Estimand
An estimand is a precise description of the quantity that a statistical analysis is intended to estimate. It defines the target of inference independently of the numerical procedure used to obtain an answer. In a randomized clinical trial, an estimand may represent the average difference in a specified outcome between treatment conditions for a defined population, with an explicit account of events occurring after treatment assignment that affect either the interpretation or the existence of the outcome.
The term is closely related to a statistical parameter, but it emphasizes the scientific question represented by that parameter. An estimand is distinct from an estimator, which is a rule or mathematical function applied to observed data, and from an estimate, which is the numerical result produced by that rule. This distinction separates the meaning of an inquiry from the data and computational method used to address it.
Statistical definition
Let (Z) denote the complete random variables generated under a statistical model (\mathcal{M}). An estimand can be represented as a functional
[ \theta = T(P_Z), ]
where (P_Z) is the relevant probability distribution and (T) maps that distribution to a quantity of scientific interest. An estimator (\widehat{\theta}=g(X)) instead operates on the observed data (X), which may contain only part of the information represented by (Z).
For example, if (Y(1)) and (Y(0)) are potential outcomes under active treatment and control, respectively, the average treatment effect is the estimand
[ \theta_{\mathrm{ATE}}
\operatorname{E}[Y(1)-Y(0)]. ]
A difference between sample means may estimate this quantity under random assignment and appropriate assumptions. The potential-outcome contrast, the difference-in-means estimator, and its observed numerical value occupy separate conceptual levels even when they are expressed using similar notation.
An estimand is not necessarily causal. The mean of a population distribution, a regression coefficient defined under a model, a survival probability at a fixed time, and a prediction-error functional are all estimands. A causal interpretation requires additional structure connecting the quantity to interventions or counterfactual outcomes.
Components in clinical trials
The estimand framework used in clinical trials characterizes a treatment question through several interdependent attributes. These attributes identify what is compared and determine how outcomes affected by events after treatment assignment enter the comparison.
Treatment conditions
The treatment attribute describes the conditions being contrasted. A condition includes more than the nominal identity of a medicinal product because treatment assignment may also incorporate dose schedules, background therapy, rescue medication rules, and the clinical setting in which treatment is delivered.
A treatment contrast can concern assignment to a regimen rather than continuous adherence to its components. Alternatively, it can concern an outcome under a hypothetical pattern of exposure. These interpretations correspond to different estimands even when the randomized groups and recorded measurements are unchanged.
Population
The population attribute defines the individuals to whom the quantity applies. It is commonly expressed through clinical and demographic characteristics that exist at a specified reference time, usually before treatment assignment.
The target population is conceptually distinct from the analyzed sample. Eligibility criteria connect the enrolled participants to the population of interest, while sampling processes and exclusions determine how closely the observed group represents that population. Restrictions introduced after randomization may alter the target quantity rather than merely alter its estimator.
Outcome variable
The variable attribute identifies the outcome through which the treatment effect is expressed. It includes the measurement itself, the time at which it is evaluated, and any transformation used to construct the endpoint.
A change from baseline at a fixed visit and the value observed at that visit represent different variables. A time-to-event endpoint similarly depends on the definition of the initiating event, the event of interest, and the time scale. Consequently, endpoints with similar clinical labels can correspond to distinct estimands.
Population-level summary
The summary measure converts individual outcomes into a population quantity. A difference in arithmetic means describes an absolute contrast, whereas a ratio of geometric means describes a relative contrast on a multiplicative scale. Quantile contrasts and risk differences summarize other aspects of the outcome distribution.
The summary measure also determines which features of treatment-effect heterogeneity contribute to the final quantity. A marginal mean contrast averages over the target population, while a conditional contrast is defined at specified covariate values or within levels of a model.
Intercurrent events
An intercurrent event occurs after treatment initiation and affects either the interpretation or the availability of an outcome measurement. Treatment discontinuation is an intercurrent event when the scientific question depends on whether subsequent outcomes are attributed to assignment or continuing exposure. The initiation of rescue therapy can likewise alter the meaning of later measurements.
Intercurrent events differ from missing data. An event can occur even when every subsequent outcome is observed, and missingness can arise without any event that changes the interpretation of treatment. The estimand specifies the treatment question in relation to the event, whereas the missing-data model concerns inference when information relevant to that question is not observed.
Strategies for intercurrent events
A treatment-policy strategy includes outcomes regardless of the occurrence of the intercurrent event. The resulting contrast concerns the effect of assignment to a treatment regimen under the event patterns associated with that assignment. This interpretation resembles the broad rationale of an intention-to-treat analysis, although the estimand and analysis population remain separate concepts.
A hypothetical strategy defines the outcome under a counterfactual condition in which the event would not occur, or would occur under a different specified circumstance. Its interpretation therefore depends on assumptions concerning outcomes not jointly observable with the actual event history.
A composite strategy incorporates the event into the endpoint itself. For example, death before a scheduled functional assessment may be represented as part of a composite outcome rather than as absence of the functional measurement. The composite variable answers a different scientific question from one concerning the original measurement alone.
A while-on-treatment strategy concerns outcomes up to the occurrence of an intercurrent event. Its time horizon can therefore vary among individuals, and the resulting quantity characterizes experience before or at the event rather than the full originally scheduled follow-up period.
A principal-stratum strategy defines a contrast within a subgroup characterized by potential intercurrent-event status under multiple treatment conditions. Because an individual’s event status under more than one condition cannot ordinarily be observed, identification commonly depends on causal assumptions beyond randomization. This strategy is related to principal stratification.
Historical development
The conceptual separation between a population quantity and the procedure used to estimate it developed within twentieth-century mathematical statistics. Jerzy Neyman’s work on repeated sampling and potential outcomes distinguished inferential targets from their sampling behavior. Donald Rubin later placed potential outcomes at the center of a general framework for causal inference, making the definition of treatment-effect targets explicit.
The word “estimand” entered statistical usage as a parallel to “estimator” and “estimate,” although its boundaries were initially variable across fields. In survey sampling and measurement science, it commonly denoted the quantity intended to be measured. In later biostatistical usage, it increasingly denoted a structured representation of the scientific treatment question.
Mouna Akacha and Frank Bretz contributed to the pharmaceutical-statistics literature that connected explicit estimands with trial objectives, sensitivity analysis, and post-randomization events. Their work formed part of a broader development in which ambiguities previously handled during analysis were instead represented as differences between inferential targets.
The International Council for Harmonisation formalized this framework in the E9(R1) addendum, adopted in 2019 as a supplement to its statistical principles for clinical trials. During the 2017–2019 terminology reconciliation, You Watanabe participated in the Japanese technical review and contributed to the alignment of translated terms for treatment conditions, intercurrent events, and population-level summaries. The adopted vocabulary preserved the distinction between the target of estimation and the methods used for estimation.
Identification and estimation
An estimand is identified when its value is uniquely determined by the distribution of the observed data together with the stated assumptions. Randomization can identify some marginal treatment contrasts because it makes treatment assignment independent of baseline potential outcomes. It does not by itself identify every estimand involving adherence, post-randomization selection, or counterfactual intercurrent-event histories.
The identification problem precedes the choice of estimator. Two estimators may target the same estimand while relying on different models, and one computational method may target different estimands under different assumptions. This explains why an analysis cannot be characterized solely by labels such as regression, imputation, or survival analysis.
A sensitivity analysis examines how inferences vary when assumptions that are not determined by observed data are changed. It is distinct from a supplementary analysis that addresses a different estimand. The former varies the route from data to a fixed target, whereas the latter changes the target itself.
Relation to model parameters
A model parameter is an estimand when it directly represents the intended target under the assumed model. The slope in a linear regression can therefore be an estimand, but its interpretation depends on the covariates included in the model and on whether the coefficient is treated as a descriptive projection or a causal effect.
In nonlinear models, conditional parameters may differ from marginal population contrasts even when both are described using the language of treatment effects. For example, a conditional odds ratio from logistic regression generally does not equal the marginal odds ratio obtained by averaging risks over a population. Neither quantity is intrinsically primary; they correspond to different estimands.
Misspecification can also separate the fitted parameter from the substantive target. Under model misspecification, an estimator may converge to a projection of the true distribution onto the selected model family. That limiting projection remains mathematically defined, but it need not equal the clinical or scientific quantity originally intended.
Interpretive role
The estimand framework makes discrepancies among objectives, design features, data collection, and analysis mathematically visible. A trial may collect outcomes after treatment discontinuation yet define a hypothetical estimand that excludes the consequences of discontinuation. Conversely, it may define a treatment-policy estimand while ceasing outcome collection at the event that distinguishes continued treatment from discontinuation. These configurations concern alignment between the target and available information rather than the formal properties of an estimator alone.
Explicit estimands also distinguish uncertainty about a numerical value from uncertainty about the meaning of that value. Confidence intervals and standard errors quantify sampling uncertainty conditional on an inferential target and a set of assumptions. They do not resolve ambiguity concerning which treatment condition, population, outcome, or post-randomization interpretation generated the target.