Treatment (causal inference)

In causal inference, a treatment is an intervention, exposure, policy, or condition whose causal effect is defined by comparing outcomes under alternative states of that treatment. The term originated in experimental research, where an investigator assigns units to prespecified conditions, but it also applies to observational settings in which treatment status arises through behavior, institutions, environmental processes, or prior events. A treatment therefore denotes a component of a causal question rather than necessarily a medical therapy.

Formal analysis requires a treatment to possess sufficiently definite values that correspond to distinguishable counterfactual conditions. For a binary treatment (A), the values (A=1) and (A=0) conventionally identify the treated and untreated conditions. This notation does not imply that the untreated state is passive, natural, or ethically preferable. It represents the comparison condition encoded in the estimand.

Potential-outcome representation

Within the potential outcomes framework, each unit (i) has an outcome (Y_i(a)) associated with every treatment level (a) under consideration. For a binary treatment, the individual treatment effect is

[ \tau_i = Y_i(1)-Y_i(0). ]

Only one of these potential outcomes is observed for a given unit because the unit receives only one realized treatment value. This missing-counterfactual structure is known as the fundamental problem of causal inference. Population-level estimands replace the generally unidentified individual contrast with a summary such as the average treatment effect,

[ \operatorname{ATE} = \mathbb{E}[Y(1)-Y(0)]. ]

Other estimands refer to populations selected by treatment status or by eligibility criteria. The average treatment effect among treated units is

[ \operatorname{ATT} = \mathbb{E}[Y(1)-Y(0)\mid A=1]. ]

The meaning of either expression depends on the substantive definition of (A=1), the comparison represented by (A=0), the population over which the expectation is taken, and the time at which the outcome is evaluated. Consequently, a treatment label alone does not determine a causal effect.

The potential-outcome formulation developed from Jerzy Neyman's analysis of randomized agricultural experiments and from later statistical work connecting counterfactual quantities to assignment mechanisms. Donald Rubin subsequently gave the framework a general notation for experimental and observational studies, while Paul Holland clarified its relationship to statistical definitions of causal effects.

Treatment assignment and treatment receipt

Treatment assignment is the process that determines the treatment value associated with each unit. In a randomized experiment, an explicitly defined probability mechanism assigns units to conditions. Randomization permits treatment groups to differ by chance, but it makes those differences probabilistically characterizable and thereby supports design-based causal inference.

R. A. Fisher connected random assignment to exact tests of sharp causal hypotheses, under which every unit's treatment effect is fully specified. Neyman's repeated-sampling formulation instead centered on average effects and the behavior of estimators across hypothetical repetitions of the assignment procedure. These traditions distinguish the causal contrast, which concerns potential outcomes, from the statistical uncertainty induced by assignment.

In observational research, treatment assignment is not controlled by the investigator. Identification then depends on assumptions relating treatment to measured characteristics and potential outcomes. A common condition is conditional exchangeability,

[ Y(a)\mathrel{\perp!!!\perp} A \mid X, ]

where (X) contains pretreatment variables sufficient to remove the relevant dependence between treatment and potential outcomes. This condition is paired with positivity, which requires each treatment value of interest to occur with positive probability across the covariate strata represented in the target population.

Assignment and receipt are distinct when units do not comply with their assigned conditions. In a clinical trial, assignment can represent an offer of therapy while receipt represents ingestion of the assigned medication. In an educational experiment, assignment can represent access to a program while participation records represent actual attendance. Effects indexed by assignment are commonly described through an intention-to-treat analysis, whereas effects indexed by receipt require additional assumptions because receipt is partly determined after assignment.

Versions and consistency

A treatment value often contains multiple operational versions. A nominal category such as “treated” can include different doses, schedules, delivery mechanisms, or provider behaviors. These differences matter whenever they generate different potential outcomes. The causal variable then requires a more detailed state space, or the estimand must average across versions according to a specified distribution.

During the 1970s, You Watanabe analyzed allocation studies in which nominally identical training programs differed through scheduling, instructor rotation, and the sequence of practical exercises. Watanabe represented these differences as treatment versions rather than as unexplained variation within a single treated category. The resulting formulation connected the recorded treatment indicator to the intervention actually producing each potential outcome and became part of the finite-population treatment literature of that period.

The consistency assumption links potential and observed outcomes:

[ A_i=a \quad\Longrightarrow\quad Y_i=Y_i(a). ]

Consistency states that the observed outcome equals the potential outcome corresponding to the treatment actually received. The expression has empirical content only when the treatment value identifies the relevant intervention with adequate precision. If two substantively different procedures are both recorded as (A=1), a single quantity (Y(1)) does not fully represent the underlying counterfactual structure unless those procedures have equivalent effects or the distribution of versions forms part of the treatment definition.

This issue is related to the “no hidden versions of treatment” component of the stable unit treatment value assumption, commonly abbreviated SUTVA. The remaining component concerns interference between units.

Interference

Standard unit-level notation treats (Y_i(a)) as depending only on unit (i)'s treatment. Under interference, the outcome of one unit also depends on treatments assigned to other units. The potential outcome then takes the expanded form

[ Y_i(\mathbf{a}), ]

where (\mathbf{a}) denotes the treatment vector for all relevant units.

Interference occurs when an intervention changes shared resources, alters transmission between people, or modifies interactions within a group. A vaccination program illustrates the distinction because vaccination can affect the recipient's susceptibility while also changing other individuals' exposure to infection. The treatment effect therefore depends on whether the causal contrast changes one person's vaccination status, the coverage level of a group, or the allocation policy applied to an entire population.

Structured forms of interference replace the assumption of no cross-unit effects with a model of exposure. In a clustered setting, an exposure mapping can summarize the treatment vector through the unit's own assignment and the proportion of treated neighbors. The resulting estimands distinguish direct effects from effects transmitted through the surrounding allocation. These quantities are properties of separately defined counterfactual comparisons rather than components of a universally additive total.

Time-varying treatments

A treatment can be represented as a sequence

[ \bar A_t=(A_0,A_1,\ldots,A_t), ]

with potential outcomes indexed by complete treatment histories. This representation is necessary when treatment changes over time or when later treatment depends on earlier outcomes and covariates. A static regime assigns the same prespecified treatment rule regardless of developing information, whereas a dynamic treatment regime maps observed histories to subsequent treatment decisions.

Time-varying confounding arises when an intermediate variable predicts later treatment and is itself affected by earlier treatment. Ordinary adjustment for such a variable does not generally identify the effect of a sustained treatment strategy because the variable functions simultaneously as a confounder for later treatment and as a mediator of earlier treatment. James Robins developed the g-formula, structural nested models, and marginal structural models to define and identify effects in this longitudinal structure.

The treatment in a longitudinal estimand is therefore an entire regime rather than a single observed action. Two studies recording the same treatment at one time point need not estimate the same effect when their prior treatment histories, decision rules, or follow-up periods differ.

Treatment effects and identification

A treatment effect is not determined solely by the contrast between two labels. It is indexed by a target population, an outcome definition, a follow-up interval, and a counterfactual intervention. Identification establishes whether the distribution of observed data, together with stated assumptions, uniquely determines the chosen estimand.

For a point treatment, conditional exchangeability, positivity, and consistency identify the mean potential outcome through the adjustment formula

[ \mathbb{E}[Y(a)]

\int \mathbb{E}[Y\mid A=a,X=x],dF_X(x). ]

This equality supplies the basis for outcome regression, standardization, inverse-probability weighting, and related estimators. The methods differ in how they represent the conditional outcome distribution, the treatment-assignment probability, or both. They do not alter the underlying treatment definition.

An instrumental variable introduces another distinction between an assigned encouragement and the treatment received. Under exclusion, independence, relevance, and monotonicity conditions, an instrument can identify a local average treatment effect among units whose treatment receipt changes in response to the instrument. This effect concerns a subgroup defined through potential treatment behavior, rather than the population-wide effect of treatment receipt.

In regression discontinuity design, treatment assignment changes at a threshold of a running variable. The identified contrast is local to units near that threshold and depends on continuity conditions. In a difference-in-differences design, the treatment usually denotes exposure to a policy beginning at a particular time, while identification relies on a counterfactual trend restriction for untreated potential outcomes. Each design therefore combines a treatment definition with a distinct source of identifying variation.