Methods of multivariate calibration like Multiple Linear Regression (MLR), Principal Component Regression (PCR) and Partial Least Squares (PLS) are typically used to determine chemical composition from NIR spectra. PLS uses a whole range of spectra and it is a method that can be recognized as a standard tool. However, if the predictors come from spectroscopy then the measurements provide hundreds or thousands of channels (variables), and often a considerable part of the variables carries no, or almost no, important information except noise. For this reason, MLR, although conceptually much simpler and not directed to explore connections between the structures of predictor and response sets, can provide results better than PLS. Therefore it is worth to check which of the methods works better in a given application of multivariate calibration. The present contribution is aimed at comparing the results of both methods when applied to the calibration of basic constituents (oil, dry matter, protein, ash, fibre) of rapeseed meal from NIR reflectance spectra. The selection of wavelength needed to apply MLR was made with two procedures: one based on the approach proposed by Brown et al., and another using an approach developed by ourselves. Comparison of the results shows that the best models can be obtained when PLS modelling is used and the spectra are pre-processed with the MSC method.