Ensemble averaging of advection-dispersion equations describing transport of a passive scalar in incompressible random velocity fields requires a closure approximation. Commonly used approaches, such as the direct interaction approximation and large eddy simulations as well as equivalent renormalization schemes, employ so-called two-point (or one-loop) closures. These approaches have proven to be adequate for transport in zero-mean (unbiased) time-dependent random velocity fields with increasing accuracy for decreasing temporal coherence. In the opposite limit of steady velocity fields with finite bias, however, these schemes fail to predict effective transport properties both quantitatively and qualitatively, leading to an obvious inconsistence for transverse dispersion in two spatial dimensions. For this case, two-point closures predict that macroscopic transverse dispersion increases as the square root of the disorder variance while it has been proven rigorously that there is no disorder-induced contribution to macroscopic transverse dispersion for purely advective transport. Furthermore, two-point closures significantly underestimate the disorder-induced contribution to longitudinal dispersion. We derive a four-point closure for stochastically averaged transport equations that goes beyond classical one-loop schemes and demonstrate that it is exact for transverse dispersion and correctly predicts an increase of the longitudinal disorder-induced dispersion coefficient with the square of the variance of the strong disorder. The predicted values of asymptotic longitudinal dispersion coefficients are consistent with those obtained via Monte-Carlo random walk simulations.