Evaluation of missing data imputation methods for human osteometric measurements

Area of Science:

• Bioarcheology
• Forensic Anthropology
• Quantitative Biology

Background:

• Bioarcheological and forensic analyses often involve incomplete skeletal specimens.
• Many multivariate statistical methods require complete data, necessitating missing data imputation.

Purpose of the Study:

• To evaluate the performance of popular statistical methods for imputing missing metric measurements in skeletal data.
• To identify accurate, robust, and efficient imputation techniques for bioarcheological and forensic applications.

Main Methods:

• Utilized William W. Howells' Craniometric Data Set and the Goldman Osteometric Data Set.
• Compared performance of single imputation methods (e.g., Bayesian Principal Component Analysis - BPCA) against multiple imputation methods.
• Evaluated methods including Bayesian linear regression (norm2), Expectation-Maximization (EM) with Bootstrapping (Amelia), and Predictive Mean Matching (PMM) (mice).

Main Results:

• Multiple imputation methods significantly outperformed single imputation methods.
• Bayesian linear regression (norm2), EM with Bootstrapping (Amelia), and PMM (mice) demonstrated strong performance in accuracy, robustness, and speed.
• Bayesian Principal Component Analysis (BPCA) was identified as a less effective single imputation method.

Conclusions:

• Multiple imputation techniques are recommended for handling missing data in skeletal analyses.
• A practical procedure for selecting appropriate imputation methods based on study findings is proposed.
• The study provides guidance for biological anthropologists working with incomplete skeletal data.