Causal Inference Ml

Causal Inference in ML

Applying causal inference methods to machine learning — from causal graphs through treatment effect estimation, counterfactual reasoning, and decision-making.

When to Use

• Determining whether X causes Y (not just correlates)
• Estimating treatment effects from observational data
• Removing confounding bias from ML models
• Answering "what if" counterfactual questions

Causal Concepts

CAUSAL_CONCEPTS = {
    'association': 'P(Y|X) — statistical correlation',
    'intervention': 'P(Y|do(X)) — causal effect of intervening',
    'counterfactual': 'P(Y_{X=x} | X=x') — what if X had been different',
}

Causal Graph (DAG)

class CausalGraph:
    def __init__(self):
        self.nodes = set(); self.edges = []
    
    def add_edge(self, cause: str, effect: str):
        self.nodes.update([cause, effect]); self.edges.append((cause, effect))
    
    def get_confounders(self, x: str, y: str) -> list:
        x_parents = set(p for p, c in self.edges if c == x)
        y_parents = set(p for p, c in self.edges if c == y)
        return list(x_parents & y_parents)

Treatment Effect Estimation

import numpy as np
from sklearn.linear_model import LogisticRegression

def propensity_score_matching(treatment, outcome, features):
    ps_model = LogisticRegression().fit(features, treatment)
    propensity = ps_model.predict_proba(features)[:, 1]
    
    treated = treatment == 1; control = treatment == 0
    effects = []
    for i in np.where(treated)[0]:
        nearest = np.where(control)[0][np.argmin(np.abs(propensity[control] - propensity[i]))]
        effects.append(outcome[i] - outcome[nearest])
    
    return {'ate': np.mean(effects), 'matched_pairs': len(effects)}

Common Pitfalls

1. Correlation ≠ causation — always consider confounders
2. Conditioning on colliders — opens spurious paths
3. Selection bias — non-random treatment assignment; use matching
4. Hidden confounders — unmeasured variables bias estimates

Verification Checklist

• Causal DAG drawn and reviewed
• Confounders identified and adjusted for
• Backdoor criterion satisfied
• Sensitivity analysis for unmeasured confounding

See Also

• ab-testing-experimentation — randomized experiments
• deep-reinforcement-learning — causal effects in RL
• ml-pipeline-design — causal inference in pipelines