Imputation of missing values in lipidomic datasets.

Area of Science:

• * Lipidomics
• * Bioinformatics
• * Statistical analysis

Background:

• * Lipidomic datasets frequently contain missing values, impacting statistical analyses.
• * Missing data can be categorized as missing completely at random (MCAR), missing at random, or missing not at random (MNAR).
• * Imputation techniques are crucial for handling missing data to enable statistical methods and enhance effect identification.

Purpose of the Study:

• * To evaluate the performance of various imputation methods for lipidomic data.
• * To compare common methods (zero, half-minimum, mean, median) with advanced techniques (k-nearest neighbor, random forest).
• * To assess imputation effectiveness using simulations and real shotgun lipidomics datasets.

Main Methods:

• * Investigated zero, half-minimum, mean, and median imputation.
• * Assessed k-nearest neighbor (KNN) and random forest imputation.
• * Utilized simulation studies and real shotgun lipidomics data analysis.
• * Focused on correlation-based and truncated normal distribution-based KNN methods.

Main Results:

• * Shotgun lipidomics data are characterized by high correlations and missing values, often MNAR due to low analyte abundance.
• * Correlation-based and truncated normal distribution-based KNN imputation methods showed the best performance.
• * These advanced KNN methods effectively imputed missing values regardless of the missingness type.
• * Imputation methods successfully controlled the type I error rate in statistical comparisons.

Conclusions:

• * K-nearest neighbor imputation, particularly correlation-based and truncated normal distribution-based approaches, is highly effective for missing data in lipidomics.
• * These methods perform well even when missingness is not at random (MNAR), a common scenario in lipidomics.
• * The chosen imputation techniques maintain statistical rigor by controlling the type I error rate.