NMR spectroscopy provides valuable data for metabolomics, but the information sought can be partly obscured by errors from hardware imperfection, causing frequency, phase, and spectral lineshape to change significantly between measurements. Clearly, this is a highly undesirable source of variation in multivariate quantitative studies such as metabolomics. Fortunately, many hardware imperfections affect all resonances in the same way. They can therefore be corrected for by comparing an experimental reference peak with the known correct peak shape, in a procedure known as reference deconvolution. This post-measurement processing method can correct many systematic errors in data. The aim of this study is to investigate how reference deconvolution can improve the results obtained by multivariate analysis of NMR data. For this purpose, 1H NMR data were recorded for a set of 136 mixture samples. Spectra were then produced with and without reference deconvolution and analyzed by principal component analysis and partial least squares methods. The results showed that reference deconvolution resulted in simpler and improved models, requiring fewer latent variables to explain the same or higher percentage of the variance. It was also evident that the recovery of the design concentrations was significantly enhanced. This confirms that reference deconvolution can significantly improve multivariate data analysis and should be considered as a standard tool in high throughput quantitative NMR spectroscopy. © 2014 John Wiley & Sons, Ltd.