that should be considered is secular trends. While it was not within the scope of the study to explore secular trends, it should be considered that secular changes in the cranium may affect the traits. Godde [45] studied individuals from the Hamann-Todd and William M. Bass Skeletal Collections and identified the presence of some secular trends in the cranial morphoscopic traits in North American males and females from 1849 to 1960. Jantz and Jantz [46] and Grine et al. [47] identified metric secular trends that affect the cranial morphology in 19th to 20th century North Americans and South Africans, respectively. Unfortunately, to date the majority of the literature focuses on metric changes and there has been little to no focus on whether (and to what extent) the morphoscopic traits are affected in this regard. Further research should be conducted to determine if secular trends are present in the morphoscopic traits and whether such trends are occurring in the South African population.

The results of the current study demonstrated only a few significant differences among the age cohorts; however, this does not negate the existence of age-related differences. It is likely that the morphoscopic methods may be unable to effectively quantify the small differences caused by age-related changes between the cohorts. For example, it is unlikely that a female’s score would change so drastically (such as from a score of 1 to a score of 3) in one lifespan. The shift from a score of a very gracile 1 to a score of an intermediate 3 would be radical and would require a large amount stress and/or remodelling for the bone to alter so significantly. As was noted in previous literature that was carried out on different populations [9, 11,12,13,14,15,16], significant differences were present between the age cohorts, but were slight and did not have a considerable effect on the sex estimate (combined accuracy of 71.90%; <40 cohort accuracy of 69.10%; and > 40 cohort accuracy of 81.30%) [3, 16]. Thus, to account for the more subtle changes, metric methods and geometric morphometrics may be better able to quantify the more nuanced changes to the craniofacial skeleton that occur due to the aging process.

Despite limited significant differences, the sample was subdivided into two broad groups (younger versus older) to see if there would be any difference in classification accuracy that could possibly justify prior knowledge of age to estimate sex. The results from the classification models indicated a slight increase in accuracy in the younger sample, which showed that the younger age cohorts classify more accurately when separated from the older age cohorts. The increase in accuracy in the younger sample is largely due to a substantial increase in the accuracy of the younger males, with the younger females only showing a slight improvement. When the age cohort was preselected in the LR model, the accuracy of the younger males substantially increased which is most likely due to the removal of the more “masculine-presenting” older females. Interestingly, the preselection of age in the LR model resulted in the inversion of the sex bias indicating that the males (in both the younger and older sample) classify better when the respective younger/older females are not included in the model. Similarly, the preselection of age in the RFM resulted in a substantial increase in the accuracy of the younger males. Contrastingly, the RFM age-preselection resulted in a decrease in the accuracy of the older sample males and a slight increase in accuracy of the older sample females. The RFMs more clearly demonstrate that the individual accuracies improve when the younger more “feminine-presenting” males are separated from the older more “masculine-presenting” females. However, the overall overlap in trait appearance for the more “feminine-presenting” younger group and the more “masculine-presenting” older group results in decreased RFM accuracies for the younger females and older males. It is essential to note that the differences between the models and related predictive performance may be due to model/sample limitations. Nevertheless, these results are consistent with the literature that states that females become more robust with age [23] resulting in a decrease in sexual dimorphism where their traits appear more masculine. However, the changes are slight and given the amount of overlap that naturally occurs with the traits, the accuracy is not greatly affected when age-specific standards are used.

Krogman and İscan [24] stated that, due to the effects of age on the skeleton, sex should only be estimated from the cranium for individuals between the ages of twenty and fifty-five years due to the increasing robusticity of females in older age cohorts. However, the classification accuracy observed with the older cohort classification model contests this, as the accuracy was not substantially lower than either the pooled or younger models. Thus, morphoscopic sex estimation can be performed with high accuracy in the South African population regardless of age. This supports other studies, such as that completed by Garvin and colleagues [10] and the study by Lesciotto and Doershuk [16], which stated that even though they found significant associations between age and the Walker [3] traits, the correlation was weak and did not influence the accuracy of the sex estimate or the traits enough to validate the creation of new standards for medicolegal contexts.

Finally, when assessing the performance of the two statistical methods—LR and RFM—both methods yielded similar overall accuracies for the pooled sample. However, closer scrutiny revealed that RFM exhibited a much greater sex bias with the age-specific samples. In the younger sample, males performed much better than females, while the older sample demonstrated the opposite trend, resulting in significant sex bias. Sex bias can introduce systematic error, where one sex is overestimated, which is undesirable for classification models. One potential reason for this is the weighting of traits and variable importance. The younger model ranked the supra-orbital margin as the most discriminatory variable, while the older model and pooled model placed more importance on the glabella. This similarity between the pooled and older sample results could be due to the older age group constituting majority of the pooled sample (N = 324/453), masking the contribution of the younger individuals in the pooled sample. Although previous studies have reported higher accuracies for RFM over LR [7], the findings of the current study show that the performance and choice of model depend heavily on dataset characteristics. When dividing the sample into age-specific subgroups, an imbalance in factors such as sample size and population affinity likely affected model performance and led to overfitting [34]. Therefore, it is crucial to carefully select samples to avoid overfitting and ensure optimal performance. The results also indicate that overall classification accuracy is not always the best measure of model performance. Other metrics, such as sex bias, sensitivity, specificity, and probability, should also be considered when developing classification models and standards.