The quadrature error in the discrete variable representation (DVR) corresponds to first-order perturbation corrections of the potential matrix elements, and is remarkably large. It causes an unphysical (ghost) level at E(ghost)approximate to(3/4)E-max in a one-dimensional problem. In a multidimensional problem the ghost components of the wave functions create large errors in levels and intensities of spectral transitions. Estimation of the quadrature error for a general formula of numerical integral is presented. The convergence direction of the DVR levels is explained based on the error analysis. Near-variational forms of the DVR are presented, which remove the ghost levels, improve convergence of the levels and intensities and are easily combined with the Lanczos method to reach high computational efficiency. As an application, vibrational band intensities and levels of H2O were calculated by using the best potential energy surface available, and two dipole surfaces.