Causal Inference Without an Experiment
Explore approaches to causal inference when randomization is unavailable, including difference-in-differences, matching, and natural experiments. Understand the key assumptions, diagnostics, and common pitfalls to accurately estimate treatment effects and defend your analysis in data science interviews.
After an A/B testing discussion, the interviewer adds a constraint: “Suppose we can’t randomize. A feature launched globally on Monday, and leadership wants to know whether it increased weekly purchases. What do you do?” Because every eligible user received the feature, a simple comparison of purchases from Sunday and the following Sunday measures change over time, not the feature’s impact. Demand may have risen because of a holiday, prices may have changed, or post-launch users may differ from pre-launch users.
The missing ingredient ...