To improve the skill of ensemble prediction systems, one must be able to estimate accurately the probability distribution of the state of an atmospheric model given a set of observations. A new ensemble assimilation methodology, the ensemble adjustment filter, has been developed and applied to observing system simulation experiments in a fully parameterized primitive equation numerical weather prediction model. The filtering method uses information from an ensemble of model integrations to obtain an estimate of the covariance between model state variables and observations. Each available observation is then allowed to impact the prior distribution for each state variable independently. However, the way in which the required product of the observational error distribution and the prior state distribution is computed maintains much of the information about covariances of the prior state variables. A one year ensemble assimilation making use of 600 randomly located column observations of a control run of the NWP model has been performed. An analysis of the ensemble mean and ensemble spread from the filter assimilation is presented. Of particular interest are ensemble assimilated fields of non-state quantities like precipitation. The potential for operational assimilations based on the ensemble adjustment filter is discussed. Dealing with systematic model error remains a potentially serious problem.