Classical inference in statistic and econometric models is typically carried out by means of asymptotic approximations to the sampling distribution of estimators and test statistics. These approximations often do not provide accurate p-values and confidences intervals, especially when the sample size is small. Moreover, even if the sample size is large, the accuracy can be poor due to model misspecification (nonrobustness). Several alternative techniques have been proposed in the statistic and econometric literature to improve the accuracy of clasical inference. In general, these alternatives address either the accuracy of the first-order approximations or the nonrobustness issue. However, the development of general procedures which are both robust and second order accurate is still an open question. In this thesis, we propose an alternative statistical test wich has both robustness and small sample properties for two large and important classes of models: Generalized Linear Models (GLM) and models on overidentifying moments conditions.