Empirical evaluation of data normalization methods for molecular classification

Background Data artifacts due to variations in experimental handling are ubiquitous in microarray studies, and they can lead to biased and irreproducible findings. A popular approach to correct for such artifacts is through post hoc data adjustment such as data normalization. Statistical methods for data normalization have been developed and evaluated primarily for the discovery of individual molecular biomarkers. Their performance has rarely been studied for the development of multi-marker molecular classifiers—an increasingly important application of microarrays in the era of personalized medicine. Methods In this study, we set out to evaluate the performance of three commonly used methods for data normalization in the context of molecular classification, using extensive simulations based on re-sampling from a unique pair of microRNA microarray datasets for the same set of samples. The data and code for our simulations are freely available as R packages at GitHub. Results In the presence of confounding handling effects, all three normalization methods tended to improve the accuracy of the classifier when evaluated in an independent test data. The level of improvement and the relative performance among the normalization methods depended on the relative level of molecular signal, the distributional pattern of handling effects (e.g., location shift vs scale change), and the statistical method used for building the classifier. In addition, cross-validation was associated with biased estimation of classification accuracy in the over-optimistic direction for all three normalization methods. Conclusion Normalization may improve the accuracy of molecular classification for data with confounding handling effects; however, it cannot circumvent the over-optimistic findings associated with cross-validation for assessing classification accuracy.