Global positioning system (GPS) data from over 260 ground-based permanent stations in China covering the period from 1 March 1999 to 30 April 2015 were reprocessed to generate homogenized precipitable water (PW) above each site with an accuracy of about 0.75 mm. The systematic PW errors in four types of radiosondes (GZZ2, GTS1, GTS1-1 and GTS1-2) used in China during this period were then investigated, and the reliability of the usage of PW from eight commonly used reanalyses (20CRv2, CFSR, ERA-Interim, JRA-25, JRA-55, MERRA, NCEP– NCAR, and NCEP–DOE AMIP-II) in multi-time scales were also evaluated. Comparisons reveal the exists of common wet biases in GZZ2 and dry biases in GTS1 which introduce artificial decreasing PW trends. As for the reanalyses, evaluations show that all reanalyses (except for 20CRv2) can generally reproduce well the climatological annual mean PW, annual amplitudes, and semiannual amplitudes, but they all show wet biases in the Tibetan Plateau. Diurnal variation amplitudes reproduced by all reanalysis products are smaller than amplitudes estimated from GPS observations over China as a whole, and none of the reanalyses can capture the diurnal phases correctly. Because of the assimilation of the unhomogenized radiosonde humidity data, most reanalyses show artificial decreasing PW trends (except in 20CRv2 and CFSR).