Calculating derivatives of spectral data by the Savitzky-Golay (SG) numerical algorithm is often employed as a preliminary preprocessing step in order to resolve overlapping signals, enhance signal properties, and to suppress unwanted spectral features that arise due to nonideal instrument and sample properties. Addressing these issues, the study on the simulated and the measured infrared data by partial least squares regression has been conducted. The simulated data sets were modelled by considering a range of undesired chemical and physical spectral anomalies and variations that can occur in a measured spectrum, such as baseline variations, noise and scattering effects. The study has demonstrated the importance of optimization of the SG parameters during the conversion of spectra into derivative form, specifically window size and polynomial order of the fitting curve. A specific optimal window size is associated with an exact component in the system being estimated, and this window size does not necessarily apply for some other component present in the system. Since the optimization procedure can be time consuming, as a rough guideline spectral noise level can be used for assessment of window size. Moreover, it has been demonstrated that, when the extended multiplicative signal correction (EMSC) is used alongside the SG procedure, the derivative treatment of data by the SG algorithm must precede the EMSC normalization.