Users who enable notifications retain twice as well — so should you force notifications on, or are you just describing people who already liked the product?
People carrying umbrellas and wet pavements go together. Confiscating the umbrellas will not dry the pavement.
Only a causal claim justifies an action, and feature-adoption metrics are confounded almost by construction.
Two variables moving together admits four explanations: A causes B, B causes A, something else causes both, or it is chance. Analytics work overwhelmingly produces the third — a confounder — because the people who adopt a feature differ from the people who don't in every way that also predicts retention. The reason this matters commercially is that only a causal claim justifies an action. If notifications cause retention, turning them on helps; if engaged users simply enable notifications, forcing them on annoys everyone and moves nothing.
Correlation has four possible causes: A→B, B→A, a confounder driving both, or coincidence. Feature-adoption metrics are almost always confounded, because adopters self-select. Only a randomised experiment cleanly separates them; when you can't run one, quasi-experimental designs — difference-in-differences, regression discontinuity, instrumental variables — get you closer, with assumptions you must state.
Correlation vs causation explained by Dr Nic with examples — Dr Nic's Maths and Stats, 4:28