The sensitivity of forecast outputs to observational inputs is quantified by utilizing the adjoint formulations of a numerical weather prediction model and data assimilation system. The reduction in short term forecast error due to the assimilation of observations is investigated using a limited area model with horizontal grid spacings appropriate for mesoscale weather systems. The relative impact of observations can vary greatly depending on the physical location of the model's domain and the area over which the error is calculated. For example, observations from radiosondes and aircraft are important for reducing error over the eastern United States, but satellite derived winds and surface reports over the eastern Pacific Ocean are most important if the target area is over the western US. There are subtle differences in the observation impacts on forecasts utilizing a range of horizontal grid spacings (20-60 km). In a few cases for the smaller grid spacings, the adjoint of the limited area model fails to account for a satisfactory amount of the nonlinear model error. Impacts on metrics other than forecast error will also be presented.