Multivariate paired data analysis: multilevel PLSDA versus OPLSDA.

Area of Science:

• Nutritional metabolomics
• Multivariate data analysis
• Experimental design

Background:

• Metabolomics data from nutritional studies often exhibit complex structures due to experimental designs.
• Cross-over designs, where subjects serve as their own controls, create paired data, complicating multivariate analysis.
• Standard methods like OPLSDA (Orthogonal Partial Least Squares Discriminant Analysis) may ignore this paired structure.

Purpose of the Study:

• To compare multilevel PLSDA with OPLSDA for analyzing cross-over metabolomics data.
• To evaluate the impact of exploiting paired data structure on analytical power and interpretability.
• To assess the additional information provided by multilevel approaches regarding treatment effects and subject variability.

Main Methods:

• Comparison of multilevel PLSDA and OPLSDA.
• Analysis of simulated and genuine cross-over nutritional metabolomics data.
• Evaluation of statistical power and interpretability of results.

Main Results:

• Multilevel PLSDA, which exploits the paired data structure, significantly improves power and interpretability over OPLSDA.
• The multilevel approach reveals diversity and abundance of treatment effects across subjects.
• Intrinsic differences between study subjects are also elucidated by the multilevel method.

Conclusions:

• Exploiting the paired data structure in cross-over designs is crucial for robust metabolomics analysis.
• Multilevel PLSDA offers a more powerful and interpretable approach for such data.
• This method provides deeper insights into treatment effects and inter-subject variability in nutritional studies.