Estimating causal treatment effects using Double Machine Learning (DoubleML), Interactive Regression Models (IRM), and modern causal inference techniques. This project demonstrates an end-to-end workflow for estimating the causal effect of answering a Get-Out-The-Vote (GOTV) phone call on voter turnout while addressing selection bias through orthogonalization, cross-fitting, and machine learning.
Estimating causal effects from observational data is substantially more challenging than standard prediction tasks due to confounding and treatment selection. This project applies modern Double Machine Learning (DML) methods to estimate the causal effect of answering a GOTV phone call on voter turnout.
The analysis follows a complete causal inference workflow, including data quality assessment, balance diagnostics, overlap analysis, benchmark estimators, and DoubleML estimation using multiple machine learning learners.
The analysis implements:
- Double Machine Learning (DoubleML)
- Interactive Regression Models (IRM)
- Orthogonal score estimation
- Cross-fitting
- Lasso nuisance learners
- Random Forest nuisance learners
- Average Treatment Effect (ATE)
- Average Treatment Effect on the Treated (ATET)
Benchmark models include:
- Naïve difference in means
- Adjusted Ordinary Least Squares (OLS)
- Data quality assessment
- Exploratory data analysis
- Covariate balance diagnostics
- Propensity score overlap analysis
- Benchmark causal estimators
- DoubleML-IRM estimation
- Comparison of causal estimates
- Interpretation of results
Comparison of benchmark estimators (naïve difference in means and OLS) with Double Machine Learning approaches. The results illustrate how causal effect estimates differ across identification strategies and machine learning learners.
Standardized mean differences between individuals who answered the phone call and those who did not, conditional on being assigned to receive a call (Z = 1). The observed imbalance highlights the presence of treatment selection and motivates the use of Double Machine Learning instead of naïve comparisons.
Estimated propensity score distributions for treated and untreated individuals within the analysis sample. The substantial overlap between groups supports the overlap assumption required for valid estimation using Interactive Regression Models (IRM).
causal-machine-learning-doubleml/
│
├── README.md
├── main.R
├── causal-machine-learning-report.pdf
├── LICENSE
├── .gitignore
├── data/
│ └── mobilization.RData
└── images/
├── estimate-comparison.png
├── overlap.png
└── love-plot.png
- R
- DoubleML
- mlr3
- glmnet
- ranger
- tidyverse
- cobalt
- ggplot2
This project demonstrates how modern machine learning methods can be combined with econometric identification strategies to estimate causal treatment effects from observational data. By combining orthogonalization, cross-fitting, and flexible nuisance models, Double Machine Learning reduces bias from high-dimensional confounding while maintaining valid statistical inference.
A detailed report describing the methodology, empirical analysis, and results is included in the repository:
causal-machine-learning-report.pdf
Robert Puselja