Bayesian additive regression trees for causal inference with multiple treatments

Michael Lopez (with Liangyuan Hu, Chenyang Gu)

BART with multiple treatments

Michael Lopez, Liangyuan Hu, Chenyang Gu https://github.com/statsbylopez/ci-bart

3 lung cancer treatments

Video assisted (VATS)

3 lung cancer treatments

Open thorectomy

3 lung cancer treatments

Robotic assisted

3 lung cancer treatments

Variable robotic VATS open.thoracotomy
Pct white 0.808 0.865 0.873
Pct high income 0.245 0.324 0.202
Pct Stage 3 cancer 0.096 0.134 0.082

3 lung cancer treatments

Variable robotic VATS open.thoracotomy
Pct white 0.808 0.865 0.873
Pct high income 0.245 0.324 0.202
Pct Stage 3 cancer 0.096 0.134 0.082
Complications 0.308 0.357 0.303

3 lung cancer treatments

Issue 1: Selection bias

Issue 2: Non-overlapping distributions

Issue 3: Large weights

Notation

Consider causal effect of \(A \in \{1, \ldots, Z\}\) on binary outcome \(Y \in \{0,1\}\)

Notation

Interest: average treatment effect among treated

Causal inference with multiple treatments

Why not binary approaches?

Causal inference with multiple treatments

Bayesian Additive Regression Trees

Why BART for causal inference? see Hill, 2012

Bayesian Additive Regression Trees

Why BART for multiple treatments?

Simulation study

6 factorial design using dbarts package in R

3 lung cancer treatments

ATT’s: generalizable to population receiving robotic treatment

Comments

Comments

Github site: https://github.com/statsbylopez/ci-bart

Twitter: @statsbylopez