"No significant effect" is one of the most expensive sentences in analytics — because it's so often wrong, or rather, incomplete. A flat aggregate can hide a strong win in one segment and a drag in another that cancel out.

First, earn the right to slice

Before cutting by segment, check the boring things: was the test powered, did randomization hold, are the guardrails clean? If those fail, segment analysis is just noise with extra steps.

Then slice — but pre-register the cuts

The danger of segmentation is that if you cut enough ways, something always looks significant. The discipline:

  • Decide the segments before you look.
  • Correct for multiple comparisons.
  • Treat anything you find post-hoc as a hypothesis for the next test, not a conclusion from this one.

Tell it as a trade-off, not a verdict

The strongest version of this analysis doesn't say "ship it" or "kill it." It says: here's who wins, here's who loses, here's the size of each — your call. That framing is what turns an analysis into a decision.

This is a placeholder article. Edit it at content/articles/reading-a-flat-experiment.md, or add your own .md file to the same folder and it will appear here automatically.