Continuum equations derived from a probability density function kinetic equation contain dispersion tensors that describe the interaction between inertial particles and the underlying turbulent flow in which they are transported. These tensors require closure treatment and recent work has shown that traditional closure approximations perform poorly when applied to the case of particle dispersion in turbulent boundary layers. The dispersion tensors are intrinsically non-local, being sensitive to both the strong inhomogeneity of wall-bounded turbulence and the influence of particle-wall collisions. A new strategy for constructing non-local closure models is presented to account for such influences. An important feature of the approach is that it utilizes exactly the same input parameters required for the traditional closures. Differences between the two approaches are therefore a reflection of the improved closure strategy, rather than a consequence of improved or additional input data. Predictions from both new and standard models are compared with benchmark statistics obtained from the simulation of particle trajectories in a synthetic boundary layer flow, generated via an inhomogeneous kinematic simulation technique. The results demonstrate the significant advance offered by the new approach.