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. Results here are based on 24h global forecast error in the Navy Operational Global Atmospheric Prediction System (NOGAPS) and use observations assimilated at 00UTC in the NRL Atmospheric Variational Data Assimilation System (NAVDAS). Largest reductions in 24h global forecast errror in the Northern Hemsiphere are produced by rawinsondes, satellite wind data, and aircraft observations. In the Southern Hemisphere largest error reductions are produced by ATOVS temperature retrievals, satellite wind data and rawinsondes. Approximately 60 (40) percent of global observation impact is attributed to observations below (above) 500 hPa. A significant positive correlation is found between observation impact and cloud cover amount at the observation location. Currently, without consideration of moisture observations and moist processes in the forecast model adjoint, the observation impact procedure accounts for about 75 percent of the actual reduction in 24h forecast error produced by the complete set of global observations.