A hybrid variational-ensemble data assimilation (denoted as hybrid GSI-EnKF) based on the Gridpoint Statistical Interpolation (GSI, a 3DVAR algorithm) and the Ensemble Kalman filter (EnKF) has been developed and successfully tested for the NCEP Global Data Assimilation System (GDAS) using the Global Forecast System model. Various results have shown that the hybrid GSI-EnKF provided better analyses and subsequent forecasts than the current operational GSI. In the hybrid GSI-EnKF, the flow-dependent ensemble covariance is incorporated during the data assimilation whereas in the GSI (a 3DVAR algorithm), a static covariance is adopted. Therefore, how the observations are assimilated and how the observation information is used by the hybrid GSI-EnKF and by the GSI can be different. In this study, the impacts of various observations assimilated using the GSI and the hybrid GSI-EnKF were compared. In the control experiments, all operational conventional and satellite data were assimilated. Data denial experiments were then conducted to access impacts of observations of interest including both conventional observations and satellite observations such as AMSU, AIRS, IASI, and GPS. Our results comparing the impacts of AMSU have shown that the amount of positive impact of AMSU assimilated by the hybrid GSI-EnKF was greater than that assimilated by the GSI, which suggested the hybrid GSI-EnKF made better use of the AMSU data than the GSI. It is also found that the amount of positive impact of assimilating AMSU on top of other observations was less than the amount of positive impact resultant from adopting the new hybrid GSI-EnKF data assimilation method as opposed to the pure GSI. The OSE results for other observation types and platforms will also be presented in the conference. In addition to the data denial experiment, the ensemble based observation impact metric has been developed to estimate the impact of observations assimilated by the hybrid GSI-EnKF. Unlike the adjoint based observation impact metric, the ensemble based observation impact does not require the tangent linear and adjoint of the forecast model. Initial result testing the ensemble observation impact metric for the conventional data has shown promising result. Further studies using the ensemble observation impact for various observation types and platforms will also be presented in the conference.