This work evaluates the feasibility of near-infrared (NIR) for determining moisture, protein and fat content of the famous Greek dish taramosalata. Furthermore, it is investigated the relative performance of the calibration model in the context of varying calibration data size and number of used wavelengths. Various calibration data sizes (n = 40, 50,... 80) and number of wavelengths (1, 2,..., 6) were considered. Calibration samples were analyzed employing traditional chemical methods and scanned using an Instalab 600-Dickey-John NIR apparatus. Calibrations were achieved with the use of multilinear regression between chemical and spectral data from each calibration data set. The optimal prediction errors (root mean square error of prediction, RMSEP), obtained with models based on six wavelengths and a calibration data size of 80, were 0.115% for moisture, 0.023% for protein and 0.088% for fat. The prediction errors decrease with an increase in calibration size and in number of wavelengths but seem to become stabilized around calibration data size 60 and number of wavelengths 3. Since a calibration set may be expensive to acquire and calibrations using many wavelengths tend to overfit the data, the accurate calibrations based on three wavelengths and a calibration data size of 60 might be preferred. The RMSEP values obtained with these models were 0.271% for moisture, 0.115% for protein and 0.222% for fat. NIR measurement as performed by the Dickey John Analyzer was proved a rapid and accurate method for analysis of taramosalata and may be used as a replacement for conventional expensive and time-consuming wet chemistry methods. (C) 2003 Elsevier Ltd. All rights reserved.