Why is it called the 'fundamental problem' of causal inference?

Because for any unit only the outcome under the exposure actually received can be observed; the outcome under the alternative exposure is missing, so the individual causal effect can never be directly measured.

How does randomization help with counterfactuals?

Random assignment makes the exposure groups exchangeable on average, so the observed outcome in one group estimates the missing counterfactual outcome of the other, allowing the average causal effect to be estimated.

Counterfactual Reasoning

Counterfactual reasoning is the foundational logic of modern causal inference: it defines the effect of an exposure as a comparison between what actually happened and what would have happened under a different, contrary-to-fact exposure for the same individuals or population. Because only one of these outcomes is ever observed, causal inference becomes a problem of recovering the missing counterfactual.

Definition

Counterfactual reasoning defines a causal effect as a contrast between potential outcomes, the outcomes a unit would experience under different exposures, of which at most one is observed for each unit.

Scope

This topic covers the potential-outcomes (Rubin causal model) framework, the definition of individual and average causal effects as counterfactual contrasts, and the fundamental problem of causal inference that only one potential outcome per unit is observable. It is a methodological reference, not clinical guidance.

Core questions

What would have happened to the same people under a different exposure?
How is a causal effect defined when only one outcome can be observed?
How do individual and average causal effects relate?

Key concepts

Potential outcomes
Individual versus average causal effect
Fundamental problem of causal inference
Counterfactual contrast
Exchangeability
Consistency

Mechanisms

In the potential-outcomes model formalised by Rubin (rubin-1974), each unit has a potential outcome under each possible exposure; the individual causal effect is the contrast between them, and the average causal effect is the population mean of those contrasts. The fundamental problem is that only the outcome under the exposure actually received is observed, so the counterfactual must be estimated using a comparison group. This is valid only when the groups are exchangeable, that is, would have had the same outcome distribution had they received the same exposure (greenland-robins-1986), and when the observed outcome matches the corresponding potential outcome (consistency). Randomisation makes exchangeability hold by design; in observational data it must be assumed and defended (hernan-robins-2006). Causal diagrams provide a complementary structural representation of these same counterfactual assumptions (greenland-pearl-robins-1999).

Clinical relevance

The counterfactual definition clarifies what a treatment or exposure effect actually means and why a valid comparison group is essential, which underpins the appraisal of all causal evidence in the health sciences. It describes the logic of effect estimation and is not a basis for individual diagnostic or treatment decisions.

Epidemiology

The potential-outcomes framework is the conceptual backbone of contemporary epidemiologic methods, from randomized trials to observational analyses using adjustment, weighting, or g-methods. It provides the language in which confounding, selection bias, and effect modification are now defined (hernan-robins-2006).

History

The potential-outcomes idea traces to Neyman's early-twentieth-century work on randomized experiments and was generalised to observational studies by Rubin in 1974 (rubin-1974). Greenland and Robins connected it to epidemiologic confounding through exchangeability (greenland-robins-1986), and the framework, later unified with causal diagrams (greenland-pearl-robins-1999), became central to how epidemiologists define and estimate causal effects (hernan-robins-2006).

Debates

How well-defined must a counterfactual intervention be?: Some argue that counterfactual contrasts are only meaningful for exposures corresponding to reasonably well-defined hypothetical interventions, raising questions about effects of attributes such as race or body weight; others take a broader view of admissible contrasts.

Key figures

Donald Rubin
Jerzy Neyman
James Robins
Sander Greenland
Miguel Hernán

Seminal works

rubin-1974
greenland-robins-1986
greenland-pearl-robins-1999

Frequently asked questions

Why is it called the 'fundamental problem' of causal inference?: Because for any unit only the outcome under the exposure actually received can be observed; the outcome under the alternative exposure is missing, so the individual causal effect can never be directly measured.
How does randomization help with counterfactuals?: Random assignment makes the exposure groups exchangeable on average, so the observed outcome in one group estimates the missing counterfactual outcome of the other, allowing the average causal effect to be estimated.