This paper describes the results of the statistical comparison of the inversions of GPS/MET radio occultation data with analysis data of the European Center for Medium-Range Weather Forecast (ECMWF). The Prime Time 4 period (February 2--16, 1997) GPS/MET data were analyzed. The data analysis algorithms include the derivation of refraction angles from the measured phase excess, the back-propagation method for handling data in multipath regions, ionospheric correction and noise reduction, and the Abel inversion of the refraction angle profiles and the derivation of dry temperature. We use a forward model of radio occultation experiments in order to produce artificial occultation data for the ECMWF global fields of atmospheric variables. The forward model exists in two variants, which are geometric optics propagator and wave optics propagator. The wave optics propagator allows for the simulation of diffraction effects and multipath propagation, which made this operator the choice for the validation. The artificial occultation data are processed by exactly the same inversion algorithm as the GPS/MET data. This allows for the adequate comparison of the retrieved temperatures. For regions with dense observational networks (Europe, China, and United States), the comparison between the satellite data inversions and the model analysis fields is characterized by biases inside 0.5 K and mean square deviations of 1.5--2 K. In the Southern Hemisphere, where observational data are sparse, the bias can reach 2 K, and the mean square deviation is 2--3 K. Because the quality of GPS/MET data must not be different in different hemispheres, the difference in the bias is assumed to be based on the used ECMWF analysis data. ¿ 2001 American Geophysical Union