Estimands as Clinical Questions

Start With the Decision, Not the Model

A clinical result is useful only if it answers a question someone actually needs answered. “Does treatment work?” is not yet precise enough: work for which patients, on what outcome, and in what sense? A difference in average symptom score may answer a different question from a difference in the chance of avoiding hospitalization. Even for the same outcome, a comparison at 12 weeks can answer a different question from one about the course of symptoms over 12 weeks.

An estimand is the target quantity a study seeks to learn about. Thinking of it as a carefully specified clinical question prevents a statistical analysis from silently deciding what “benefit” means. The question comes first; a model or calculation is then chosen to estimate the quantity it defines.

Definition: A clinical estimand is a precisely defined target treatment contrast for a stated population and clinical outcome. It specifies whose outcomes are relevant, the treatment conditions being compared, and the outcome summary that gives the comparison its clinical meaning. It is a target quantity, not a calculation method or an observed numerical result.

The emphasis here is on translating the question into a target contrast: identify the people whose outcomes matter and say exactly what clinical quantity is to be compared. A complete question also needs to make the outcome meaningful in context, including its measurement and timing. More formal organization of the attributes of a trial estimand builds on these ideas.

Expressing a Target Contrast

Potential outcomes provide a concise way to write a treatment comparison. Let \(Y_i(1)\) be the outcome person \(i\) would have under treatment 1 and \(Y_i(0)\) the outcome that same person would have under treatment 0. These are defined for a particular outcome and time point, and for the treatment conditions as specified in the clinical question. In general, only one of these two outcomes is observed for each person.

Let \(S\) denote the target population. If the clinical question concerns average outcome, one possible target is the difference in population means:

\Delta_{\text{mean}}=E[Y(1)\mid S]-E[Y(0)\mid S].

This contrast answers: among people in \(S\), how different would the average outcome be under treatment 1 versus treatment 0? The sign and units depend on how the outcome is defined. If higher symptom scores mean worse symptoms, a negative mean difference may favor treatment 1. If the outcome is improvement, a positive difference may favor treatment 1.

A clinical question may instead concern a binary outcome, such as hospitalization by a specified time. The risk difference is

\Delta_{\text{risk}}=P(Y(1)=1\mid S)-P(Y(0)=1\mid S),

where \(Y(a)=1\) means that the event occurs under treatment condition \(a\). This contrast is measured in probability units, often reported as a percentage-point difference. Neither the mean difference nor the risk difference is universally the right choice: each answers a different clinical question.

Potential outcomes clarify what is being compared, but they do not by themselves guarantee that the target can be learned from observed data. For a causal interpretation, the treatment conditions must be sufficiently well defined, observed treatment and outcome data must correspond to the conditions in the question, and the design and analysis must support comparison of the relevant potential outcomes. For example, random assignment can support such a comparison for the population enrolled, but generalizing to a different population requires further justification. Defining the target and identifying it from data are separate tasks.

Worked Examples: Making the Question Concrete

Worked Example: Average Symptom Improvement at 12 Weeks

Suppose a hypothetical trial asks whether a new treatment improves chronic pain more than standard care among adults eligible for the trial. Pain is recorded on a 0–10 scale, and improvement is defined as baseline score minus the score at week 12. The target question is the difference in mean improvement under assignment to the new treatment versus assignment to standard care in the eligible population.

Assume, for illustration, the target mean improvement is 2.4 points under the new treatment and 1.7 points under standard care. The target contrast is

\Delta_{\text{mean}}=2.4-1.7=0.7\text{ points}.

The positive value means that mean improvement is 0.7 points greater under the new treatment condition, on the stated scale and at week 12. This is a hypothetical target contrast, not a claim that a real trial observed this result. If the question instead concerned the mean pain score at week 12, rather than improvement from baseline, the outcome quantity would be different; its contrast could not be substituted without changing the clinical question.

Worked Example: Risk of Hospitalization by 90 Days

In a second hypothetical question, clinicians want to know whether a preventive treatment changes the chance of hospitalization within 90 days among adults with a specified condition. Let the binary outcome equal 1 for hospitalization by day 90. Suppose the target risks are 0.18 under the preventive treatment and 0.26 under comparison care.

The risk difference, treatment minus comparison, is

\Delta_{\text{risk}}=0.18-0.26=-0.08.

Equivalently, the risk difference is \(-8\) percentage points. In this hypothetical example, the target risk is 8 percentage points lower under the preventive treatment condition. The unit matters: this is an absolute difference in risks, not an 8% relative reduction. A relative-risk question would use a different contrast, \(0.18/0.26\approx0.692\), and would express the comparison on a relative rather than an absolute scale. Which summary better serves the decision depends on the clinical use of the answer.

Worked Example: Choosing the Outcome Summary

Suppose a hypothetical study records a symptom score at weeks 4, 8, and 12 for a target population of patients with the same condition. The clinical team asks whether treatment reduces symptom burden across the 12-week period. One possible summary is the mean of the three recorded scores for each treatment condition, averaged over the target population. Suppose those target summaries are 5.0 points under treatment and 6.2 points under comparison care, with higher scores indicating worse symptoms.

The contrast is

\Delta=5.0-6.2=-1.2\text{ points}.

The target average symptom score is 1.2 points lower under treatment across the specified measurement occasions. This answers a different question from the treatment contrast at week 12 alone. If the week-12 means were, for example, 4.3 and 5.0 points, the week-12 contrast would be \(4.3-5.0=-0.7\) points. Both calculations can be correct, but only one matches a given clinical question. The values here are hypothetical and illustrate how the choice of summary changes the target.

Why the Contrast Follows From the Question

The difference between mean potential outcomes has a useful interpretation. For a common target population \(S\), linearity of expectation gives

E[Y(1)\mid S]-E[Y(0)\mid S] =E[Y(1)-Y(0)\mid S].

To see why, write the conditional expectation of a difference as the difference of conditional expectations:

E[Y(1)-Y(0)\mid S] =E[Y(1)\mid S]-E[Y(0)\mid S].

Thus, the mean contrast can also be described as the average of the individual treatment differences in the target population. This is a mathematical identity; it does not mean that both potential outcomes are observed for each person, nor does it say that the average contrast is the same for every subgroup. It also does not establish that the contrast is identifiable from a particular data set.

The common population in both terms is essential. Comparing the mean under treatment among one group with the mean under comparison care among a different group may combine treatment differences with population differences. A clinical question should therefore make clear whose outcomes are being contrasted before a numerical comparison is interpreted.

Common Ambiguities and Failure Modes

A useful way to check an estimand is to restate it in ordinary clinical language: “For [which patients], compare [which treatment conditions] on [which outcome summary] at or over [which time period].” If the sentence admits multiple materially different interpretations, the target is not yet clear enough.

Changing the population unnoticed. “Patients in the trial” and “all patients who might receive the treatment” need not refer to the same group. A result for enrolled participants does not automatically answer a question about a broader clinical population.
Using a vague outcome. “Better health” does not specify whether the target is a mean score, event risk, time to an event, or another quantity. Different summaries can lead to different clinical interpretations.
Leaving time unstated. A treatment contrast at week 12 need not match one at week 4 or an average over repeated visits. State the time point or period the decision concerns.
Treating the observed-data calculation as the question. The mean among people with recorded outcomes is a statistic from a selected set of records. It is not automatically the target mean for the intended population.
Ignoring events after treatment begins. Treatment discontinuation, rescue medication, or other events can affect what outcome a clinical question intends to compare. If the question does not say how such circumstances relate to the target, the intended contrast may remain ambiguous. Their detailed classification and handling require separate specification.

A frequent mistake is to reason, “The analysis compares the groups among people with observed week-12 scores, so that comparison is the estimand.” The correction is to ask first which population and outcome quantity the clinical decision concerns. The observed-data analysis is a route to an answer, and may or may not successfully estimate that target. Missing outcomes can make estimation difficult; they do not by themselves redefine the clinical question.

Clinical Target, Statistical Analysis, and Alternatives

Three items are easy to confuse. The estimand is the target clinical quantity. An estimator is a statistical procedure used to estimate that quantity from data. An estimate is the numerical result produced by applying the procedure to a particular data set. For example, a mean difference may be the target; a specified model-based procedure may estimate it; and the calculated difference in points is the estimate. Choosing a sophisticated model cannot repair an unclear target.

There is no single contrast that answers every treatment question. A mean difference focuses on average outcome in the outcome’s original units. A risk difference focuses on absolute change in event probability; a risk ratio focuses on relative change. A contrast at a fixed visit addresses that visit, whereas a summary across visits addresses the chosen period. These are alternatives because they answer different questions, not merely because they use different statistical machinery.

Similarly, a contrast based on treatment assignment and one based on treatment actually received need not target the same clinical question. The former can address the consequences of assigning a treatment policy, while the latter may address a different question and requires its own clear definition and assumptions. The names and formal strategies for handling events after treatment starts are specified separately; at this stage, the practical lesson is to notice when a question leaves those circumstances unresolved rather than silently choosing a convenient analysis.

The estimand framework matters because treatment decisions depend on the meaning of the comparison, not just its numerical value. A negative 1-point score difference, an 8-percentage-point risk reduction, and a relative risk of 0.69 are not interchangeable descriptions. Clear targets make it possible to judge whether an analysis answers the decision, identify why analyses disagree, and communicate what a result does and does not establish.

Check Your Understanding

For each question, identify or calculate the target contrast, then compare your reasoning with the solutions below.

A clinic wants to compare mean fatigue-score improvement at 8 weeks among eligible patients receiving two treatment conditions. What must be specified to make this a clear clinical question?
In a hypothetical population, the target event risks by 6 months are 0.12 under treatment and 0.20 under comparison care. Calculate and interpret the risk difference, treatment minus comparison.
Why is a mean difference at week 12 not automatically an answer to a question about average symptoms over 12 weeks?
Explain why the mean difference among people with observed outcomes is not, merely by being calculated, the target estimand for all eligible patients.

Solutions. (1) State which eligible patients form the target population, define fatigue and how improvement is measured, specify the 8-week time point, identify the two treatment conditions, and choose the mean improvement contrast. (2) \(0.12-0.20=-0.08\), or \(-8\) percentage points. The target risk is 8 percentage points lower under treatment in this hypothetical population. (3) A week-12 value summarizes one time point; an average over 12 weeks combines outcomes across a period according to a specified summary. They are different outcome quantities. (4) The observed-outcome subset may differ from the full eligible population. The observed-data mean is a calculation on available records; additional assumptions and an appropriate analysis are needed to relate it to the target for all eligible patients.