Age-at-death estimation based on algorithms for multi-class discriminant analysis using binary and ordinal predictors

Area of Science:

• Statistical modeling
• Forensic anthropology
• Biostatistics

Background:

• Accurate age-at-death estimation is crucial in forensic science and anthropology.
• Developing robust statistical methods for age estimation using demographic data is an ongoing challenge.
• Binary and ordinal predictors offer a flexible approach to modeling age-related data.

Purpose of the Study:

• To develop and evaluate multi-class discriminant analysis algorithms for age-at-death estimation.
• To assess the performance of algorithms using binary and/or ordinal predictors.
• To determine the effectiveness of these algorithms in scenarios with correlated predictors.

Main Methods:

• Examined algorithms based on uncorrelated predictors (crude assumption) and latent models assuming multivariate normal distribution.
• Utilized tetrachoric/polychoric correlations estimated from training data or generated from correlated data algorithms.
• Applied algorithms to age-at-death estimation datasets, analyzing classification accuracy.

Main Results:

• Crude algorithms showed poor performance with highly intercorrelated ordinal predictors.
• Latent model-based algorithms demonstrated similar and satisfactory classification performance.
• Overall classification accuracy was good, even with small sample sizes, but cross-validated accuracy was low for small samples.

Conclusions:

• Latent model algorithms are suitable for age-at-death estimation; crude algorithms should be avoided with correlated ordinal data.
• Increasing class size significantly enhances classification accuracy.
• High cross-validated accuracies (over 90%) are achievable with larger class sizes.