Laser-Induced Breakdown Spectroscopy (LIBS) is a widely accepted technique used for both classification and quantification purposes considering complex and heterogeous samples. Based on a set of training spectra acquired from diverse and representative samples within a specific application domain, it becomes possible to apply various data processing techniques and modeling methods to construct the predictive model in question. Naturally the complexity of both the laser-matter and the laser-plasma interactions and the heterogeneity of natural samples often requires the development of various predictive models, which are then compared based on figures of merit such as the RMSEP (Root Mean Square Error of Prediction) value for quantification or the classification rate for qualitative analysis. Our ultimate goal is, of course, to select the model that appears to be the most accurate, which ultimately boils down to searching for the lowest RMSEP value or the highest classification rate. This is precisely where the whole problem lies because even if we observe a different level of error for two models, for example, this difference is not necessarily statistically significant. In such a case, we are therefore not allowed to say that the lower error indicates the best predictive model to consider. The purpose of this article is to provide a tutorial on introducing a statistical model comparison procedure, whether they are quantitative or qualitative. Two LIBS data sets have been used to illustrate the principles of the proposed method.