Using multivariate adaptive regression splines to estimate subadult age from diaphyseal dimensions

Abstract

Subadult age estimation is considered the most accurate parameter estimated in a subadult biological profile, even though the methods are deficient and the samples from which they are based are inappropriate. The current study addresses the problems that plague subadult age estimation and creates age estimation models from diaphyseal dimensions of modern children. The sample included 1,310 males and females between the ages of birth and 12 years. Eighteen diaphyseal length and breadth measurements were obtained from Lodox Statscan radiographic images generated at two institutions in Cape Town, South Africa between 2007 and 2012. Univariate and multivariate age estimation models were created using multivariate adaptive regression splines (MARS). K-fold cross-validated 95% prediction intervals (PIs) were created for each model and the precision of each model was assessed.The diaphyseal length models generated the narrowest PIs (two months to six years) for all univariate models. The majority of multivariate models had PIs that ranged from three months to five and six years. Mean bias approximated zero for each model, but most models lost precision after 10 years of age. While univariate diaphyseal length models are recommended for younger children, multivariate models are recommended for older children where the inclusion of more variables minimized the size of the prediction intervals. If diaphyseal lengths are not available, multivariate breadth models are recommended. The present study provides applicable age estimation formulae and explores the advantages and disadvantages of different subadult age estimation models using diaphyseal dimensions.