Causal Inference with DAGs

Causality

High
Causal DAGs use directed acyclic graphs to represent assumptions about what causes what. They are powerful because they make hidden causal structure visible, especially confounding, mediation, and what should or should not be controlled for.
Reasoning type
Causal-structural
Certainty level
Assumption-dependent
Cognitive load
High
Formality
High

Core Idea

Definition

A causal DAG is a graph of variables connected by directional arrows that encode assumed causal relationships and help determine how causal effects can be identified from data.

In Plain English

Draw the cause structure first, then use that map to decide what comparisons are informative and which ones are misleading.

Framework Structure

Components

Variables
Directed Causal Arrows
Confounders and Mediators
Adjustment Strategy

Flow

List variables -> Draw assumed causal directions -> Identify backdoor paths and confounding -> Decide what to condition on or avoid

How to Apply

  • 1.List the key variables related to the causal question
  • 2.Draw arrows for your best current assumptions about causal direction
  • 3.Identify confounders, mediators, and colliders in the graph
  • 4.Use the DAG to decide what should be adjusted for and what should not
  • 5.Revise the graph when new domain knowledge challenges the structure

When to Use

  • Causal analysis with observational data
  • Designing studies or experiments
  • Clarifying confounding and identification assumptions
  • Policy, product, or analytics questions where correlation is not enough
  • Any setting where people are arguing past each other about what the real mechanism is

When NOT to Use

  • When no defensible causal assumptions can be articulated
  • When the graph is being treated as truth rather than a representation of assumptions
  • When the question is purely predictive rather than causal
  • When the structure is so oversimplified that it hides the very problem you need to analyze

Example

Problem

An analyst wants to know whether a customer success program reduces churn.

Application

  • 1.Map variables such as program participation, baseline customer quality, product usage, and churn
  • 2.Draw arrows to represent plausible causal relationships
  • 3.Notice that customer quality may confound both program participation and churn
  • 4.Use the DAG to decide what needs adjustment before estimating the causal effect

Conclusion

The DAG improves the analysis by revealing what comparisons are biased before any regression is run.

Takeaway

Causal graphs sharpen thinking by making assumptions inspectable instead of leaving them implicit.

Common Mistakes

  • Drawing arrows casually without domain knowledge
  • Conditioning on colliders and introducing bias
  • Treating the graph as data rather than as a model of assumptions
  • Ignoring omitted variables that materially affect the conclusion
  • Confusing a helpful simplification with complete causal reality

How to Practice

arrow justification

For each arrow in a causal graph, state why you believe the direction makes sense.

backdoor scan

Practice identifying alternate causal paths that could confound the effect you care about.

graph revision

After new evidence or feedback, redraw the DAG rather than defending the original map.

Related Cognitive Biases

illusory correlation

Without explicit causal structure, correlated variables can be mistaken for direct causes.

control illusion

People often believe adjustment for more variables is always better, even when it can create bias.

oversimplification

A graph can look clean while still omitting an important driver of the system.

Related Frameworks

Related Skills

identifying components
fact inference separation
comparing evidence
systems thinking

Variants & Extensions

Backdoor adjustment reasoning
Collider and confounder analysis
Structural causal graphing
Identification mapping

Typical Failure Modes

  • Bad graph assumptions
  • Collider bias
  • Missing key variables

Further Reading

  • The Book of Why by Judea Pearl and Dana Mackenzie
  • Causal Inference in Statistics by Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell
  • Mostly Harmless Econometrics by Joshua D. Angrist and Jörn-Steffen Pischke