https://www.bradyneal.com/causal-inference-course
Introduction to Causal Inference라는 강의를 듣고 정리했습니다.
Z로 Y가 정의 되도록
8-1. What is an Instrument?
Assumption 1 : Relevance
Z → T
Assumption 2 : Exclusion Restriction
T가 mediate
edge를 지우는 것은 assumption을 추가하는 것과 같다
Assumption 3 : Instrumental Unconfoundedness
Z is unconfounded
U에서 Z로가는 path가 없어야 함.
Conditional Instruments
Z is unconfounded after conditioning on observed variables
8-2. No Nonparametric Identification of the ATE
backdoor path를 block할 수 없어서 nonparameteric identification을 써야함.
non-parametric?
8-3. Warm-Up: Linear Setting
Linear outcome
- Z 없음, exclusion restriction assumption
Binary Linear Setting
Multiplying Path Coefficients in Linear Setting
- relevance asssumption 때문에 denominator는 non-zero
이걸로 뭘 알 수 있음?
Wald Estimator
Continuous Linear Setting
T랑 Z가 continuous하면
- relevance asssumption 때문에 denominator는 non-zero
증명은
Two-Stage Least Squares Estimator
- binary setting에서도 쓸 수 있다.
8-4. Nonparametric Identification of Local ATE
Linear Outcome Assumption as Homogeneity
모든 unit이 같은 treatment effect를 가진다.(homogeneity) → very restricting
Potential treatment
Principal Strata
data → 4 strata (instrument가 어떻게 영향을 주는지)
Monotonicity Assumption
- No Defiers
- ∀i,Ti(Z=1)≥Ti(Z=0)
Deriving Local ATE Identification w/ Monotonicity Assumption
Nonparametric Identification of LATE Under Monotonicity Assumption
Problems:
• Monotonicity 항상 성립되는 것은 아니다
• Complier에서의 average effect(CACE)가 아니라 whole population에서의 average effect(ATE)를 알고 싶다.
8-5. More General Settings for the ATE
Nonparametric Outcome with Additive Noise
Set Identification of ATE (rather than point identification)
이해 못함