🔍 CausalLens

LLM-Augmented Causal Inference Pipeline

Causal question answering over tabular data. Combines DoWhy, EconML Double ML, and the Claude API to produce rigorous ATE estimates and business-readable reports — end to end.

Pipeline

Raw CSV  →  Causal DAG  →  ATE Estimation  →  Refutation Tests  →  LLM Report
              (DoWhy)     PSM · IPW · DML      Placebo · Boot      Claude API
                           Causal Forest

✨ What It Does

Step	Method	Library
1. Causal DAG construction	Domain-knowledge graph	networkx · DoWhy
2. ATE estimation × 4 methods	PSM · IPW · Double ML · Causal Forest	dowhy · econml
3. Robustness checks	Placebo · Random confounder · Bootstrap	dowhy
4. LLM graph interpretation	DAG → plain English	anthropic Claude API
5. Business report generation	Full markdown report	anthropic Claude API
6. Interactive UI	Dataset explorer + visualisations	streamlit

📦 Datasets

1. Lalonde NSW Job Training

• Source: Lalonde (1986) — classic RCT benchmark in causal inference
• Treatment: treat — participated in subsidised job training
• Outcome: re78 — 1978 earnings (USD)
• Ground truth ATE: ~$886 (from the original experiment)
• Use case: Validate that observational methods recover the experimental truth

2. IBM HR Employee Attrition

• Source: IBM Watson Analytics / AIF360 (n = 1,470)
• Treatment: HighIncome — monthly income above median
• Outcome: Attrition_num — employee left the company (1/0)
• Use case: Estimate causal impact of compensation on retention

🚀 Quickstart

# 1. Clone / unzip
cd CausalLens

# 2. Install dependencies
pip install -r requirements.txt

# 3. Set your Anthropic API key
cp .env.example .env
# edit .env: ANTHROPIC_API_KEY=sk-ant-...

# 4a. Run the Streamlit app
streamlit run app/streamlit_app.py

# 4b. Or run the notebooks
jupyter notebook notebooks/

📁 Project Structure

CausalLens/
├── data/
│   ├── __init__.py
│   ├── load_data.py          # auto-fetch IBM + Lalonde datasets
│   └── cache/                # local CSV cache (auto-created)
├── causal/
│   ├── __init__.py
│   ├── discovery.py          # build & visualise causal DAGs
│   ├── effect_estimation.py  # PSM · IPW · Double ML · Causal Forest
│   └── refutation.py         # placebo · random confounder · bootstrap
├── llm/
│   ├── __init__.py
│   ├── graph_interpreter.py  # LLM explains the causal DAG
│   └── report_generator.py   # LLM writes a full business report
├── app/
│   └── streamlit_app.py      # interactive Streamlit UI
├── notebooks/
│   ├── 01_lalonde_analysis.ipynb
│   └── 02_ibm_attrition_analysis.ipynb
├── .env.example
├── requirements.txt
└── README.md

🧠 Methods Detail

Propensity Score Matching (PSM)

Matches each treated unit to the most similar control unit based on the estimated propensity score P(T=1|X). Implemented via DoWhy's backdoor estimator.

Inverse Probability Weighting (IPW)

Re-weights the sample so that treated and control groups have the same covariate distribution. Uses stabilised IPS weights.

Double / Debiased Machine Learning (LinearDML)

Removes the effect of confounders via two nuisance models:

• model_y: GBM predicts outcome from confounders → residual Ỹ
• model_t: GBM predicts treatment from confounders → residual T̃

Then regresses Ỹ on T̃ to obtain the ATE. Provides valid confidence intervals via cross-fitting (5-fold CV).

Causal Forest DML

Non-parametric extension of Double ML using a causal forest as the final stage estimator. Estimates heterogeneous treatment effects (CATE) and reports the average.

Refutation Tests

Test	What it checks
Placebo Treatment	Replace treatment with noise → ATE should collapse to ≈ 0
Random Common Cause	Add a spurious confounder → ATE should be stable
Bootstrap (Data Subset)	Re-estimate on 80 % subsets → ATE should be stable

📝 Sample LLM Report Output

## Executive Summary
Participation in the NSW job-training programme causally increased 1978
earnings by approximately $921 (Double ML, 95% CI [$412, $1,430]).  This
estimate is consistent across PSM ($876), IPW ($904), and Causal Forest
($935), and passes all three refutation tests, giving high confidence
that the effect is not a statistical artefact.

## Business Implications
Policy-makers should continue funding similar job-training programmes...

📊 Resume Bullet

\item LLM-augmented causal inference pipeline on IBM HR Attrition and
Lalonde datasets; combined DoWhy (PSM, IPW) with EconML Double ML and
Causal Forest (GBM nuisance models) for ATE estimation and refutation
testing (placebo, random confounder, bootstrap); Claude API
auto-generates business reports from causal estimates via a
planner--estimator--reporter workflow; deployed as an interactive
Streamlit app.

🔑 API Key

The LLM features (graph interpretation + report generation) require an Anthropic API key. All causal estimation runs without one.

Get a free key at: https://console.anthropic.com

📚 References

• Lalonde, R. J. (1986). Evaluating the econometric evaluations of training programs.
• Pearl, J. (2009). Causality: Models, Reasoning, and Inference.
• Chernozhukov et al. (2018). Double/debiased machine learning.
• DoWhy: https://github.com/py-why/dowhy
• EconML: https://github.com/py-why/EconML