As a preliminary step for assessing the impact of global positioning system (GPS) refractive delay data in numerical weather prediction (NWP) models, the GPS zenith tropospheric delays (ZTDs) are analyzed from 51 permanent GPS sites in the western Mediterranean. The objectives are to estimate the error statistics necessary for future assimilation of GPS ZTD data in numerical models and to investigate the variability of the data in this area. The time series, which were derived continuously from November 1998 to June 2001, are compared with independent equivalent values derived from radiosonde profiles and the High-Resolution Limited-Area Model (HIRLAM) NWP model. Based on over two years of data, the difference between radiosonde and GPS ZTD has a standard deviation of 12 mm of delay and a bias of 7 mm of delay. Some sites have biases as high as 14 mm of delay. The bimodal distribution of residuals, with a higher bias for daytime launches, indicates these biases may be due to radiosonde day-night measurement biases. The biases between the GPS ZTD and HIRLAM estimates are smaller, but the 18-mm ZTD standard deviation is significantly greater. The standard deviation of the residuals depends strongly on the amount of humidity, which produces an annual signal because of the much higher variability of water vapor in the summer months. The better agreement with radiosonde data than HIRLAM estimates indicates that the NWP models will benefit from the additional information provided by GPS. The long-term differences between the observational data sources require further study before GPSderived data become useful for climate studies.