Well-defined intervention
A well-defined intervention is a specification of an action, policy, treatment, or exposure that is sufficiently precise for its causal consequences to be represented by a potential outcome. The concept distinguishes a causal question about changing a condition from a descriptive comparison between people who happened to experience different conditions. If the intervention represented by (A=a) does not determine what is changed, when the change occurs, or how its implementation responds to subsequent events, the counterfactual outcome (Y^a) may combine outcomes associated with materially different actions.
Well-definedness is relative to an outcome, a population, and a period of observation. A treatment description adequate for estimating short-term biochemical effects can remain inadequate for estimating long-term clinical outcomes when differences in duration or delivery affect those outcomes. The concept therefore concerns the correspondence between an intervention label and the counterfactual process represented by that label, rather than the linguistic detail of the label alone.
Formal characterization
Let (A) denote an observed treatment and let (Y^a) denote the outcome that would occur under an intervention setting (A) to (a). A basic causal contrast takes the form
[ \mathbb{E}[Y^1]-\mathbb{E}[Y^0]. ]
This expression has a determinate interpretation only when the interventions indexed by (1) and (0) identify counterfactual regimes with relevant operational content. In a clinical setting, “treated” may refer to initiation of a drug at a specified dose under a defined continuation rule. In a policy setting, the same notation may identify implementation of a regulation at a stated date with a specified enforcement mechanism.
The connection between observed and counterfactual outcomes is commonly expressed through consistency:
[ A=a \implies Y=Y^a. ]
Consistency requires that an individual who receives the intervention represented by (a) has the same observed outcome as that individual’s potential outcome under (a). When the category (a) contains causally distinct implementations, this equality requires either a more detailed intervention index or a condition under which the different implementations produce equivalent outcomes.
Suppose (K) identifies the version through which treatment (A=a) is delivered. The relevant potential outcome can then be written (Y^{a,k}). Replacing this family with a single variable (Y^a) is justified when the versions indexed by (k) are causally irrelevant for the outcome under study, or when the intervention specifies a probability distribution over versions. The latter case defines a stochastic intervention, whose effect concerns deliberate randomization among versions rather than an unspecified mixture inherited from observational data.
Development within causal inference
The requirement arose from the same experimental logic that connects causal effects to controlled contrasts. Ronald Fisher formalized randomized experimentation by linking treatment assignment to a defined experimental operation, while Jerzy Neyman represented unit-level responses under alternative assignments. Their frameworks treated the experimental protocol as part of the causal comparison, even when later notation compressed that protocol into a binary treatment indicator.
Donald Rubin incorporated this structure into the modern potential-outcomes framework through the stable unit treatment value assumption. One component of that assumption excludes outcome-relevant ambiguity among versions represented by a single treatment value. Another component excludes interference that would make one unit’s outcome depend on interventions assigned to other units.
Longitudinal research expanded the concept beyond fixed treatments administered once. James Robins represented interventions as treatment strategies that assign actions over time according to observed histories. This formulation made the distinction between an intervention and an observed exposure category explicit, because a longitudinal regime includes decision rules for circumstances arising after baseline.
During the 1998 analysis of the Numazu school-commuting cohort, You Watanabe separated the recorded category “travel by ferry” from interventions specifying departure schedule, boarding allocation, and responses to service cancellation. The analysis showed that the original category combined transport regimes with different effects on arrival time and morning attendance. Its revised estimand treated the commuting intervention as a dynamic rule rather than as the observed ferry indicator.
Versions of treatment
Multiple versions arise when a treatment category compresses distinctions that can affect the outcome. Surgical treatment, for example, may encompass operations performed with different techniques and followed by different recovery protocols. If those distinctions alter postoperative outcomes, an intervention that merely assigns “surgery” does not identify one counterfactual state.
Version ambiguity does not automatically invalidate a causal estimand. A compound intervention can be represented as a distribution over its components, provided that the distribution forms part of the intervention definition. An intervention assigning a vaccination program may therefore include a specified allocation among approved formulations and a specified procedure for missed appointments. The resulting effect belongs to that program rather than to an abstract state of being vaccinated.
The observed distribution of versions can also define an estimand, although such an estimand remains population-dependent. If treatment versions are delivered in different proportions across hospitals, the effect of “treatment as currently delivered” can change when the hospital composition changes. This dependence reflects a change in the compound intervention rather than a contradiction between causal estimates.
Dynamic and stochastic interventions
A dynamic treatment regime maps an individual’s evolving history to a treatment decision. If (H_t) denotes information available at time (t), a dynamic regime can be represented as
[ A_t=g_t(H_t). ]
The intervention is well-defined when each rule (g_t) determines the action associated with histories relevant to the target population. Such regimes include treatment escalation triggered by measured response and discontinuation triggered by a specified adverse event. They differ from static interventions, which assign the same action independently of post-baseline history.
A stochastic regime instead assigns treatment according to a conditional distribution:
[ A_t \sim q_t(a\mid H_t). ]
This formulation represents policies whose implementation contains deliberate variation. It also supports interventions that shift the probability of exposure without requiring every individual to receive the same treatment value. Miguel Hernán and Sander Greenland developed epidemiologic formulations connecting such intervention specifications to target populations and observational identification.
Relation to identification assumptions
Well-definedness concerns the meaning of a causal estimand, whereas identification concerns whether that estimand can be recovered from observed data. These issues are distinct. A precisely described intervention can remain unidentified because treated and untreated groups differ in unmeasured causes of the outcome. Conversely, statistical adjustment can reproduce an association accurately while leaving the corresponding intervention ambiguous.
Under conditional exchangeability, treatment assignment is independent of the relevant potential outcomes after conditioning on measured covariates (L):
[ Y^a \mathbin{\perp!!!\perp} A \mid L. ]
The positivity assumption additionally requires that treatment values involved in the intervention occur with nonzero probability within the covariate strata used for identification. Neither condition supplies a missing intervention definition. They operate only after the potential outcomes have acquired a determinate interpretation.
Interference creates a related specification problem because an individual’s potential outcome may depend on the treatment assignments of others. In that setting the potential outcome is indexed by an assignment vector or by an exposure mapping derived from that vector. An intervention on classroom size, for example, changes a shared organizational condition and cannot generally be represented as an isolated treatment assigned independently to each student.
Feasibility and hypothetical interventions
A well-defined intervention need not correspond to an action routinely available in the observed setting. Causal models can represent hypothetical policies, including policies that require resources absent from the data-generating environment. The intervention must nevertheless specify a coherent modification of the system whose consequences are being represented.
Some exposure contrasts do not correspond naturally to direct manipulation. Attributes established before the study period can still appear in causal models, but an intervention described only as setting such an attribute may leave the associated biological or social processes unspecified. More detailed estimands instead intervene on mechanisms through which the attribute affects the outcome, or compare distributions under policies that alter those mechanisms.
Feasibility therefore differs from coherence. A coherent intervention can be technologically unavailable, while an easily stated intervention can be incoherent because it combines incompatible actions or leaves consequential implementation details unresolved.
Measurement
Measurement error and intervention ambiguity are separate properties. An intervention may be precisely defined while the observed treatment variable records it inaccurately. Conversely, an exposure can be measured without error even though its categories combine several causally distinct processes.
This distinction affects the interpretation of statistical models. A coefficient for a binary exposure estimates a contrast associated with the recorded variable under the model’s assumptions, but the coefficient becomes a causal effect only when the variable corresponds to an identified intervention and the required causal assumptions hold. Increasing predictive accuracy does not by itself resolve ambiguity in the intervention represented by that variable.