Although it might be thought that the determination of lignin content in wood by near infrared (NIR) spectroscopy is well known and has been used for several years, model statistics (mainly errors of prediction) found in the literature recommend further study. It is shown that partial least squares regression (PLS-R) models can be improved, namely the number of PLS vectors and the error of prediction can be substantially decreased by careful selection of the combination of wavenumber range(s) and pre-processing methods and validation of the models. To cover a wide range of the natural variability, the total lignin content of 200 Norway spruce wood samples was determined by wet-laboratory chemical methods. From the same milled samples Fourier transform near infrared (FT-NIR) spectra were recorded using a NIR fibre-optic probe. NIR bands, property weighting spectra and correlation coefficients were used to pre-select convenient wavenumber ranges. PLS regressions were carried out to establish a mathematical correlation between the data sets of wet-laboratory chemical methods and the FT-NIR spectra, leading to a number of “good” models with similar coefficients of determination ( r2)>0.90 and root mean square errors of cross-validation ( RMSECV) below 0.3%. External validation of the models also gave r2>0.90 and root mean square errors of prediction ( RMSEP) below 0.3%. As several models showed similar statistics, a further simple step, called evaluation, was introduced to assist in being able to make a decision about what model should be used. Therefore, in addition, FT-NIR spectra of another 366 wood samples were measured to evaluate the pre-selected combinations of wavenumber ranges and pre-processing methods. With this simple additional step (evaluation) the model with the highest predictability could be selected easily.