The purpose of this dissertation is to contribute to research taking place in the field of encompassing in regression models. The main idea underlying this principle is to select a model only if it can account for, or explain the results of a rival model. The validation of the model is then done by comparing the results obtained with those of another model. First we define the notion of exact encompassing ; it rests on the existence of a function linking the estimators of each models. The study of the pseudo-true value in the second model enables us to define the approximate encompassing. We then propose encompassing statistics based on the difference between one estimator of the second model and an estimator of the pseudo-true value. We test the validation of the encompassing model by studying asymptotically these statistics, once normalized. We then apply the concept of approximate encompassing to the problem of non-nested regressors choice. We present parametric encompassing tests and link them to classical tests. The results obtained in this parametric setting are then extended by using non-parametric technics of regression estimation. We propose four statistics (parametric or functional) by combining parametric and non-parametric specifications for each of both models. We show that each of the statistics is normally distributed with zero mean. We also study the choice of the window-width affecting those results and we define the pseudo-true window-width connected to the pseudo-true value estimator. Finally, in the context of two non-parametric models, we propose a global encompassing criterion and we analyze its asymptotic behavior.