The data assimilation system of the Global Modeling and Assimilation Office at NASA/GSFC applies a multivariate wind-height error covariance in the analysis. This covariance model can in principle accommodate a flexible three- dimensional specification of the error since the analysis equation is solved in physical and observation space. There are two components of the wind error: One is coupled to the geopotential height error through a basic mass-wind balance relation, while the other component is mass-decoupled and thus by definition uncorrelated with errors in the geopotential height field. The uncoupled error is further specified as a sum of uncorrelated rotational and divergent components, respectively. There are a couple of different methods that are widely used to provide estimates of parameters used in covariance models for meteorological analysis schemes. However, each of these methods has its own limitations. It is therefore necessary to test the estimates in assimilation/forecast experiments and modify them based on empirical considerations where necessary. We describe an improved wind-mass error analysis scheme in which the parameter estimates are based on the statistics of the observation-minus-forecast residuals. The impact of the scheme is illustrated by results from a series of assimilation and forecast experiments. The assessment of these results includes both statistical and dynamical aspects. The former is based on the bias and standard deviation of the short-range forecast error as well as standard medium-range skill scores, while the latter is based on the representation of features and phenomena of the atmospheric circulation in the analysis and forecast fields.