The goal of this article is to stress the importance of considering the phase in QSPR studies. It is found that the phase plays a fundamental role in the QSPR models from both a statistical and a physical point of view. From a statistical point of view, it is observed that the predictive performance drops drastically when the QSPR model, obtained for a given phase, is used to predict the solubility of the same set of compounds but in another phase. From a physical point of view, when the compounds in the training set are in different phases, the physical interpretation of the descriptors involved in the model is obscured because the descriptors which appear in the correlation equation are a sort of average encoding the different physical mechanisms underlying the property. It is shown that the use of compounds in the same phase, instead, allows a more transparent physical interpretation of the descriptors involved in the model and, at the same time, improves the statistics of the models.