Interactive Tool: Classifying Causal Mechanisms

The Problem

You observe a correlation between X and Y. Someone asks, "Why might this be?" You list several possible explanations. (Scenario is illustrative.)

Example Scenario

"Counties with community health worker (CHW) programs have lower hospitalization rates. What explains this correlation?"

Possible Explanations

1 CHWs prevent hospitalizations. The program itself reduces emergency visits through care coordination and patient education.
2 Well-organized counties do both. Counties with strong public health infrastructure implement CHW programs AND have lower hospitalization rates for unrelated reasons.
3 Low hospitalizations enable CHW programs. Counties with fewer emergencies have more capacity to implement preventive programs.
4 Sick patients leave CHW counties. People who need more hospital care move to counties with better hospital access, leaving CHW counties with healthier populations.

Next: Before we can match explanations to designs, we need to classify them. Each type of causal mechanism requires a different research strategy.

Mechanism Types

Causal explanations fall into four categories. Each represents a different causal structure with different implications for research design.

Alternative Explanation

Confounding

"Well-organized counties do both." A third factor (county capacity) causes both CHW programs and low hospitalizations.

Key question: Is there a common cause driving both X and Y?

Wrong Direction

Reverse Causation

"Low hospitalizations enable CHW programs." The outcome actually causes the treatment, not the other way around.

Key question: Did Y happen before X, or could Y have caused X?

Sample Problem

Selection Bias

"Sick patients leave CHW counties." Who enters or exits the sample depends on both X and Y, creating a spurious correlation.

Key question: Does being in the sample depend on X, Y, or both?

The Target

True Causal Effect

"CHWs prevent hospitalizations." X actually causes Y through a mechanism we can understand and potentially replicate.

Key question: After ruling out alternatives, does X genuinely cause Y?

Why Classification Matters

Each mechanism type requires a different research strategy:

Confounding requires finding variation in X that is independent of the confounder
Reverse causation requires establishing timing or finding X variation that could not be caused by Y
Selection bias requires either changing the sample definition or modeling the selection process
True effects can only be confirmed by ruling out the alternatives

Next: How do you classify an explanation? Use the interactive decision tree to categorize any causal story.

Classifier Tool

Use this decision tree to classify any causal explanation. Answer each question based on the mechanism being proposed.

Q1 Q2 Q3 Result

Question 1 of 3

Does this explanation involve a third variable that affects both the treatment and the outcome?

Look for phrases like "both happen because of..." or "a common factor drives..." or "the same underlying reason causes..."

Question 2 of 3

Does this explanation suggest Y might cause X, rather than X causing Y?

Look for phrases like "enables," "allows," "creates conditions for," or any suggestion that the outcome influences the treatment.

Question 3 of 3

Does this explanation involve who enters or leaves the sample based on the treatment or outcome?

Look for phrases like "people leave," "patients move," "cases drop out," "only those who... are observed," or any suggestion that sample composition depends on X or Y.

Classification

Confounding

This explanation proposes a third factor that causes both the treatment and the outcome. The observed correlation between X and Y may be entirely spurious.

Recommended Design Strategy

Find variation in X that is independent of the confounder. Consider instrumental variables, regression discontinuity, or natural experiments where the confounder does not influence treatment assignment.

Classification

Reverse Causation

This explanation suggests the outcome influences the treatment, not the reverse. The arrow points the wrong direction.

Recommended Design Strategy

Establish timing with panel data, use leads and lags, or find treatment variation that occurred before the outcome could have influenced it. Consider Granger causality tests or event studies.

Classification

Selection Bias

This explanation involves sample composition changing based on treatment, outcome, or both. The correlation may exist only in the observed sample.

Recommended Design Strategy

Redefine the sample to avoid conditioning on post-treatment variables, model the selection process explicitly, or use bounds analysis to estimate the range of possible effects.

Classification

Potential Causal Effect

This explanation proposes that X genuinely causes Y. To confirm, you must rule out confounding, reverse causation, and selection bias.

Recommended Design Strategy

Design studies to rule out each alternative mechanism. Use experimental or quasi-experimental methods that create as-if-random variation in X. The effect is identified only after alternatives are eliminated.

Next: Once you have classified each explanation, match it to a research design. Different mechanisms require different methods.

Design Matching

Each mechanism type has research designs suited to testing it. The table below maps classifications to recommended approaches.

Mechanism	Design Approach	What It Does
Confounding	Instrumental Variables Uses a variable that affects X but not Y directly	Isolates variation in X independent of confounders
Confounding	Regression Discontinuity Compares units just above/below a threshold	Creates as-if-random assignment near cutoff
Confounding	Difference-in-Differences Compares changes in treated vs. control groups	Removes time-invariant confounders
Reverse Causation	Panel Data with Lags Uses past X to predict future Y	Establishes temporal precedence
Reverse Causation	Event Study Tracks outcomes before and after treatment	Shows Y was flat before X changed
Reverse Causation	Granger Causality Tests if past X predicts future Y beyond Y's own history	Distinguishes which variable leads
Selection Bias	Bounds Analysis Estimates range of effects under different assumptions	Shows how much selection could matter
Selection Bias	Heckman Selection Model Jointly models selection and outcome	Corrects for non-random sample
Selection Bias	Intent-to-Treat Analysis Analyzes by initial assignment, not actual treatment	Avoids conditioning on post-treatment behavior
True Effect	RCT (Gold Standard) Randomly assigns treatment	Eliminates all confounding by design
True Effect	Natural Experiment Exploits as-if-random variation from real-world events	Approximates random assignment

Before Choosing a Design

> Have you listed all plausible mechanisms?

> Have you classified each one?

> Does your design address the most likely threat?

> Can you explain what variation you are using?

> What threats does your design NOT address?

The 80/20 Rule

Most observational studies fail because of confounding. Designs that address confounding (IV, RD, DiD) will solve most problems. Reverse causation and selection bias matter, but less often.

No design solves everything. Be explicit about what threats your chosen design addresses and what threats remain.

Next: The key insight ties it all together. Classification is the first step; the goal is identification.

Key Insight

Classifying causal mechanisms is not the end goal. It is the first step toward designing research that can distinguish between competing explanations.

Flowchart showing the classification process: from observed correlation through mechanism classification to matched research design, with four mechanism types (confounding, reverse causation, selection bias, true effect) each pointing to specific design strategies

Classification transforms a list of possible explanations into a research strategy. Each mechanism type maps to designs that can test it.

The Classification-to-Design Workflow

The economist's approach to observational data follows a systematic process:

List: Write down every plausible explanation for the observed correlation
Classify: Sort each explanation into confounding, reverse causation, selection, or true effect
Match: Identify research designs that can distinguish between the classified mechanisms
Design: Choose the design that addresses the most plausible threat given available data
Acknowledge: Be explicit about which threats your design cannot address

Questions That Drive Better Research

What would make this explanation false? Every mechanism should be testable in principle.

What variation am I using? The source of treatment variation determines what confounders are ruled out.

What is my identifying assumption? State explicitly what must be true for your design to work.

What threats remain? No design addresses all mechanisms. Honest research acknowledges limitations.

Key Takeaway

Correlation invites multiple explanations; research design narrows them down. A statistics-focused approach would be to list possible confounders and control for them statistically. Economics classifies explanations by type and designs studies that can rule out each category. The goal is not to find the "right" adjustment, but to find variation in treatment that is independent of the threats you have classified. This is what economists mean by "identification strategy."