Abstract: The distribution of the precipitable water vapor (PW) over the Tibetan Plateau has an important impact on the regional weather and climate. The reliability of PW measurements is of concern. The characteristics of the systematic and random errors of the radiosonde (RS) PW data in comparison with groundbased GPS measurements at Lhasa during the period from 1999－2008 and at Naqu in 2003 are studied. The results show that the RS_PW is significant dryer than the GPS_PW at Lhasa in recent 10 years. Different types of humidity sensor show a different magnitude of the dry bias of PW. The GZZ 2 (goldbeater's skin hygrometer) and the GTS-1 (carbon hygristor) have the relative mean dry bias of 8.8% and 4.4%, respectively, and have the relative mean random errors of 19.8% and 13.3%, respectively. The relative PW difference is apparently reduced in recent 10 years, from 12.7% to 2.4%. The main reason is that high performance humidity sensors (GTS-1) have been introduced. The variation characteristics of the RS_PW dry bias are also investigated. The results show that the RS_PW dry bias exhibited pronounced diurnal and annual variations. The dry bias of the RS_PW is much larger at 12:00 UTC than that at 00:00 UTC and larger in summer than that in winter. The GZZ-2 and the GTS-1 have the relative mean PW bias of 15.8% and 8.3%, respectively, at 12:00 UTC at Lhasa. The GZZ-2 and the GTS-1 have the relative mean PW bias of 1.6% and 0.5%, respectively, at 00:00 UTC at Lhasa. The GZZ-2 has the relative mean PW bias of 12.4% and 0.3% at 12:00 and 00:00 UTC respectively. The causes of diurnal and annual variation of the RS_PW dry bias are analysed. It can be seen that the diurnal variations of the RS_PW dry bias are significant mainly because the air temperature is higher at 12:00 UTC than that at 00:00 UTC. The annual variations of the RS_PW dry bias are pronounced mainly because the air temperature is higher in summer than that in winter.