We have evaluated the location accuracy of the U.S. National Lightning Detection Network (NLDN) via comparison of individual NLDN stroke locations with their corresponding stroke terminations as determined from multiple-site video recordings of local cloud-to-ground lightning flashes. The video records used in this analysis were acquired in the vicinity of Albany, New York, during the summers of 1994 and 1995, seasons corresponding to the initial and final phases of the upgrade of the NLDN to the improved accuracy from combined technology (IMPACT) configuration. For 1994 a total of 751 strokes were located in common between the video and NLDN data sets, yielding median and mean values of the NLDN-video location separation of 2.21 km and 3.74 km, respectively. Constraint of the 1994 data to the best video locations (those with a semimajor axis value of the 50% error ellipse less than 500 m) yielded a subset of 53 strokes with median and mean values of the NLDN-video location separation of 2.61 km and 4.74 km, respectively. For 1995 a total of 219 strokes were located in common, yielding median and mean values of the NLDN-video location separation of 442 m and 865 m, respectively. Constraint of the 1995 data to that year's best video locations (those with a semimajor axis value of the 50% error ellipse less than 200 m) yielded a subset of 79 common strokes with median and mean values of the NLDN-video location separation of 435 m and 625 m, respectively. The IMPACT upgrade of the NLDN apparently has resulted in about a fivefold increase in location accuracy. A special subset of 11 strokes occurred in 1995; these were located by the NLDN and observed to terminate on local structures of accurately known location. For these 11 strokes, the median and mean values of the NLDN location error are 518 m and 484 m, respectively, a result quite consistent with the overall NLDN-video location comparison and the claimed 500-m median location accuracy for the upgraded network. Indeed, we argue that removal of an apparent residual bias in the NLDN locations ultimately could yield a median location accuracy in this region approaching 250 m. Finally, we find that the NLDN location error is a function of the estimated stroke peak current, with the weakest peak currents having the poorest locations for both years. For the 1995 data, strokes with peak currents of intermediate magnitude (20--30 kA) were the most accurately located. ¿ 1998 American Geophysical Union