Testing the cross-applicability of juvenile sex estimation from diaphyseal dimensions

Abstract

Sex estimation is a crucial component of the biological profile. Stull et al. (2017) have proposed a promising juvenile sex estimation method using long bone measurements taken from a South African sample, providing relatively high classification accuracies and made easy to use via the KidStats web-based app. In this study, we test the models developed by Stull et al. (2017) on an external historic population from Lisbon, Portugal, in order to determine whether the models can be reliably applied to archeological and forensic populations outside of the original population sample. The study sample consisted of 102 individuals (45 females and 57 males) aged under 13 years at death from the Lisbon identified skeletal collection. Measurements from these individuals were used to test the flexible discriminant analysis (FDA) models given by Stull et al. (2017). Allocation accuracies were calculated for boys and girls and children over and under 2 years separately and combined. Our findings show that the models developed by Stull et al. (2017) yield poor accuracy when applied to our external population and thus can potentially be misapplied on archeological skeletal remains or forensic remains of unknown origin. A number of statistical issues may explain why models fail to be transportable or even generalizable, namely multicollinearity, model overfitting and overly optimist bootstrapped cross-validation rates. It is also likely that population differences in size and sexual size dimorphism also affected the applicability of the models. We emphasize the importance of externally validating prediction models, particularly if they are intended to be applied across populations. Our study addresses Stull and co-worker’s request for further validation of the method on populations outside of South Africa, as the models cannot be confidently applied in the field until it has been externally validated.