Causal Inference
Causal inference is the branch of epidemiology and biostatistics concerned with deciding when an observed association between an exposure and an outcome reflects a genuine cause-and-effect relationship rather than chance, bias, or confounding. It supplies the conceptual frameworks, graphical tools, and analytic methods that let researchers state causal questions precisely and judge whether the available data can answer them.
Definition
Causal inference is the set of formal frameworks and methods used to define causal effects, state the assumptions under which they can be estimated from data, and assess how robust the resulting conclusions are to violations of those assumptions.
Scope
This area gathers the core machinery of modern causal reasoning in health research: causal criteria and theories of causation, directed acyclic graphs for encoding assumptions, the identification conditions that connect a causal estimand to estimable quantities, sensitivity analysis for unmeasured bias, and the counterfactual (potential-outcomes) framework that underpins them all. It is a methodological reference, not clinical guidance.
Sub-topics
Core questions
- When does a statistical association support a causal conclusion?
- What assumptions are needed to identify a causal effect from observational data?
- How can those assumptions be made explicit and checked?
- How sensitive is a causal conclusion to unmeasured confounding or other bias?
Key concepts
- Counterfactuals and potential outcomes
- Confounding and exchangeability
- Directed acyclic graphs
- Identification and estimands
- Sensitivity analysis
- Hill's viewpoints on causation
Mechanisms
Modern causal inference rests on the potential-outcomes (counterfactual) model formalised by Rubin (rubin-1974), in which a causal effect is a contrast between outcomes that would occur under different, mutually exclusive exposures for the same units. Directed acyclic graphs (greenland-pearl-robins-1999) translate substantive assumptions about how variables relate into a graph whose structure determines which adjustments block confounding and which would introduce bias. Identification asks whether, under stated assumptions such as exchangeability, positivity, and consistency, the counterfactual contrast equals a function of the observed data (hernan-robins-2006). Where assumptions cannot be guaranteed, sensitivity analysis quantifies how strong an unmeasured bias would have to be to overturn the finding.
Clinical relevance
Causal inference frameworks shape how observational evidence on treatments, exposures, and risk factors is generated and appraised; understanding them helps readers judge whether a reported effect is credible. This area describes how evidence is reasoned about and is not a source of individual diagnostic or treatment recommendations.
Epidemiology
Causal-inference methods are now standard across observational epidemiology, pharmacoepidemiology, and comparative-effectiveness research, where randomisation is often impossible and investigators must instead make and defend explicit assumptions. The pluralistic tradition stresses that no single method or criterion settles causation on its own (vandenbroucke-2016).
History
Twentieth-century epidemiology moved from informal association-versus-causation debates, crystallised in Hill's 1965 viewpoints (hill-1965), toward an explicit mathematical theory of causation. Rubin's 1974 potential-outcomes formulation (rubin-1974) and the subsequent development of causal diagrams by Greenland, Pearl, and Robins (greenland-pearl-robins-1999) unified counterfactual reasoning with graphical models, and by the 2000s these tools had become central to how epidemiologists frame and answer causal questions (hernan-robins-2006).
Debates
- Is there a single correct framework for causal inference?
- Some argue the counterfactual model with graphical methods provides a unified foundation, while others defend a pluralistic view in which different criteria and methods complement one another and no single rule resolves causation.
Key figures
- Austin Bradford Hill
- Jerome Cornfield
- Donald Rubin
- James Robins
- Sander Greenland
- Judea Pearl
- Miguel Hernán
Related topics
Seminal works
- hill-1965
- rubin-1974
- greenland-pearl-robins-1999
- hernan-robins-2006
Frequently asked questions
- How is causal inference different from ordinary statistical association?
- Association describes how variables move together in data; causal inference adds explicit assumptions about how the data were generated so that an association can be interpreted as the effect of changing one variable on another.
- Can causal effects be estimated without a randomized trial?
- Yes, but only under stated and often untestable assumptions such as no unmeasured confounding; causal-inference methods make those assumptions explicit and let researchers test how robust conclusions are to them.