A real data-driven simulation strategy to select an imputation method for mixed-type trait data

Area of Science:

• Evolutionary Biology
• Bioinformatics
• Comparative Genomics

Background:

• Missing observations in biological trait datasets hinder analyses across various disciplines.
• Existing imputation methods yield mixed results, necessitating a framework for selecting appropriate techniques for diverse, real-world datasets.
• Trait datasets often contain mixed data types (categorical, count, continuous), complicating imputation strategies.

Purpose of the Study:

• To develop and validate a real data-driven simulation strategy for selecting the optimal imputation method for mixed-type trait datasets.
• To evaluate the performance of candidate imputation methods, including mean/mode, k-nearest neighbour, random forests, and MICE, with and without phylogenetic information.

Main Methods:

• A squamate trait dataset was used as a target, with missing data simulated under missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR) mechanisms.
• Imputation was performed using candidate methods, incorporating phylogenetic information from nuclear, mitochondrial, or multigene trees.
• Performance was assessed using mean squared error for numerical traits and proportion falsely classified rates for categorical traits.

Main Results:

• The random forest method, enhanced with a nuclear-derived phylogeny, demonstrated the lowest error rates across most traits.
• Imputed datasets more accurately reflected the original data's characteristics and distributions compared to complete-case datasets.
• Phylogenetic information did not consistently improve performance for all traits or scenarios, highlighting the need for careful method selection.

Conclusions:

• A real data-driven simulation strategy is effective for selecting suitable imputation methods for mixed-type trait datasets.
• Random forests combined with appropriate phylogenetic data offer a robust approach for trait data imputation in evolutionary biology.
• Caution is advised, as the utility of phylogenetic information in imputation varies by trait and missingness mechanism.