Great question. Seth Roberts has studied this as part of self-experimentation. One of the major issues is finding tasks that don't have too large a learning effect, which somewhat confounds the data.