Imputation of Missing Values for Multi-Biospecimen Metabolomics Studies: Bias and Effects on Statistical Validity

Area of Science:

• * Metabolomics
• * Bioinformatics
• * Statistical Analysis

Background:

• * High-throughput metabolomics mass spectrometry data analysis is challenged by missing values.
• * Omitting samples with missing data can cause data loss and biased group comparisons.
• * Imputing missing data may disrupt intra-subject correlation, impacting biomarker discovery.

Purpose of the Study:

• * To investigate imputation strategies for metabolomics data from multiple biospecimens per subject.
• * To compare a novel combined data matrix imputation approach with separate matrix imputation.
• * To evaluate the impact on intra-subject correlation and statistical significance testing.

Main Methods:

• * Compared two imputation strategies: combined vs. separate data matrix imputation.
• * Evaluated five imputation methods: Random Forest, k-NN, Expectation-Maximization, Quantile Regression, Half-Minimum.
• * Assessed bias in intra-subject correlation estimation and statistical test validity.

Main Results:

• * Combining biospecimen data matrices for imputation offered minimal improvement in correlation or statistical accuracy for most methods.
• * Random Forest imputation generally outperformed other methods across performance metrics, excluding specificity.
• * The novel combined approach showed no significant advantage over separate imputation.

Conclusions:

• * Simple imputation strategies, including combining data matrices, may not substantially enhance metabolomics analysis of multi-specimen studies.
• * Random Forest is a promising imputation method for metabolomics data, warranting further investigation.
• * Careful consideration of imputation methods is crucial for accurate biomarker discovery in multi-specimen metabolomics studies.