Visualizing Causality: A Guide to Directed Acyclic Graphs (DAGs) in Epidemiological Causal Inference

Directed Acyclic Graphs (DAGs) are a powerful conceptual framework used in epidemiology to formalize and visualize assumptions about causal relationships between variables. By representing variables as nodes and direct causal effects as arrows, DAGs provide a clear, non-parametric language to map out the complex web of factors influencing an exposure-outcome relationship. The core utility of a DAG is to help researchers distinguish causal pathways from non-causal (spurious) associations that arise from bias. The main sources of bias, confounding and selection bias, have distinct graphical representations. Confounding arises from 'backdoor paths'—non-causal connections created by common causes (forks). These paths can be blocked by statistically adjusting for the confounding variable. In contrast, selection bias often results from conditioning on a 'collider' (a common effect), which paradoxically creates a spurious association where none existed. By applying a set of graphical rules known as d-separation, researchers can use a DAG to identify the minimally sufficient set of covariates that must be adjusted for to isolate a causal effect, while also identifying variables that should not be adjusted for (such as colliders or mediators) to avoid introducing bias or blocking the effect of interest. Ultimately, DAGs are an indispensable tool for designing more robust observational studies and for the critical interpretation of their results.

1. Introduction: The Challenge of Causal Inference

In epidemiology, a central goal is to determine if an exposure (e.g., a medication, a lifestyle factor, an environmental pollutant) causes an outcome (e.g., a disease, recovery, death). In an ideal world, we would use a Randomized Controlled Trial (RCT) to answer this question. Randomization ensures that, on average, the groups being compared are similar in all respects except for the exposure they receive. However, RCTs are often unethical (e.g., randomizing people to smoke), impractical, or too expensive.

Consequently, epidemiologists rely heavily on observational studies (cohort, case-control, etc.). The major challenge in these studies is confounding: a mixing of effects where the association between an exposure and an outcome is distorted by a third variable. Directed Acyclic Graphs (DAGs) have emerged as an essential tool to help researchers think systematically and transparently about causal relationships, identify potential sources of bias like confounding, and determine the best analytical strategy to mitigate them.

2. What is a Directed Acyclic Graph (DAG)?

A DAG is a visual representation of the assumed causal relationships between a set of variables. It is a formal, non-parametric framework that forces researchers to make their assumptions explicit. Let's break down the name:

• Graph: It's a collection of nodes (or vertices) connected by edges (or arrows).

- Nodes: Represent variables of interest (e.g., Exposure, Outcome, Age, Sex, Socioeconomic Status). - Edges: Represent direct causal effects. An arrow from variable A to variable B (A → B) signifies that A is a direct cause of B. The absence of an arrow is a very strong assumption: it asserts that there is no direct causal effect.

• Directed: The edges have a direction, indicated by an arrowhead. This represents the hypothesized flow of causality (e.g., Smoking → Lung Cancer, not Lung Cancer → Smoking).
• Acyclic: There are no feedback loops. You cannot start at a node, follow the arrows, and end up back at the same node. This reflects the common-sense notion that a variable cannot be its own cause (e.g., you can't have A → B → C → A).

3. The Three Building Blocks of Causal Paths

All relationships within a DAG can be broken down into three fundamental structures that determine how information (or statistical association) flows between variables.

a) The Chain (Mediation):

• Structure: A → M → B
• Description: The effect of A on B is mediated through M. A causes M, which in turn causes B.
• Example: Smoking (A) → Tar Deposition in Lungs (M) → Lung Cancer (B).
• Path Status: The path is open by default, meaning A and B will be statistically associated. If you condition on the mediator M (e.g., by stratifying your analysis by levels of tar deposition), the path becomes blocked. This is because, within a given level of M, any remaining variation in A is no longer associated with B.

b) The Fork (Confounding):

• Structure: A ← C → B
• Description: C is a common cause of both A and B. This is the classic structure of confounding.
• Example: Let's test the hypothesis that drinking coffee (A) causes heart disease (B). It's known that many coffee drinkers are also smokers, and smoking (C) causes heart disease. The DAG would be: Coffee (A) ← Smoking (C) → Heart Disease (B). The variable Smoking (C) is a confounder.
• Path Status: The path is open by default, creating a non-causal (spurious) association between A and B. To estimate the true causal effect of A on B, you must block this "backdoor path" by conditioning on the confounder C.

c) The Collider:

• Structure: A → Z ← B
• Description: Z is a common effect of both A and B. A and B both cause Z.
• Example: Having a specific gene variant (A) might cause a disease, and having a poor diet (B) might also cause that same disease. The disease itself (Z) is a collider. In this structure, A and B are independent and not associated by default.
• Path Status: The path is closed by default. A and B are not associated through Z. However, if you condition on the collider Z (or a descendant of Z), you open the path and create a spurious, non-causal association between A and B. This is a critical and often counter-intuitive concept.

4. Using DAGs to Identify and Control for Bias

The primary use of a DAG in study design and analysis is to identify the set of variables that must be controlled for (or "adjusted for") to obtain an unbiased estimate of the causal effect of an exposure on an outcome.

a) Identifying Confounding (Backdoor Paths): An association between an exposure (E) and an outcome (D) can be due to the causal path (E → D) or a backdoor path. A backdoor path is a non-causal path between E and D that begins with an arrow pointing into E.

• Example: In our coffee/smoking/heart disease example (Coffee ← Smoking → Heart Disease), the path Smoking → Heart Disease is a backdoor path from Coffee to Heart Disease because it starts with an arrow pointing into Coffee.
• The Goal: To estimate the causal effect of E on D, we must block all open backdoor paths. We do this by conditioning on (adjusting for) one or more variables along each backdoor path.
• The Rule: A set of variables Z is a sufficient adjustment set if it blocks all backdoor paths from the exposure to the outcome. Crucially, this set Z must not contain any colliders that would be opened by conditioning.

b) Understanding Selection Bias (Collider-Stratification Bias): Selection bias occurs when the selection of subjects into a study (or into the final analysis) is dependent on both the exposure and the outcome. In a DAG, this is almost always represented by conditioning on a collider.

• Example (Berkson's Bias): Suppose we are studying the association between two diseases, Cholecystitis (A) and Arthritis (B), using only hospital patients. Both diseases increase the probability of hospitalization (Z). The structure is: Cholecystitis (A) → Hospitalization (Z) ← Arthritis (B). By restricting our study to hospitalized patients, we are conditioning on the collider (Z). This will create a spurious association between cholecystitis and arthritis among hospital patients, even if no such association exists in the general population.

5. A Step-by-Step Practical Application

Let's assess the causal effect of physical activity (E) on depression (D), considering age (A) and diet quality (Q).

1. Draw the DAG based on a priori knowledge:

Age (A) influences physical activity (E) and depression (D). (A → E; A → D). This makes Age a classic confounder (Fork: E ← A → D). Physical activity (E) may improve diet quality (Q). (E → Q). Diet quality (Q) affects depression (D). (Q → D). We hypothesize physical activity (E) directly affects depression (D). (E → D). * The resulting DAG has two causal paths from E to D: the direct path (E → D) and an indirect, mediated path (E → Q → D). It also has one backdoor path: E ← A → D.

1. Identify Causal and Non-Causal Paths:

Causal Paths: E → D (direct); E → Q → D (indirect). We want to estimate the total effect from both of these. Backdoor Path: E ← A → D. This path is open and will create confounding bias.

1. Identify the Adjustment Set:

To get an unbiased estimate of the total causal effect of E on D, we must block the backdoor path E ← A → D. We can do this by conditioning on the confounder, Age (A). We should not* condition on Diet Quality (Q), because it is a mediator. Adjusting for a mediator would block a part of the causal effect we want to measure.

1. Conclusion: To estimate the total causal effect of physical activity on depression, we should adjust for age in our statistical model (e.g., via regression or stratification) but not for diet quality.

Browse public study notes