What's the difference between correlation and causation? How do you test for causation?
StatisticsLevel: IntermediateStatistics and probability
What's the difference between correlation and causation? How do you test for causation?
Hints
- Compare joint variation of two quantities and the effect of their units.
- What changes when one quantity is rescaled? For a matrix, check symmetry, rank and sample size.
Answer
Correlation measures the strength/direction of a linear (or monotonic) relationship between two variables; it says nothing about one causing the other — confounders, reverse causation, or coincidence can produce correlation. To establish causation you typically need a randomized controlled experiment (A/B test), or, when experiments aren't feasible, quasi-experimental methods like difference-in-differences, instrumental variables, regression discontinuity, or propensity score matching — combined with a causal graph (DAG) to reason about confounders.
Follow-up: Give an example of a confounder that creates spurious correlation.
Source and licence: Ishan Dutta — Awesome Data Science Interview Q&A · MIT