Abstract: Realistic dynamic analyses of flexible pavements under moving vehicle loads can only be done on the basis of three-dimensional (3D) layered models. However, analyses on the basis of two-dimensional (2D) models, which overestimate the pavement response, are very attractive because of their much lower computational cost. In this paper, ordinary and improved 2D plane strain layered models are presented and compared against 3D layered models in the framework of linearity and analytical/numerical methods of solution. The improved 2D model is constructed by introducing viscous body forces for absorbing the waves propagating along the direction perpendicular to the plane of the model. The accuracy of above two 2D models is determined by comparing their response results against those coming from the 3D model and for the following pavement modeling cases: elastic or viscoelastic, isotropic or cross-anisotropic, layered half-plane and half-space. The vehicle load is uniformly distributed over a finite length or a rectangular area for the 2D and 3D models, respectively, and moves at a constant speed. All the cases are analyzed by the method of complex Fourier series used with respect to time and one or two horizontal coordinates. Thus, the partial differential equations of motion for the 2D or 3D models reduce to ordinary differential equations, which can be easily solved analytically/numerically. The results show that the ordinary 2D model overestimates the surface vertical displacement by up to 284% and the critical strains by 10–43% relative to the 3D model, while the improved 2D model only approaches the 3D response at very high load speeds in a not well defined manner. For this reason, extensive parametric studies are conducted on the basis of both ordinary 2D and 3D models for the creation of comprehensive response databanks, which are used in conjunction with nonlinear regression analyses for constructing accurate response prediction expressions for 3D models using results from 2D models in a very efficient manner. The simplified expressions, derived using a genetic algorithm for optimal term selection, predict 3D responses with R² values ranging from 0.92 to 0.99.