Skeletal age-at-death estimation: Bayesian versus regression methods Efthymia Nikita a, *, Panos Nikitas b a Science and Technology in Archaeology and Culture Research Center, The Cyprus Institute, 2121 Aglantzia, Nicosia, Cyprus b Department of Chemistry, Aristotle University of Thessaloniki, University Campus, 54124 Thessaloniki, Greece A R T I C L E I N F O Article history: Received 4 September 2018 Received in revised form 14 January 2019 Accepted 24 January 2019 Available online 5 February 2019 Keywords: Forensic anthropology Skeletal age estimation Auricular surface Bayesian statistics Transition analysis Regression analysis A B S T R A C T Age-at-death estimation in a skeletal assemblage (target sample) is biased by the demographic profile of the material used for age prediction (training sample) when this profile is different from that of the target sample. This bias is minimized if the demographic profile of the target sample is properly taken into account in the method developed for age-at-death estimation. In the Bayesian approach this is accomplished via the informative prior. For methods based on regression, we propose two techniques: (a) using weighting factors taken from the demographic profile of the target sample, and (b) creating a new hypothetical training sample that has a demographic profile similar to that of the target sample. The two techniques, as well as the Bayesian approach, were tested using 532 artificial systems in which the age marker exhibited an eight-grade expression. It was found that depending on the criteria used for evaluation, the proposed approaches and especially the one based on a hypothetical training sample, may give better results than the Bayesian method in more than 90% of the systems studied. A basic prerequisite for the good performance of the proposed approaches is to select carefully the training sample. This sample should exhibit a uniform demographic profile or a profile with almost equal numbers of young and older individuals. All the above hold if the training and the target samples have different demographic profiles. If the profiles are the same or very similar, the best aging method is the direct regression using simple linear models. © 2019 Elsevier B.V. All rights reserved. 1. Introduction Age-at-death estimation based on skeletal remains is a key parameter in forensic anthropology and bioarchaeology. In forensic contexts it contributes to the establishment of an individual’s biological profile and the ultimate identification of this individual [1,2]. In bioarchaeological contexts, age-at-death estimation provides key information pertaining to past demogra- phy and constitutes an important covariate in the study of pathological, activity and other skeletal markers of past life quality. Despite its importance, age-at-death estimation for adults is challenging because the skeletal degeneration on which age estimation is based progresses in a nonlinear manner as it is affected by pathology, activity and other factors. Therefore, it becomes progressively more and more disassociated from chronological age [3]. Various methods have been proposed for age estimation in adult remains, most of which are focused on diarthrodial and amphiarthrodial joints (e.g. Refs. [4–6]). A number of studies have stressed the high level of inter-population variation in the ageing process and the need for enhancing current approaches and developing novel ones [7–12]. Most of the methods used for age-at-death prediction have the following general structure. There are two samples, a training and a target one. The training sample should be large enough in order to be representative of the reference population, whereas the target sample may consist of one to several hundreds of individuals. Based on the training sample, the relationship between chronological age-at-death and one or more age markers is established, usually adopting regression or Bayesian approaches. Then this relationship is used to estimate the age-at-death of the individuals in the target sample. This procedure, and especially the use of regression models, was questioned by Bocquet-Appel and Masset [13], who argued that the ages estimated in a target sample are biased by the demographic profile of the training sample. This phenomenon is known as ‘age mimicry’ and Bayesian age estimation has been developed in order to minimize it [14–16]. In particular, the Bayesian method uses in the computations information about the demographic profile of the target sample via the age-at-death distribution function of the informative prior . The question that * Corresponding author at: Science and Technology in Archaeology and Culture Research Center, The Cyprus Institute, 20 Konstantinou Kavafi Street, Aglantzia, Nicosia, 2121, Cyprus E-mail addresses: e.nikita@cyi.ac.cy (E. Nikita), nikitas@chem.auth.gr (P. Nikitas) . https://doi.org/10.1016/j.forsciint.2019.01.033 0379-0738/© 2019 Elsevier B.V. All rights reserved. Forensic Science International 297 (2019) 56–64 Contents lists available at ScienceDirect Forensic Science International journal homepage: www.elsevier.com/locate/forsciint