Improving adult age-at-death estimates using visible features of the human skeleton has been the subject of much research because such assessments are critical elements of forensic investigations and bioarchaeological studies. Beginning in 2014, a National Institute of Justice (NIJ)–funded research team developed a set of age-informative skeletal traits; collected reference data from collections in the United States, Portugal, Thailand, and South Africa; evaluated traits for their applicability; and developed alternative ways to generate age estimates from those traits. Here we present a comparison of two ways to produce age estimates: Stephen D. Ousley’s machine learning approach available as beta version computer software (TA3-ML) and an analytical procedure that originated with the version of transition analysis introduced two decades ago with different skeletal characteristics (TA3-TA). The two approaches are evaluated using the same 41 modern Portuguese and American skeletons. Both methods rely on NIJ-project skeletal data (TA3), but the number of traits used differs, as do reference sample sizes and compositions. Estimates generated through TA3-TA more closely approximate reported ages throughout adulthood than those from TA3-ML. Nevertheless, there remains a problem with underestimation in the TA3-TA approach, and neither method is ready for widespread implementation. Ongoing work is being directed toward resolving these issues by adjusting the mix of NIJ-project traits used in TA3-TA.
Read full abstract