Causal Identification

Causal identification is the step in causal inference that asks whether a causal quantity of interest can, even in principle, be recovered from the available data under stated assumptions. It separates the conceptual question of what is being estimated from the statistical question of how precisely it can be estimated, and is logically prior to any analysis.

Definition

Causal identification is the demonstration that, under explicit assumptions, a defined causal effect equals a quantity that can be computed from the distribution of the observed data.

Scope

This topic covers the identification conditions that link a causal estimand to a function of observed data, exchangeability (no unmeasured confounding), positivity, and consistency, together with strategies such as covariate adjustment, instrumental variables, and target-trial emulation. It is a methodological reference, not clinical guidance.

Core questions

• Can the causal effect of interest be recovered from the data at all?
• What assumptions are required for identification, and are they plausible?
• Which strategy, adjustment, instruments, or trial emulation, identifies the target effect?

Key concepts

• Estimand and identification
• Exchangeability (no unmeasured confounding)
• Positivity
• Consistency
• Instrumental variables
• Target trial emulation

Mechanisms

Identification requires linking a counterfactual contrast to observed quantities through assumptions. Greenland and Robins (greenland-robins-1986) formalised exchangeability, the requirement that exposed and unexposed groups would have had comparable outcomes had they shared the same exposure, as the core condition that, together with positivity (every relevant subgroup can experience each exposure) and consistency (the observed outcome under the actual exposure equals the corresponding counterfactual), allows confounding to be removed by adjustment. When unmeasured confounding makes adjustment-based identification implausible, an instrumental variable, a factor affecting the outcome only through exposure, can sometimes identify an effect under its own strong assumptions (hernan-robins-2006-iv). Framing an observational analysis as the explicit emulation of a hypothetical randomized target trial helps make identification assumptions transparent and avoids common design biases (hernan-robins-2016-trial). Graphical criteria provide a systematic way to check identifiability from assumed causal structure (pearl-1995).

Clinical relevance

Whether a causal effect is identifiable determines whether observational evidence on a treatment or exposure can be interpreted causally at all, which is central to appraising such evidence. This topic describes the logic of evidence generation and is not a basis for individual diagnostic or treatment decisions.

Epidemiology

Identification thinking is now embedded in observational epidemiology and comparative-effectiveness research, where investigators state their assumptions before estimating effects. The target-trial framework has become a widely used device for organising identification in studies using routinely collected health data (hernan-robins-2016-trial).

History

Greenland and Robins's 1986 paper gave epidemiology a rigorous account of identifiability through exchangeability (greenland-robins-1986), and graphical methods later supplied general criteria for checking it (pearl-1995). The instrumental-variable and target-trial literatures then extended identification strategies to settings where simple adjustment fails (hernan-robins-2006-iv, hernan-robins-2016-trial).

Debates

How credible are instrumental-variable assumptions?

Instrumental variables can identify effects under unmeasured confounding, but their key assumptions, that the instrument affects the outcome only through exposure and shares no common cause with it, are largely untestable and often debated in applications.

Key figures

• Sander Greenland
• James Robins
• Miguel Hernán
• Judea Pearl

Seminal works

• greenland-robins-1986
• hernan-robins-2006-iv
• hernan-robins-2016-trial

Frequently asked questions

What is the difference between identification and estimation?

Identification asks whether the causal quantity can be expressed in terms of observable data under stated assumptions; estimation asks how to compute it precisely from a finite sample once it is identified.

What is the 'no unmeasured confounding' assumption?

Often called exchangeability, it states that, conditional on measured covariates, exposed and unexposed groups would have had the same outcome distribution had they received the same exposure; it is required for adjustment-based identification and is generally untestable.