Causality, Experimental Design, and Intervention Science

Manual: General · Subject: General Science

This lesson explores how scientists identify causal structure, design interventions, and distinguish mechanisms from mere correlation.

From Association to Causation

Causal questions in science

Science often seeks interventions: what happens if we change a variable? Causal inference formalizes this through counterfactual reasoning, structural models, and experimental or quasi-experimental designs. The central problem is that observed data alone do not automatically reveal causal direction.

Core causal concepts

Confounder

A variable influencing both treatment and outcome.

Mediator

A variable on the causal pathway between treatment and outcome.

Collider

A variable influenced by two causes; conditioning on it can induce bias.

Counterfactual

The outcome that would have occurred under an alternative intervention.

Why are randomized experiments so powerful?

They tend to balance confounders across groups They always eliminate measurement error They guarantee mechanistic understanding They remove the need for sample size calculations

DAGs and identification

Directed acyclic graphs (DAGs) provide a language for encoding assumptions about causal structure. They help scientists reason about adjustment sets, backdoor paths, colliders, and identification conditions. A DAG is not itself proof of causality; rather, it is a disciplined representation of causal assumptions.

What is a collider in causal inference?

Experimental and observational strategies

Randomized controlled trial

Highest internal validity when feasible
Can be expensive or ethically impossible
May have limited external validity

Observational study

Often more feasible and scalable
Requires careful adjustment for confounding
Typically more vulnerable to bias

💡

Research frontier

Current methods include causal discovery from time series, transportability across populations, synthetic controls, and double/debiased machine learning for high-dimensional settings.

Designing a causal study

1

Step 1: State the causal effect of interest.
2

Step 2: Draw a causal diagram and identify threats to validity.
3

Step 3: Choose between experiment, quasi-experiment, or observational design.
4

Step 4: Predefine adjustment variables and exclusion criteria.
5

Step 5: Perform sensitivity analysis for hidden bias.

Conditioning on a collider can:

Introduce spurious association Guarantee unbiased estimation Increase randomization Remove all confounding automatically

Neuroanatomy Explorer

Memory Deck

Match the Pairs

Concept Constellation