Testing of the effect of missing data estimation and distribution in morphometric multivariate data analyses.

Area of Science:

• * Paleontology and Evolutionary Biology
• * Quantitative Biology and Biostatistics

Background:

• * Missing data are a pervasive issue in biological datasets, particularly in morphometrics.
• * The performance of various missing data imputation and deletion techniques is not well understood for these data types.

Purpose of the Study:

• * To evaluate the error introduced by different missing data handling techniques in morphometric datasets.
• * To compare random, anatomically biased, and taxonomically biased missing data introduction methods.

Main Methods:

• * Analysis of a large dataset of extant crocodilian skulls using principal component analysis (PCA).
• * Introduction of 23 different proportions of missing data using three distinct methodologies.
• * Comparison of results using Procrustes superimposition to quantify introduced error.

Main Results:

• * Gower's distance was the top-performing non-estimation method; Bayesian PCA excelled in estimation.
• * Higher estimation error was observed in taxa with small sample sizes or significant morphological differences.
• * The distribution of missing data significantly impacted estimation error across most methods.

Conclusions:

• * Missing data imputation methods perform differently based on data distribution (random, anatomical, or taxonomic bias).
• * Anatomically biased missing data introduced greater deviation than random or taxonomic biases.
• * Understanding missing data patterns is crucial for accurate morphometric analyses.