IS 507 · Lecture 12 · Yonghan Jung

Causal Inference - I

A simulated hospital: does Remdesivir save lives?

drug (A = 1)no drug (A = 0)truth / structurebias
X = wealthy (1) or not (0) · A = got Remdesivir · Y = died within 28 days
P(X=1) = 0.5 · P(A=1 | X) = 0.8 wealthy, 0.2 not wealthy · P(Y=1 | A, X) = 0.23 − 0.18·X + 0.02·A
1Simpson

Drug patients die less. Does the drug help?

all patients
wealthy
not wealthy
Death rate (top) and share wealthy per arm (bottom).
2Suzy & Billy

But-for fails. Hold BH fixed.

Suzy throws (ST)
Billy throws (BT)
but-for: change ST, BS changes?
holding BH = F: change ST, BS changes?
3Association vs causation

Who gets the drug decides what we compare

treateduntreatedwealthy (top half)
causation
E[Y(1)] − E[Y(0)]
association
E[Y | A=1] − E[Y | A=0]
4A/B test

A coin flip balances the arms

Estimates over 500 samples. Navy = truth +0.02.
5Same data

Two models. Same data. Opposite answers.

SCM 1 · drug does nothing
U ~ Bern(0.5) · A = U · Y = 2·U + ε
observed Y, by A
SCM 2 · drug adds 2
A ~ Bern(0.5) · Y = 2·A + ε
observed Y, by A

Same data.

Which one made your data? ?

6Conditional RCT

A different coin for each age

X = old (≥ 60) or young · P(old) = 0.5 · P(Y=1 | A, old) = 0.30 + 0.02·A · P(Y=1 | A, young) = 0.08 + 0.02·A
naive E[Y|A=1] − E[Y|A=0]
adjusted Σₓ (E[Y|A=1,x] − E[Y|A=0,x]) P(x)
truth+0.020
7Ignorability

Adjusting for X fails if U is hidden

X A Y U
adjusted for X
truth+0.020
bias
Bottom: Y(1) rate, treated vs untreated, within X. Equal bars = Y(a) ⫫ A | X.
8Estimators

IPW · Outcome regression · AIPW

X ~ N(0,1) health score · P(A=1|x) = sigmoid(−x + 0.5(x² − 1)) · Y = 10 + 4x + 3x² + 5·A + N(0, 4²) · true ATE = 5
view
propensity model
outcome model
9Double robustness

AIPW: right if either model is right

propensity model
outcome model
Average bias (estimate − 5). Rust = |bias| > 0.3.
10Errors multiply

Two rough models, one sharp estimate

IPW bias
OR bias
AIPW bias
bias(AIPW) ≈ error(ê) × error(μ̂)
|bias| vs δ, exact over x ~ N(0,1).