"I'll tell you another thing," Roscoe continued. "You'll never see that kind of squib with an error bar of, say, 1 percent."
I thought about that for a couple of seconds, but Roscoe wasn't patient enough to wait for my thoughts on the subject.
"Doesn't take an Einstein to figure out why, either. If you use the same logic, you'll see that they have to talk to 10,000 people to get the error down to 100, or 1 percent. That's 10 times as expensive as talking to 1,000 people. So to push the error down from 3 percent to 1 percent, it costs you an incremental factor of 9."
"Ten," I thought to myself. Then I realized you could reuse the first thousand in the second sample. Never misses a trick, that Roscoe.