Show & Tell 40.4 — CausalInferenceOptimizer: DML, Matching & Quasi-Experimental Methods

[web3guru888]

CausalInferenceOptimizer — Show & Tell

Implementation Highlights

The CausalInferenceOptimizer provides a comprehensive toolkit of statistical estimation methods for causal inference.

Double/Debiased Machine Learning (DML)

Following Chernozhukov et al. (2018), the DML estimator achieves √n-consistency even with high-dimensional nuisance parameters estimated by ML:

1. Split data into K folds
2. For each fold k:
   a. Train outcome model: μ̂(X) = E[Y|X] on data \ fold k
   b. Train propensity model: ê(X) = E[T|X] on data \ fold k
   c. Compute residualized scores on fold k:
      Ỹ = Y - μ̂(X)
      T̃ = T - ê(X)
3. Estimate θ = (Σ T̃·Ỹ) / (Σ T̃·T̃)   — Neyman orthogonal score

Key insight: Cross-fitting (step 1-2) prevents overfitting bias; Neyman orthogonality ensures the estimator is insensitive to first-stage estimation errors.

Method Selection Guide

Method	When to Use	Key Assumption
IPW	Known treatment mechanism	Correct propensity
Doubly Robust	Either model may be wrong	One model correct
DML	High-dimensional covariates	Sparsity / ML works
Matching	Small-sample, visual inspection	Common support
Synthetic Control	Panel data, single treated unit	No spillover
DiD	Panel data, multiple units	Parallel trends
RDD	Threshold-based treatment	Continuity at cutoff

Related: #812