The missing comparison
A shop’s sales rise after it introduces an advert. That sequence does not reveal what sales would have done without it. A holiday, a competitor’s closure or changing prices could contribute. The unobserved no-advert outcome is the counterfactual we need to estimate.
Design matters more than a impressive line
Random assignment can make treatment and comparison groups comparable in expectation. Natural experiments use external variation, but require assumptions about what changed and for whom. Regression can describe associations; adding variables does not automatically create a causal design.
A before–after example
A treated store rises from 100 to 130 sales while a comparison store rises from 100 to 120. The difference in changes is 10 sales, not 30. Interpreting that as causal still requires a defensible comparison, including assumptions about trends and other differences.
Read the limits
Ask about selection, measurement, uncertainty and whether the result applies elsewhere. “Statistically significant” is not the same as large, economically important or causal. Econometrics is a discipline of explicit assumptions, not a machine that turns any dataset into proof.