I photographed the following scene with my new 18-200 VR (actually, VRII) lens @200mm, using a Nikon D80 body. I focussed on the group of cars in the center. Camera setting was: ISO 400, shutter priority, shutter speed from 1/125s to 1/15s (4 shutter speed combinations). All shots are JPEG L, normal quality, with no post-processing of any kind.
For each shutter speed I used three different shooting configuration:
For each configuration I took ten continuous shots with VR off and 10 continuous shots with VR on (normal setting). For each of the 4x3x2=24 different shooting combinations (total 240 shots) I selected the 4 best shots, for which I am comparing below crops 240x160 pixel wide, shown at 100% magnification.
When the quality difference among the four best crops is small, I am also indicating how many good shots (considering the given shutter speed, not in absolute terms!) I got out of 10, to give an idea of the behaviour of the shot distribution beyond these four. Of course the definition of "good" here is subjective, and this number is only meant to compare results at fixed shutter speed, but different shooting configurations.
For comparison, on the last column of each row I show - bordered in red - the crop from a reference shot, taken with a shutter speed of 1/800s VR On, to show what one can expect under best conditions.
The white dot covering part of the car licence plate was added for privacy protection reasons.
At the end of this page I summarize my conclusions. If you wish, feel free to drop me an email. Thanks!
At 1/15s, where hand holding a 200mm lens - equivalent to 300mm for the Nikon D80 - is really hard (at least for me!), VR does make the largest difference when shooting unsupported. This way probably the VR algorithm can work the way it was designed for, and so it works best. If one looks for support (either vertical or horizontal), the VR is less effective.
At 1/30s and 1/60s VR still makes a significant difference, and it does so regardless of the support used (wall, table or nothing). Also, I see no difference in quality in the three cases, so I would still prefer to shoot with no support for consistency.
At shutter speeds where hand-held shooting gives good results (for me this is 1/125s), VR does not make appreciable difference, for good or for bad, at least shooting unsupported. Therefore it seems that one can leave VR on and forget about it. Shooting with horizontal support, the VR even provides results which are marginally worse compared to those obtained with VR off.
Comparing the results at 1/15s - no support - VR On, with those at 1/60s and 1/125s - no support - VR Off, I would conclude that 1/15s VR On provides better results than 1/60 VR Off, but worse than 1/125s VR Off. This is like saying that the VR improves results by roughly 2.5 stops, at least in this test.
In principle one should also consider the fact that at 1/15s f25 diffraction starts to kick in and reduces the overall sharpness - see my test on sharpness at different apertures. If one allows for this, then the 2.5 stops may be seen as a conservative lower limit, and say 3 stops an upper limit.
If you wish to know more on the VR algorithm, see the Nikon documents: here and here. They explain how the VR actually works. It is interesting - and reassuring - to note that my conclusion on the 2.5 to 3 stops improvement is in fair agreement with what found on the first of these documents.
However, I do not seem to agree on the further statement that VR can cope with tripod-based vibrations, since supporting the camera did not improve my results.