Is every estimand causal?

No. Descriptive quantities can also be estimands.

Can different estimators target the same estimand?

Yes, though they may rely on different assumptions or have different statistical properties.

Is an estimate the true estimand?

It is a data-based approximation whose error and uncertainty need assessment.

AI Glossary / Causal inference and experimentation / Intermediate

Estimand

Definition

An estimand is the precise quantity that a study aims to learn about. It defines the target of estimation, such as a population mean, a difference in risks or an average causal effect, before choosing the statistical method used to estimate it.

What it means

The estimand answers “What exactly are we trying to know?” An estimator answers “What calculation will we use?” An estimate is the numerical result obtained from a particular dataset. Keeping these separate prevents a convenient statistical output from quietly replacing the question that motivated the study.
An estimand can be descriptive or causal. A population's average delivery time is descriptive; the difference in average delivery time under two scheduling policies is causal.
Precision requires specifying the relevant population, outcome definition, time period and comparison where applicable. Two studies can use similar data and methods yet answer different questions because their estimands differ.

How it works

Start with the decision or scientific question. Translate it into a target quantity defined independently of the particular sample. For a causal question, specify intervention alternatives and the population over which the contrast is averaged. Clarify how events such as nonparticipation, switching strategies or incomplete follow-up affect the target.
Next ask whether the estimand is identifiable from the proposed data and assumptions. Choose an estimator suited to that target and evaluate its properties. Report the resulting estimate with uncertainty and limitations. A technically correct calculation can still be unhelpful if the estimand is poorly chosen or does not match the decision being made.

Three practical examples

A service team defines its target as the median waiting time among all eligible requests during a particular quarter.
A product experiment targets the difference in 30-day completion probability under two interface versions.
A program evaluation distinguishes the effect of offering training to everyone eligible from the effect among people who actually participate.

Think of it this way

Think of the estimand as the destination, the estimator as the route and the estimate as where one trip ends. Different routes can aim at the same destination, and a carefully traveled route can still lead to the wrong destination if the question was misdefined. The analogy is conceptual: statistical estimation also involves uncertainty and assumptions about what the data reveal.

Worked example

Suppose an organization asks whether a new reminder improves course completion. One possible estimand is the difference in completion probability within 30 days if all eligible learners are offered the reminder versus if none are offered it. Another is the difference among learners who actually open the reminder.
These are different targets because opening may depend on motivation and other characteristics. Restricting analysis to openers does not simply provide a cleaner estimate of the original offer effect.
For the first target, a hypothetical randomized offer study observes completion rates of 0.62 and 0.56. The difference-in-proportions estimator produces an estimate of 0.06, or six percentage points, before uncertainty is quantified. The estimand is the underlying population contrast; 0.06 is the sample-based estimate, and subtraction of the two sample proportions is the estimator.

Illustrative case study

Fictional case: Maple Learning commissions an evaluation of a reminder system. Managers want to know whether enabling it for all eligible learners improves completion, but the draft analysis includes only people who clicked a message.
The statistician identifies the mismatch and writes the estimand before revising the analysis plan. The team decides how to handle learners who never read the message and how long to follow completion. It then selects a design and estimator aligned with that target. This case illustrates avoiding a question switch; it does not assert that the reminder works or that one estimand is always preferable.

Common misconceptions and mistakes

An estimand is not the observed numerical estimate or the software function used to compute it. A regression coefficient may be an estimand in some settings, but it should not be assumed to answer a causal question without further definition and assumptions.
Do not change the target silently when excluding observations, altering follow-up or restricting to participants. Missing data concern what is observed; they should not automatically redefine whose outcomes matter. A narrow confidence interval around the wrong target does not make the study useful for the intended decision.

Business use

Explicit estimands help business teams agree on what success means before collecting or analyzing data. They reduce disputes caused by one team reporting a metric for active users while another expects a result for everyone eligible.
Write the target in plain language and, where useful, in notation. Include population, outcome, comparison and timing. Distinguish policy effects from effects among selected users. Record any changes to the target and explain why they were needed. Once the estimand is clear, the team can assess whether the available data and method can support it, rather than letting the dashboard's easiest calculation determine the conclusion.

Check your understanding

A study aims to estimate average waiting time, uses the sample mean, and obtains 12 minutes. Which item is the estimand?

Answer: The target population's average waiting time. The sample-mean calculation is the estimator, and 12 minutes is the estimate produced from the observed data.

Related terms

Learn next

Learn estimand, estimator and estimate together using one concrete question. Then compare how changing the population, follow-up period or treatment strategy changes the target. Only after that choose a statistical method and examine what assumptions it needs.

Estimate

Sources

Frank Harrell, Glossary of Statistical Terms, estimand/estimator/estimate: https://hbiostat.org/glossary/ Miguel A. Hernán and James M. Robins, Causal Inference: What If: https://miguelhernan.org/whatifbook Author glossary excerpts consulted. Examples are original and nonclinical; full methodological and editorial review remain pending.

Previous: Essential matrix Next: Estimate