I haven't done the stats--and I suspect you'd have a hard time doing a statistically valid study--but keep in mind the ostensible objective of producing working, error-free code--not just quantity of code. A good pair should be able to create at least as much correct code as two programmers working separately.