Technically speaking, binning can only occur at the hardware level. "Binning" is done with per-pixel logic that adds the charge of adjacent pixels (i.e. 2x2) together to produce a stronger output charge. This cannot be done after the data has been digitized, at best, all you can do is average together pixel data from a RAW image to get a similar result.
There is a key difference between hardware binning and post-process averaging. While they produce a similar result, binning is actually producing an image with a stronger real signal with lower spatial resolution. In certain applications, such as say astrophotography, where you may be imaging something with an extremely low photon flux such that the signal from 1x1 binned pixels is buried in read noise, yet the signal from 2x2 binned pixels is strong enough to overpower read noise. Furthermore, as focal length increases, it becomes very easy to "oversample" the resolved image too much, which directly results in low photon flux per pixel. Binning effectively creates larger pixels, reducing image scale (which in best practice should be 2x to 3x the size of the average star in the image...with some of my equipment, for example, I am able to oversample more than 5x, requiring me to GREATLY increase exposure time to get enough signal strength to overpower read noise.)
It should also be noted that for true binning, you need a monochrome sensor. There is no real way to "bin" pixels in a bayer array. Again, at best, all you could really do is average, which will reduce noise, but it won't improve signal strength.
