The Land Data Assimilation System (LDAS) project is a multi-institutional effort to produce better land-surface hydrology simulations over the continental US for use in weather and climate models. Several land-surface models (NOAH, Mosaic, VIC, CLM, and Sacramento) participate in the project. Simulations from these models differ even when they are driven by exactly the same forcing and same surface parameters. Understanding the reasons for these differences will allow us to improve land surface models. We take advantage of actual observations of fluxes, soil and vegetation properties, soil moisture, and soil temperature at more than 40 stations from around the US for the period of 1997-1999 to perform experiments with each model to investigate the reasons for these differences. We use different combinations of standard LDAS and locally observed forcing, surface parameter specification, and meteorology, and evaluate the results from the standard and special runs with observed land state variables and fluxes. The results provide guidance for model improvement and the design of future observing networks.