Abstract: The observation error covariance matrix plays a key role in the variational data assimilation. Assumptions for using the innovation vector method to estimate the observation error of the COSMIC (Constellation Observing System for Meteorology, Ionosphere, and Climate) GPS (Global Positioning System) RO (radio occultation) data are analyzed and the algorithm for estimating the observation error of refractivity is composed. Using the H-L method, the observation errors of the COSMIC refractivity are estimated over the low-latitudes (0°-30°), mid-latitudes (30°-60°) and high-latitudes (60°-90°) of the north and south hemisphere respectively in the different seasons. The impacts of the refractivity observation error on the GRAPES (Global/Regional Assimilation and Prediction Enhanced System) assimilation and forecast skill are assessed by the assimilation and forecast experiments of GRAPES. The fractional refractivity observation errors vary with the latitude and height significantly. The results illustrated that a single error profile is not a good global representation of the refractivity errors. On the other hand, the fractional refractivity observational errors vary with the season over the mid-latitudes and high-latitudes of the south and north hemisphere, and in those locations the error in the boreal summer is twice of that in the boreal winter. The observation error profiles are symmetric in the south and north hemisphere in both the boreal spring and autumn. Compared with the single fractional refractivity error profile, the new refractivity observation error estimated in this paper results in an improvement in the anomaly correlation scores and a reduction of geopotential height errors, temperature errors and wind errors of the GRAPES forecasts.