It's not that the scale changes, it's that the statistics count too few events to properly handle external factors. When you're looking a cohorts of around 4000 people, small things can change things by multiple percent.