auto_causal_inference/
├── agent/                 # LangGraph agent source code
│   ├── data/              # Sample data (bank.db)
│   ├── app.py             # Main entry point for LangGraph causal agent
│   ├── generate_data.py   # Data generation script for causal inference
│   ├── requirements.txt   # Dependencies for LangGraph agent
│   └── ...                # Other helper modules and notebooks
│
├── mcp_agent/             # MCP agent implementation
│   ├── data/              # Sample data (bank.db)
│   ├── server.py          # MCP causal inference server
│   ├── client.py          # MCP client to call the causal inference server
│   ├── requirements.txt   # Dependencies for MCP agent
│   └── ...                # Additional files
│
└── README.md              # This documentation file

pip install requirements.txt

cd agent
python app.py

langgraph dev

cd mcp_agent
python client.py

User asks: "Does offering a promotion increase digital product activation ?"

age -> promotion_offer;
age -> activated_ib;
income -> promotion_offer;
income -> activated_ib;
education -> promotion_offer;
education -> activated_ib;

region_code -> promotion_offer;

promotion_offer -> branch_visits;
branch_visits -> activated_ib;

promotion_offer -> customer_engagement;
activated_ib -> customer_engagement;

channel_preference -> activated_ib;
promotion_offer -> activated_ib

{
  "confounders": ["age", "income", "education"],
  "mediators": ["branch_visits"],
  "effect_modifiers": ["channel_preference"],
  "colliders": ["customer_engagement"],
  "instruments": ["region_code"],
  "causal_graph": "...DOT format...",
  "dowhy_code": "...Python code..."
}

import dowhy
from dowhy import CausalModel

model = CausalModel(
    data=df,
    treatment='promotion_offer',
    outcome='activated_ib',
    common_causes=['age', 'income', 'education'],
    instruments=['region_code'],
    mediators=['branch_visits']
)

identified_model = model.identify_effect()
estimate = model.estimate_effect(identified_model, method_name='backdoor.propensity_score_matching')
print(estimate)

estimators = ["S-learner", "T-learner", "X-learner"]
# base_learners = ["random_forest", "neural_network"]

cd = CausalityDataset(data=df, treatment=state['treatment'], outcomes=[state["outcome"]],
                    common_causes=state['confounders'])
cd.preprocess_dataset()

estimators = ["SLearner", "TLearner"]
# base_learners = ["random_forest", "neural_network"]

ct = CausalTune(
    estimator_list=estimators,
    metric="energy_distance",
    verbose=1,
    components_time_budget=10, # in seconds trial for each model
    outcome_model="auto",
)

# run causaltune
ct.fit(data=cd, outcome=cd.outcomes[0])

print(f"Best estimator: {ct.best_estimator}")
print(f"Best score: {ct.best_score}")

refute_results = []
refute_methods = [
    "placebo_treatment_refuter",
    "random_common_cause",
    "data_subset_refuter"
]
for method in refute_methods:
    refute = model.refute_estimate(identified_estimand, estimate, method_name=method)
    refute_results.append({"method": method, "result": str(refute)})

pass_test = all("fail" not in r["result"].lower() for r in refute_results)

| Role                | Variable                     | Why it's assigned this role                                      |
| ------------------- | ---------------------------- | ---------------------------------------------------------------- |
| **Confounder**      | `age`, `income`, `education` | Affect both the chance of receiving promotions and IB usage.     |
| **Mediator**        | `branch_visits`              | A step in the causal path: promotion → visit → IB activation.    |
| **Effect Modifier** | `channel_preference`         | Alters the strength of the effect of promotion on IB activation. |
| **Collider**        | `customer_engagement`        | Affected by both promotion and IB usage; should not be adjusted. |
| **Instrument**      | `region_code`                | Randomized promotion assignment at the regional level.           |

Best estimator: backdoor.econml.metalearners.TLearner, score: 483.1930697900207

Refutation passed: True.
[   
    {'method': 'placebo_treatment_refuter', 
    'result': 'Refute: Use a Placebo Treatment Estimated effect:0.23849549989874572
                New effect:-0.0004960408910311281
                p value:0.96'}, 
    {'method': 'random_common_cause', 
    'result': 'Refute: Add a random common cause
                Estimated effect:0.23849549989874572
                New effect:0.23847067700750038
                p value:0.98'}, 
    {'method': 'data_subset_refuter', 
    'result': 'Refute: Use a subset of data
                Estimated effect:0.23849549989874572
                New effect:0.23749715031525756
                p value:0.96'}
]

Result Summary:
1. There is a causal effect between offering promotions and activating internet banking services, with a 15% increase of activating internet banking if we open the promotion for everybody. This shows a strong positive impact of the promotion offer on activation.

2. Factors like age, income, education level could have influenced both the decision to offer promotions and the likelihood of activating internet banking services. These factors may have affected the outcome regardless of the promotion offer.