Potential outcomes and treatment effects
Give every unit two outcomes, one for each treatment, and causal inference becomes a missing-data problem. ATE, ATT and CATE; ignorability and overlap; propensity scores; weighting, matching and doubly robust estimation; and two designs for when confounders are hidden.
55 min
After this lesson you can
- Define potential outcomes, the individual, average and conditional treatment effects, and the ATT
- Explain the fundamental problem of causal inference and decompose the naive difference into effect plus selection bias
- State ignorability and positivity, and show how they identify the ATE
- Define the propensity score, prove its balancing property, and derive inverse propensity weighting
- Implement IPW, matching and the doubly robust (AIPW) estimator, and say when each fails
- Connect potential outcomes to the do-operator
- Describe instrumental variables and difference-in-differences, and their key assumptions
In this lesson
- Where the idea came from
- The vocabulary
- What the naive comparison measures
- Ignorability: randomisation within strata
- The propensity score
- Inverse propensity weighting
- Matching
- Doubly robust estimation
- When confounders are hidden: two designs
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.