Augmenting travel-time data with backazimuth information can improve regional seismic event locations for which only a few recording stations are available. This study examines the achievable accuracy of backazimuths derived from 164 regional short-period P waves recorded at Regional Seismic Test Network (RSTN) stations. The data represent 62 events with a good distribution of distance, azimuth, and signal-to-noise ratio (SNR) at each of four RSTN stations. Using Preliminary Determination of Epicenter locations to generate reference backazimuths, we examine the effects of computational method, signal window length, frequency content, and SNR on RSTN backazimuth estimates. The variety of geologic environments at the receivers also allows assessment of the effect of receiver structure on backazimuth estimation. Our results indicate that regional three-component recordings can produce useful backazimuths; more than 75 per cent of the records yield backazimuths within 20° of the correct value when reasonable parameters are used. The accuracy of the calculated backazimuths does vary with the receiver, however. RSTN stations in New York and Ontario, located in Precambrian terranes, produce very accurate backazimuths for SNRs greater than 5 dB. At the South Dakota and Tennessee stations, located in sedimentary rocks, complicated receiver structure is the controlling factor in backazimuth estimation. Propagational complexities at the South Dakota site preclude recovery of reasonable backazimuth values; poor statistics are obtained for all parameter combinations. At the Tennessee site, backazimuth accuracy depends strongly on the computational method used. In general, the choice of frequency band is important and varies by stations, while the length of the signal window is less critical. If the problematic South Dakota records are excluded, we achieve an rms backazimuth error of about 6° for recordings with SNR above 10 dB. While a single optimized choice of parameters produces good results for the 164-record data set, a further 5 to 10 per cent improvement can be achieved if the parameters are customized on a station-by-station basis.