We should compare all above metrics in conditional slice of - [Credit card form detected * User has cards on file (local or payments)]. At this time ablation happens without knowing whether user has cards on file and that would dilute metrics. So current ablation experiment is finding metrics for the slice of “credit card form detected” and excludes “user has cards on file” condition. Based on current metrics (shown suggestion/interacted with form) rate is 29%, so metrics are diluted by 3X. Counterfactual logging (User should have shown suggestions) should be implemented to compare conditional metrics properly
Refer Ablation experiment setup for autofill proposal - https://docs.google.com/document/d/1vMGO7sfzYDXpev_5ftyLH-WoKpCHA4I9lVLYuesVT9c/edit#heading=h.k6sbad93o9zk
Comment 1 by bugdroid1@chromium.org
, Oct 12 2017