Causal Inference

The statistical and epistemological methods used to determine whether an observed association between an exposure and outcome reflects a genuine causal relationship.

Source: Hernán & Robins 2020

Finding that two things vary together is straightforward, but showing that one causes the other is not, because an association can arise from confounding, reverse causation, or chance rather than a true causal link. Causal inference provides the reasoning and methods to distinguish genuine cause from mere correlation, ideally through randomisation, or through careful assumptions and design when only observational data exist. The core difficulty is that the counterfactual, what would have happened otherwise, is never directly observed. Correlation alone cannot settle cause. Hernan and Robins (2020) set out this framework.

Source: Hernan & Robins 2020

Causal inference is challenging because association does not imply causation: an observed relationship between exposure and outcome can arise from confounding, where a third factor influences both, from selection or information bias, or from reverse causation, so establishing that the exposure causes the outcome requires ruling these out. In observational data especially, unmeasured confounding threatens causal conclusions. Because what would have happened under a different exposure for the same individuals cannot be observed, causal inference relies on assumptions and study designs to approximate that counterfactual comparison, which is inherently demanding.

Source: Hernán & Robins 2020

Methods supporting causal inference include randomised experiments, which balance confounders and provide the strongest basis for causal claims; and, for observational data, approaches such as adjustment for confounders through regression, stratification, or matching; propensity score methods; instrumental variables; and quasi-experimental designs such as difference-in-differences and regression discontinuity. Causal diagrams help identify what to adjust for. These methods aim to approximate the comparison that randomisation provides, addressing confounding and bias. The choice depends on the data and assumptions, and each rests on conditions that must hold for the causal interpretation to be valid.

Source: Rubin 1974

Causal inference from observational data rests on assumptions such as no unmeasured confounding, meaning all common causes of exposure and outcome are measured and accounted for; positivity, that individuals with each combination of characteristics could receive each exposure; and consistency, that the exposure is well defined. These assumptions allow the observed data to identify the causal effect. Because they cannot generally be verified and may fail, particularly no unmeasured confounding, causal conclusions from observational data are held with caution, and sensitivity analyses assess how violations could affect them. So the validity of the inference depends on these assumptions.

Source: Hernán & Robins 2020

Causal inference is important because decisions about interventions, treatments, and policies depend on knowing whether they actually cause changes in outcomes, not merely whether they are associated with them, so distinguishing causation from association is central to acting effectively. In health and economics, mistaking a non-causal association for a causal one could lead to ineffective or harmful decisions. Causal inference provides the frameworks and methods to draw causal conclusions responsibly, with their assumptions made explicit. Because sound decisions require understanding causes, causal inference is fundamental to research that aims to inform action.

Source: Hernán & Robins 2020