An important source of error in very-long-baseline interferometry (VLBI) estimates of baseline length is unmodeled variations of the refractivity of the neutral atmosphere along the propagation path of the radio signals. We present and discuss the method of using data from a water vapor radiometer (WVR) to correct for the propagation delay caused by atmospheric water vapor, the major cause of these variations. Data from different WVRs are compared with estimated propagation delays obtained by Kalman filtering of the VLBI data themselves. The consequences of using either WVR data or Kalman filtering to correct for atmospheric propagation delay at the Onsala VLBI site are investigated by studying the repeatability of estimated baseline lengths from Onsala to several other sites. The lengths of the baselines range from 919 to 7941 km. The repeatability obtained for baseline length estimates shows that the methods of water vapor radiometry and Kalman filtering offer comparable accuracies when applied to VLBI observations obtained in the climate of the Swedish west coast. For the most frequently measured baseline in this study, the use of WVR data yielded a 13% smaller weighted-root-mean-square (WRMS) scatter of the baseline length estimates compared to the use of a Kalman filter. It is also clear that the ''best'' minimum elevation angle for VLBI observations depends on the accuracy of the determinations of the total propagation delay to be used, since the error in this delay increases with increasing air mass. For use of WVR data along with accurate determinations of total surface pressure, the best minimum is about 20¿; for use of a model for the wet delay based on the humidity and temperature at the ground, the best minimum is about 35¿. ¿1991 American Geophysical Union