causal-inference

Causal Inference

Overview

Causal inference provides methods for estimating causal effects rather than merely correlational associations. Use this skill whenever you need to claim that X causes Y, not just that X and Y are correlated.

When to Use This Skill

• When your research question is causal ("Does X increase Y?")
• When you have observational data and want to make causal claims
• When designing a study and choosing between experimental and observational approaches
• When a reviewer asks about confounding, endogeneity, or causal identification

When NOT to Use This Skill

• When your goal is purely predictive (correlations are sufficient for prediction)
• When you have a true RCT with perfect compliance (standard t-test is enough)

---

Hard Rules

1. Never use causal language (causes, increases, reduces) without causal identification — observational correlation alone is insufficient
2. Draw the DAG before choosing a method — the causal graph determines what you can and cannot control for
3. Conditioning on a collider opens a backdoor path — it can create spurious associations
4. Always run a placebo test — if your method claims an effect where no effect should exist, the method is flawed

---

1. Causal DAGs (Directed Acyclic Graphs)

Drawing the causal graph is the first step before any analysis.

Node types:

• Treatment (T): the variable you want to study
• Outcome (Y): the variable you want to affect
• Confounder (C): common cause of T and Y — must control for
• Mediator (M): on the causal path T → M → Y — do NOT control for if you want total effect
• Collider (K): caused by both T and Y (or their descendants) — do NOT condition on

# Simple DAG visualization with networkx
import networkx as nx
import matplotlib.pyplot as plt

G = nx.DiGraph()
G.add_edges_from([
    ("Education", "Income"),      # T → Y (causal path)
    ("Family_SES", "Education"),  # C → T (confounder)
    ("Family_SES", "Income"),     # C → Y (confounder)
])
pos = nx.spring_layout(G, seed=42)
nx.draw(G, pos, with_labels=True, node_color="lightblue",
        node_size=2000, arrows=True, arrowsize=20)