Seven Fundamental Rules for Causal Inference: A Practitioner’s Guide

In the world of data analysis, understanding causal relationships is crucial. This article breaks down seven fundamental rules that govern how real-world causal mechanisms translate into observable associations in your data. These rules serve as building blocks for more complex causal inference techniques and are essential for anyone working with or learning about causal inference.

Four Fundamental Causal Structures

Before diving into the rules, let's review the four basic causal structures that form the foundation of these principles:

1. Complete Independence

   • No path can be traced between variables A and B.
   • Example: The color of a car (A) and the taste of an apple (B) are completely independent.

2. Chain

   • A directed path can be traced from A to B, with all arrows pointing from A to B.
   • Example: Smoking (A) → Lung Cancer (M) → Shorter Lifespan (B). Here, lung cancer is a mediator.
   • Chains are "open" and transmit correlation between A and B.

3. Fork

   • An undirected path can be traced from A to B through a common causal ancestor C.
   • Example: Genetics (C) → Height (A) and Genetics (C) → Weight (B). Here, genetics is a confounder.
   • Forks are "open" and transmit correlation between A and B.

4. Collider

   • An undirected path can be traced from A to B through a causal descendant D.
   • Example: Intelligence (A) and Study Habits (B) → Test Scores (D). Here, test scores are a collider.
   • Colliders are "closed" by default but can become "open" if conditioned on the collider or its descendants.

The Seven Rules of Causal Inference

1. Rule 1: Association Does Not Imply Causation

   • Just because A and B are correlated does not mean A causes B.
   • Example: Ice cream sales (A) and drowning incidents (B) are correlated, but neither directly causes the other. Both are influenced by a common ancestor: warm weather.

2. Rule 2: Conditioning on Mediators

   • Conditioning on a mediator in a chain can block the path between A and B.
   • Example: If you control for lung cancer (M), the direct correlation between smoking (A) and shorter lifespan (B) is reduced or eliminated.
   • Implementation Note: In regression, this means including M as an independent variable.

3. Rule 3: Conditioning on Confounders

   • Conditioning on a confounder in a fork can block the path between A and B.
   • Example: If you control for genetics (C), the correlation between height (A) and weight (B) is reduced or eliminated.
   • Implementation Note: In regression, this means including C as an independent variable.

4. Rule 4: Conditioning on Colliders

   • Conditioning on a collider can open a path between A and B that was previously closed.
   • Example: If you control for test scores (D), the correlation between intelligence (A) and study habits (B) might increase.
   • Implementation Note: Be cautious when including colliders in your models, as this can introduce spurious correlations.

5. Rule 5: Conditioning on Descendants of Colliders

   • Conditioning on a descendant of a collider can also open a path between A and B.
   • Example: If you control for a variable influenced by test scores (e.g., college admissions), the correlation between intelligence (A) and study habits (B) might increase.
   • Implementation Note: This is similar to Rule 4 but applies to variables downstream from the collider.

6. Rule 6: Backdoor Paths

   • Confounding can occur through backdoor paths, which are indirect causal paths that introduce bias.
   • Example: In the fork structure (Genetics → Height and Genetics → Weight), there is a backdoor path from Height to Weight through Genetics.
   • Implementation Note: Use techniques like backdoor adjustment to control for confounders.

7. Rule 7: Frontdoor Paths

   • Mediation can occur through frontdoor paths, which are indirect causal paths that transmit the effect of A on