From a pure data acquisition standpoint, replacing an interpolation by a direct measure will, by definition, provide a better result, unless in the practically unfeasible case where the result of the interpolation is equal to the measured value.
There is another case where it won't make any difference: where there isn't any data.
The idea that the world has "real" colours, which you can capture data about, which can be more or less accurate, is just wrong. The colour we see does not exist outside our heads: it is purely subjective. There is no such thing as objectively "better colour" or "more accurate colour", only subjectively better or more accurate colour.
Of course, although there is no "colour" in the real world there is light of different wavelengths, and the wavelengths
are data. But a silicon-based detector can't measure wavelength: we have no access to that data. We can use an RGB system and, by experiment, develop an algorithm that reproduces the effect of light of a particular wavelength on the RGB system in our retina. The catch is that the micro-lenses do not transmit light of only one colour. The micro-lens means that the quantum efficiency of a pixel with a (say) green lens is less for red and blue light than for green light, but it is not zero. The same number of captured photons could result from pure green light, or from less intense green light plus some blue and/or red light. You have to find a way of working out, or guessing what it
probably was.
If you had a measured RGB value for every pixel you could calculate it directly, since you know the spectral transmission of the micro-lens. That would get rid of interpolation. That is
not what you get with pixel shift. With pixel shift you get a value from the pixel's neighbour with a different coloured micro-lens when it looked (you hope) at the same part of the image. You are still interpolating: assigning the pixel a value for the colours it did not measure based on its neighbours which did measure those colours. That is no different to the Bayer mosaic for detail bigger than four pixels, so for detail on that scale pixel shift won't change the RGB values. Detail smaller than two pixels is not resolved in either case. So there is a window where you
might get some benefit, if the sensor shift is precise enough, and you have really good tripod and a lens with no lateral colour.
It is reasonable to ask whether you get
different answers by looking at a pixel's neighbours compared to pixel shift. We will have to wait and see, although you certainly won't where the detail is either larger than eight pixels or smaller than two pixels.
