06.5.6 · Skill in Statistical reasoning
Confounding
A hidden factor that affects both the cause and the effect.
Users who use a feature may retain better because they were already engaged, not because of the feature. Confounders make correlations look causal. Randomized experiments are the most reliable way around them.
Resources
2- articleGo deeperHow and when to control for confounders during product usage analyses ↗Heap · 2021 · 10 min
Uses the familiar "users of this feature retain better" case to show how a confounder fakes an effect and how to adjust for it.
- bookGo deeperCausal inference for the brave and true ↗Matheus Facure · free online book
A free, rigorous and readable book on causal inference with Python code, from randomized tests to methods for when you cannot run one.
Tools
none yetTools for this topic are being researched.