Numerical ray tracing by BowSim shows that departures from sphericity by a mere 1-2% noticeably distort the primary bow. We have to marvel that rainbows are ever seen!

The BowSim computation at right is for a mixture of spherical drops and oblate spheroids that have their vertical axis shortened by 2.5%. This could for example be produced by two showers falling through an observer's rainbow cone, one of smaller spherical drops and the second of large drops flattened by air resistance. The lower bow is from the oblate spheroids. At the simulation solar altitude of 5º the secondary bow is undistorted. This example is not to suggest that oblate spheroids, merely convenient for the calculation, are required to make twinned bows but rather that twinned bows could be produced by some type of non-sphericity.

We need more observations of twinned bows to better know their nature; close-up images at known focal lengths of the 'twins', wider angle images around the bow and if possible the secondary in order to measure the primary-secondary spacing and thus any departures from circularity, exposures to show the detailed structure of the secondary, all combined of course with the local weather conditions.