An estimand is a precise description of the treatment effect that a clinical trial is designed to measure. Think of it as the specific question a trial is trying to answer, spelled out in enough detail that everyone involved, from doctors to statisticians to regulators, agrees on exactly what “success” means before the data are even collected. The concept became a formal requirement in drug development after international regulators endorsed a framework for defining estimands in November 2019, with the FDA publishing its final guidance in May 2021.
Why the Term Exists
Clinical trials seem straightforward on the surface: give one group a drug, give another group a placebo, and compare results. But real trials are messy. Patients drop out. Some switch to a different medication partway through. Others start taking additional “rescue” treatments because the study drug isn’t working fast enough. Someone might die from an unrelated cause before the trial ends.
Each of these events creates an ambiguity. If a patient stopped taking the drug after two weeks due to side effects, should their final health outcome still count toward the drug’s score? What about a patient who added a second medication on top of the study drug? Depending on how you handle these situations, the same trial data can tell very different stories about how well a treatment works. Before the estimand framework existed, trial teams often made these decisions late in the process, sometimes inconsistently, and the reported treatment effect could easily be misunderstood by regulators, doctors, or patients reading the results.
The framework forces everyone to define the target of measurement upfront, before data collection begins, so the trial design, statistical analysis, and interpretation all point at the same question.
The Five Parts of an Estimand
Every estimand is built from five attributes. Together, they form a complete, unambiguous description of the treatment effect a trial aims to quantify.
- Treatment: What intervention is being compared to what? This includes the specific drug, dose, and schedule, as well as the comparator (placebo, standard care, or another drug).
- Population: Who are the patients? This defines the group the results should apply to, such as adults over 50 with moderate heart failure, or children with a specific genetic condition.
- Variable: What outcome is being measured, and when? For example, change in blood sugar levels from the start of the trial to week 26, or survival time from randomization.
- Handling of intercurrent events: What happens when real-world complications arise during the trial? This is the attribute that makes the framework distinctive, and it gets its own section below.
- Population-level summary: How will the results be compared across groups? Common choices include the difference in average outcomes between groups, a risk ratio, an odds ratio, or a hazard ratio for time-to-event outcomes like survival.
What Intercurrent Events Are
Intercurrent events are things that happen to patients after they start the trial that affect either the interpretation or the existence of their outcome data. They are not errors or protocol violations. They are predictable complications of running a study with real human beings. Common examples include:
- A patient discontinues treatment because of side effects
- A patient stops treatment because it isn’t working
- A patient starts taking rescue medication on top of the study drug
- A patient switches to a completely different treatment while still enrolled
- A patient dies before the trial ends (in studies where death is not the primary outcome)
Context matters for how these events are classified. Death is an intercurrent event in a pain relief trial because it prevents the pain measurement from being collected. In an oncology trial studying survival, death is the outcome itself.
The estimand framework requires trial designers to anticipate which intercurrent events are likely and to decide, in advance, how each one will be handled in the analysis. This is a major shift from older approaches where these decisions were often made after the data came in.
A Real-World Example
The PIONEER 1 trial, which tested an oral diabetes drug against placebo, illustrates how a single trial can define more than one estimand for the same outcome. The trial measured changes in blood sugar (HbA1c) and body weight over 26 weeks, but it asked two distinct questions about those outcomes.
The first estimand, called the treatment policy estimand, asked: what is the average difference in blood sugar change between the drug group and the placebo group, regardless of whether patients stopped taking the study drug or started rescue medication? This reflects what happens in the real world, where some patients inevitably switch or add treatments, and it captures the overall effect of a strategy that starts with the study drug.
The second estimand, called the trial product estimand, asked a different question: what would the average difference have been if all patients had continued taking only the study drug for the full 26 weeks, with no rescue medication? This isolates the drug’s biological effect under ideal conditions.
Both questions are legitimate. The first tells regulators and doctors what to expect when the drug is prescribed in routine practice. The second reveals how potent the drug is when patients actually take it as directed. Neither answer is “right” or “wrong.” They simply target different aspects of the treatment effect, and the estimand framework makes the distinction explicit rather than leaving it buried in statistical footnotes.
Estimand vs. Estimator vs. Estimate
These three terms are easy to confuse, but they refer to different stages of the same process.
The estimand is the target: the precise treatment effect you want to know. It exists at the population level and, in practice, can never be observed directly because you would need complete data on every eligible patient, with no missing information.
The estimator is the method you use to get at that target. It is the statistical formula or algorithm applied to trial data. For example, calculating the difference in sample means between two groups, or using a survival curve method to estimate how long patients live. Choosing the wrong estimator for a given estimand is like using a bathroom scale to measure your height. The tool needs to match the target.
The estimate is the actual number you get when you run the estimator on your data. It is always an approximation of the estimand because trial participants are a sample of a larger population, not every patient completes the study, and real-world messiness introduces uncertainty. A trial might produce an estimate of “2.1 percentage points greater reduction in blood sugar with the drug versus placebo.” That number is the estimate. The thing it is trying to approximate is the estimand. The formula that produced it is the estimator.
Why It Matters Beyond Statistics
The estimand framework sounds technical, but its impact is practical. Before this framework, different stakeholders in a trial could hold different mental models of what the trial was measuring. A clinician might assume the results reflect what the drug does when patients take it consistently. A regulator might assume the results reflect real-world outcomes including patients who stopped treatment. A statistician might handle missing data in a way that aligns with neither assumption. The final number could satisfy everyone superficially while actually answering a question nobody quite intended to ask.
By requiring an explicit, five-part definition of the treatment effect before the trial begins, the framework aligns these perspectives. It pushes clinical teams and regulators into early conversations about what matters most: are we asking what happens when patients take this drug as directed, or what happens when doctors prescribe it knowing some patients will stop? These are fundamentally different questions with different implications for treatment guidelines, insurance coverage, and patient expectations.
A single trial can define multiple estimands, each capturing a different facet of the treatment’s effect. This also makes published results more transparent for anyone reading them. When a trial report specifies its estimand, you can judge whether the question it answered is the question you care about, rather than assuming all trials measure the same thing.