Informative Priors Produce Precision

Informative priors in Bayesian analyses improve estimates of adult age, but effects in developmental age estimation are unknown. Accounting for residual correlations between developing teeth is supported, but how characteristics of the correlation matrix might affect estimates of age is unknown. Bayesian priors and correlation matrices are systematically tested for their effects on dental developmental age estimates. Nine cumulative probit models were fit to training samples from 880 dental score sets of London dental patients: the full sample, half samples split by sex or ancestry (Bangladeshi or European), and quarter samples split by sex and ancestry. A target sample of dental development scores from computed tomography images in the New Mexico Decedent Image Database was randomly divided into validation (n = 381) and test samples (n = 188) before analysis. All age ranges were 3 to 21 years. Four priors were tested: uniform and kernel densities derived from all-cause mortality, homicides, and NamUs missing persons data. Thirty-six rounds of age estimation were performed using nine correlation matrices from training sample splits and four priors. Correlation matrices with lower variability measures produced higher age interval success rates. Informative priors produced larger residual error and narrower age intervals without a corresponding decrease in success rate relative to the uninformative prior. Model correlation matrices can and should be evaluated for suitability before estimating ages. The choice of an informative or uninformative prior depends on practitioner priorities of age interval precision versus point estimate accuracy. The model discussed here is published as an R package.