Nonrandom missing data can bias Principal Component Analysis inference of population genetic structure

Area of Science:

• Population genetics
• Genomics
• Bioinformatics

Background:

• Next-generation sequencing (NGS) generates substantial missing data in population genetics studies.
• Principal Component Analysis (PCA) is a common tool for visualizing population structure, often using mean imputation for missing data.

Purpose of the Study:

• To investigate the impact of missing data, particularly nonrandomly distributed missingness, on PCA in population genetic analyses.
• To evaluate how biased missing data affects the interpretation of population structure in both simulated and empirical datasets.

Main Methods:

• Simulated population genetic data with varying levels and patterns of missingness (random vs. biased).
• Empirical data from big brown bat (Eptesicus fuscus) using restriction site-associated DNA sequencing (RADseq) with different missing data filters.
• Principal Component Analysis (PCA) applied to simulated and empirical datasets.

Main Results:

• Biased missing data in individuals shifts them towards the origin in PCA plots, mimicking admixed individuals.
• This effect was observed in both simulated data and empirical RADseq data from big brown bats.
• Nonrandom missing data is common in non-model organisms due to sample quality variations.

Conclusions:

• Missing data, especially when nonrandomly distributed, can significantly bias PCA results and lead to misinterpretation of population structure.
• Researchers should carefully consider and visualize missing data patterns when using PCA in population genetics.
• Recommendations include plotting PCA with missingness gradients, cautious interpretation of central samples, and using complementary analyses for validation.