Shirobokov Ivan Grigorievich

Introduction. There are no fewer than two hundred algorithms for sex estimation based on cranial morphology, relying on statistical analysis of non-metric, linear, angular traits, and their combinations. Nevertheless, many physical anthropologists prefer to rely on visual observations. The objectives of this research encompass exploring potential reasons behind the preference for a visual approach and conducting an analysis of the comparative effectiveness of visual and statistical methods for sex estimation.

Materials and methods. The study is grounded in an analysis of publications related to methods of sex estimation based on cranial traits, spanning the past 70 years. Comparison of accuracy estimates was conducted using non-parametric tests, considering differences in statistical methods, validation approaches (no validation, cross-validation, independent test), and variable types (non-metric traits, craniometry, geometric morphometrics).

Results. General reasons for skepticism towards algorithms include unrealistic expectations regarding their capabilities, greater susceptibility to errors by models compared to humans, lack of control over classification. However, algorithms generally surpass experts in predicting the target variable. The average accuracy of visual sex estimations based on cranial traits is slightly lower than the estimates of statistical models and exhibits noticeable variability. The accuracy of estimations made by experienced anthropologists is comparable to the average performance of models. Nevertheless, the effectiveness of algorithms significantly diminishes when applied to datasets originating from sources other than the training set, particularly when dealing with craniometric traits. In a substantial portion of studies, the size of the training datasets is insufficient for a reliable assessment of model effectiveness, and the sex distribution is skewed towards male skulls, leading to some inflation of the accuracy of their estimates. Model effectiveness can also decline due to errors in the evaluation of non-metric traits, and the assessment of inter-researcher discrepancies does not allow for an evaluation of their impact on model accuracy.

Conclusion. Despite an extensive bibliography, there remains a lack of data on both the accuracy of the visual approach to sex estimation and the reliability of models with claimed high effectiveness. The adoption of flexible methodologies enabling researchers to independently control both variable selection and the composition of the training set will help overcome algorithm aversion and enhance the quality of estimates. © 2023. This work is licensed under a CC BY 4.0 license.