The accuracy with which MODFLOW simulates surface water-groundwater interaction is examined for connected and disconnected losing streams. We compare the effect of different vertical and horizontal discretization within MODFLOW and also compare MODFLOW simulations with those produced by HydroGeoSphere. HydroGeoSphere is able to simulate both saturated and unsaturated flow, as well as surface water, groundwater and the full coupling between them in a physical way, and so is used as a reference code to quantify the influence of some of the simplifying assumptions of MODFLOW. In particular, we show that (1) the inability to simulate negative pressures beneath disconnected streams in MODFLOW results in an underestimation of the infiltration flux; (2) a river in MODFLOW is either fully connected or fully disconnected, while in reality transitional stages between the two flow regimes exist; (3) limitations in the horizontal discretization of the river can cause a mismatch between river width and cell width, resulting in an error in the water table position under the river; and (4) because coarse vertical discretization of the aquifer is often used to avoid the drying out of cells, this may result in an error in simulating the height of the groundwater mound. Conditions under which these errors are significant are investigated.