Stating the Question and Target Estimand

From a Broad Question to a Target Estimand

A statement such as “Does treatment improve outcomes?” is not yet a report-ready research question. It leaves open who the result concerns, what treatment is compared with what alternative, which outcome is meant, when it is measured, and how “improve” will be quantified. Those choices determine the target quantity, not merely the wording of the report.

In an observational comparative study, the target estimand is the precise population-level quantity the analysis aims to learn about. It is distinct from the numerical estimate calculated from the available data. A clear estimand lets readers assess whether the analysis and its interpretation answer the question the authors intended to ask.

For a binary treatment, let \(A=1\) denote the treatment strategy and \(A=0\) the comparator. Let \(Y(1)\) and \(Y(0)\) denote the outcome that an individual would have under each strategy, measured at the specified time. These potential outcomes define a causal contrast. They are not both observed for the same individual, so a causal interpretation of an estimated contrast requires assumptions connecting the observed data to the target quantity. Naming the estimand does not itself establish that those assumptions hold.

For example, a report might ask about the average difference in 30-day symptom score under treatment versus usual care among adults eligible for either strategy at a specified clinic. The outcome, time, population, contrast, and scale are all stated. Whether a particular analysis can identify that difference is a separate question.

Specify Each Part of the Question

The population should be defined substantively, not just as “the study sample.” State the group to which the answer is intended to apply: for example, adults with a specified condition who meet stated eligibility criteria at treatment initiation. The sample is the set of people whose records enter the analysis; it may be a subset of the target population, or may represent a different population altogether.

The treatment contrast names both sides of the comparison. In observational research, treatment can mean a treatment strategy, an exposure, or receipt of a particular intervention. For this tutorial, the key reporting task is to identify the contrast precisely enough that the two potential outcomes have meaning. Avoid labels such as “treated” and “control” when they do not say what people actually received or what alternative is being compared.

The outcome needs a defined measure and the follow-up time needs a defined endpoint or interval. “Clinical outcome” and “after treatment” are too vague to identify a single target quantity. The effect scale also matters: a risk difference, risk ratio, and odds ratio are different estimands, even if they use the same population, treatment contrast, and outcome.

Common Target Estimands

A familiar target is the average treatment effect (ATE) in a specified population \(P\): the average difference between the potential outcomes under treatment and comparator in that population. Its interpretation depends on which population \(P\) denotes; “the ATE” without a population is incomplete.

Another target is the average treatment effect among the treated (ATT), the average contrast among people who received treatment. It answers a different question from the ATE: the ATT concerns the treated group, not necessarily everyone eligible for either strategy.

The average treatment effect among the untreated (ATU) averages the contrast among people who received the comparator. These targets can differ when treatment effects vary between people or groups. A report should name the target directly instead of relying on readers to infer it from a method or software setting.

There are also targets defined for a population with particular treatment-propensity characteristics, such as an overlap population emphasizing people for whom either strategy was plausible. Such a target is not automatically the full eligible population. The population definition and the effect scale should still be reported explicitly.

When the Analysis Targets a Narrower Population

A stated target and an analysis target can diverge. The study may begin with a broad intended population but exclude people without comparable treatment options, with incomplete records, or outside a prespecified eligibility range. Matching or other analysis choices may also restrict which individuals contribute information. If the effect is then estimated for a selected subgroup, readers should not be told that it automatically represents the original population.

Let \(S=1\) indicate membership in the narrower group actually targeted, such as people who remain eligible after a restriction. The contrast in that group is \(E[Y(1)-Y(0)\mid S=1]\), whereas the original population target is \(E[Y(1)-Y(0)]\). These are equal only under conditions that make the average effect the same in the two populations; equality should not be assumed merely because both contrasts use the same treatment and outcome.

The distinction can be seen by splitting the original population into the included and excluded groups. If \(p=P(S=1)\), then:

This follows from the law of total expectation, applied to the individual contrast \(Y(1)-Y(0)\) across the two mutually exclusive groups. Unless \(p=1\) or the two group-specific average effects are equal, the effect in the included group need not equal the original population average. Thus, a restriction can change the estimand, not just reduce the sample size.

Choosing and Checking the Target

The choice among the ATE, ATT, ATU, or a restricted-population effect should follow the scientific question and intended use of the result. If the question is about offering treatment to all eligible patients, a population-wide target may be natural. If the question concerns the effect for people who actually received treatment, the ATT may be more directly aligned. These are alternatives, not interchangeable names for one quantity.

A useful reporting check is to compare the proposed interpretation with the population that supplied the identifying information. If an analysis excludes a group or depends on comparisons available only for a subset, ask whether the estimand has changed. State the narrower target explicitly when it has. A sample size reduction alone does not prove that the target changed, but a restriction defining who the result concerns does.

Causal interpretation also depends on assumptions. At a minimum, treatment and outcome must be defined consistently enough for the potential outcomes to be meaningful; the measured treatment must correspond to the stated strategies; relevant confounding must be adequately addressed; and both strategies must be possible in the target population where their effects are compared. These are often described using consistency, conditional exchangeability, and positivity. Their plausibility is not established by writing a precise estimand, and a report should not present an association as a causal effect solely because it has named a causal target.

Common Mistakes and Alternatives

• Calling the sample the population. “Among patients in the analysis” may describe who contributed data, but it does not necessarily identify the population the study aims to inform. Name the target population and disclose restrictions.
• Leaving out the effect scale. “The treatment reduced the outcome” is ambiguous. State whether the contrast is, for example, a difference in mean score, a risk difference, or a ratio of risks.
• Confusing the ATE with the ATT. An effect among treated individuals is not automatically an effect among all eligible individuals. Name which average is intended.
• Assuming a precise question guarantees a valid answer. Defining the target is necessary for interpretation, but it does not remove confounding or establish that the observed data identify that target.
• Generalizing a restricted result without justification. If the analysis targets a subgroup, describe that group and avoid implying the result automatically applies to people outside it.

The main alternative to a broad population target is not an imprecise question; it is a clearly stated narrower target. A narrower estimand may be scientifically useful, especially where treatment choices are genuinely comparable, but readers need to know that it is narrower and why it was selected. Reporting both the intended population and the analysis target prevents a method-driven restriction from disappearing in the interpretation.