I had similar empirical results on one of my PhD projects for medical image classification. With small data sets, we got better results on 8-bit data sets compared to 16-bit. We viewed it as a form of regularization that was extremely effective on smaller data sets with a lot of noise (x-rays in this case).