Two statistical criteria to choose the method for dilution correction in metabolomic urine measurements

Introduction

Different normalization methods are available for urinary data. However, it is unclear which method performs best in minimizing error variance on a certain data-set as no generally applicable empirical criteria have been established so far.

Objectives

The main aim of this study was to develop an applicable and formally correct algorithm to decide on the normalization method without using phenotypic information.

Methods

We proved mathematically for two classical measurement error models that the optimal normalization method generates the highest correlation between the normalized urinary metabolite concentrations and its blood concentrations or, respectively, its raw urinary concentrations. We then applied the two criteria to the urinary ¹H-NMR measured metabolomic data from the Study of Health in Pomerania (SHIP-0; n?=?4068) under different normalization approaches and compared the results with in silico experiments to explore the effects of inflated error variance in the dilution estimation.

Results

In SHIP-0, we demonstrated consistently that probabilistic quotient normalization based on aligned spectra outperforms all other tested normalization methods. Creatinine normalization performed worst, while for unaligned data integral normalization seemed to most reasonable. The simulated and the actual data were in line with the theoretical modeling, underlining the general validity of the proposed criteria.

Conclusions

The problem of choosing the best normalization procedure for a certain data-set can be solved empirically. Thus, we recommend applying different normalization procedures to the data and comparing their performances via the statistical methodology explicated in this work. On the basis of classical measurement error models, the proposed algorithm will find the optimal normalization method.