CompoundBrief

Research literacy

How to Read a Clinical Trial Result: Endpoints, Estimands and Confidence Intervals

Published October 15, 2026 · 8 min read

Scope: This is a general guide to reading clinical trial reports. It uses no real trial results and makes no claim about any compound. Terminology follows the ICH E9(R1) estimand addendum and the CONSORT reporting statement, both linked in the sources.

A trial headline usually boils down to one number. The number is only meaningful once you know who was counted, what counted as success, how uncertain the estimate is, and what the study could not show. This guide walks through the pieces that most often change how a result should be read.

Start with the design

Random assignment is what lets a trial compare groups fairly. If participants are allocated by chance, differences between groups at the start are due to chance alone, so a later difference in outcomes can more plausibly be attributed to the intervention. Blinding (participants, clinicians and assessors not knowing who got what) protects against expectation shaping what is reported and measured. A trial without randomization or blinding can still be informative, but it answers a weaker question, and its results should be read as such.

Check the basics in any report: how many people were randomized, who was eligible, how long they were followed, and what the comparison group received (placebo, active comparator, or nothing). A result against placebo and a result against another treatment are not the same kind of evidence.

Primary endpoint, secondary endpoints, and surrogates

A well-run trial states its primary endpoint in advance, the one outcome the study was sized to test. Secondary endpoints are additional outcomes, and because each extra comparison raises the chance of a false positive, they are usually treated as supporting rather than decisive. Be cautious when a press release leads with a secondary or exploratory finding.

A surrogate endpoint is a measurement used as a stand-in for an outcome that matters more but takes longer to observe, such as a laboratory value standing in for a clinical event. Surrogates are convenient, but a change in the marker does not guarantee a change in the outcome it is meant to predict.

Intention-to-treat and per-protocol

In an intention-to-treat analysis, participants are analyzed in the groups they were randomized to, regardless of whether they completed treatment or followed the protocol. This preserves the benefit of randomization and reflects what happens when a treatment is offered in the real world, including people who stop early. A per-protocol analysis restricts attention to people who adhered, which can show what the intervention does when taken as directed but can also reintroduce bias, because the people who adhere may differ from those who do not.

Estimands: the question behind the number

Two analyses of the same trial can give different numbers because they answer different questions. The ICH E9(R1) addendum formalized this idea as the estimand: a precise description of the treatment effect being estimated. The key concept is the intercurrent event, something that happens after treatment starts and affects the outcome, such as stopping the study drug, switching to another treatment, or needing rescue therapy.

Different strategies handle those events differently. A “treatment policy” approach keeps all the data regardless of what happened, which resembles intention-to-treat and answers “what is the effect of being assigned this treatment?” Other approaches estimate the effect as if participants had stayed on treatment, which answers “what is the effect while the treatment is taken?” Neither is wrong. They are different questions, and the second usually gives a larger effect than the first when many participants stop early. When a report shows two figures for the same endpoint, check which estimand each one uses before comparing them.

Effect sizes and confidence intervals

A point estimate, such as an average difference between groups, is the best single guess from one study. A confidence interval shows the range of values reasonably compatible with the data. A wide interval means the estimate is imprecise, which is common in small trials. An interval that includes “no difference” means the study could not rule out no effect. A narrow interval around a tiny effect can be statistically solid and still not matter much in practice. Look for both the size of the effect and the interval around it, not just whether a result was called “significant.”

Dropouts and missing data

Every trial loses participants. What matters is how many, how evenly across groups, and how the analysis handled the gaps. If one arm loses far more people than the other, or the reasons for leaving differ, the remaining groups may no longer be comparable. Reports should include a participant flow diagram, as recommended by the CONSORT statement, showing how many were randomized, treated, followed up and analyzed.

Safety data deserve equal attention

Efficacy results get the headlines, but a trial’s adverse-event and discontinuation data are part of the result. Look at how many people stopped because of side effects, how serious events were defined, and whether the study was large and long enough to detect uncommon problems. A trial powered to show benefit on one measure is often too small to say much about rare harms.

A short checklist

Sources

← Back to all articles