It is necessary to calibrate the frequency response of each telescope, in both amplitude and phase. In principle, one could apply bandpass corrections in a baseline- or an antenna-based scheme. The former is the more conceptually accessible since the visibilities are baseline dependent. However, for large arrays such as the VLA, this would mean the bandpass table could become very large as there are N(N-1)/2 interferometers for N antennas. This is the main reason why the AIPS bandpass generating task, BPASS, decomposes the baseline-based responses into antenna-based responses and writes these to the BP table. This is done in a manner very similar to the general calibration discussed previously.
To make a good bandpass calibration, it is common to observe a strong source for some period of time. This is especially important for high dynamic range spectral-line applications. Usually, the bandpass calibrator is observed before and after the rest of the program, so that a check on the temporal stability of the bandpass can also be made. If you don't have a special bandpass calibrator, then you will probably need to use your secondary calibrator(s), or perhaps the primary, if you observed it long enough. So, you must decide whether to try and work out the bandpass response as a function of time, or whether you really only have sufficient signal-to-noise ratio if you average all your calibrator data together (usually it is the latter).
A useful technique in determining bandpasses is that which divides the spectral data by the channel 0 data first. This essentially removes any source structure and atmospheric gain (phase and amplitude) errors and reduces the source to a pseudo point source. Therefore the standard calibration equation (see § 3) can be used to work out the gains for each channel. This means that the bandpass calibrator does not necessarily have to be a point source, just that it is bright. However, this technique relies on their being minimal source structure across the bandpass. If you are doing multi-frequency synthesis work, then it is quite possible that there will be source structure across the 128 MHz bandwidth. In these cases, you should make sure that your bandpass calibrator is indeed a point source (such as 1934-638).
If this techniques is applicable, there is no need to calibrate the data, because the division removes the gain errors. In any case, even if you could calibrate with this option, it would also just divide out.
It is also important to realize that this type of bandpass calibration implicitly assumes that the bandpass calibrator's flux density does not vary significantly over the band. Remember this when and if you attempt multi-frequency synthesis, as it affects the validity of the whole procedure. Note that MIRIAD can do a better job of this because you can input the spectrum of the source if you know it. MIRIAD has built in knowledge about some calibrators, in particular, 1934-638.