Mixed Multivariable Models to Improve Dental Age Estimation in a Worldwide Sample

Thumbnail Image

Authors

Cirillo, Laura

Issue Date

2022

Type

Dissertation

Language

Keywords

Age , Age estimation , Dental , Development , Forensic anthropology , Method

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

Despite popular acknowledgment that dental development is the age estimation method least affected by environment, there has been a persistent call for population-specific methods designed to tease out minute differences for increased precision. Many of these methods fail to provide adequate statistical justification for the need to limit the population and do not use a realistic sample, limiting their application. Because both forensic and bioarchaeological contexts necessitate generalizable methods that can be applied to unknown demographic context, this research explores the impacts of compounding variables and incorporating diversity on precision and accuracy. Development and eruption data was collected on a cross-sectional sample of 2,656 individuals with known chronological age between birth and 26 years old. The sample consisted of individuals from 10 countries (Angola, Australia, Brazil, Colombia, France, Netherlands, Saudi Arabia, Spain, South Africa and the United States). Ordinal data was recorded from a combination of dental radiographs, Lodox scans, and CT scans; continuous data was collected using 3D interlandmark distance on CT samples only. A Mixed Cumulative Probit algorithm was used to create univariate models for each tooth, and a subset of multivariate models for comparison. Training samples from six countries (Brazil, Colombia, France, Saudi Arabia, Spain, South Africa and the United States) were used to develop the age estimation models, each tested on a holdout sample and independent samples from the remaining countries. Two main comparisons were made: 1. Pooled, global model performance vs. all population-specific model performance and 2. Model performance of different variables (i.e., univariate or multivariate, ordinal or continuous). Accuracy, precision, bias, and generalizability were used as measures of performance. Pooled models often matched or out-performed the population-specific models and were the only consistently generalizable option. Multivariate and continuous/mixed models showed promise for increasing the accuracy and precision of the method, but more variations and increased samples are needed for an adequate comparison of performance. Method adjustments and future directions to build on the present work are recommended and discussed.

Description

Citation

Publisher

License

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN