An adjoint-based procedure for assessing the impact of observations on short-range forecast error in numerical weather prediction is described. The method is computationally inexpensive and allows observation impact to be partitioned for any set or subset of observations, by instrument type, observed variable, geographic region, vertical level or other category.