Two headlines can use different percentages and still describe the same trial honestly. They may answer different statistical questions. A good reading method starts with the study design and then returns to the number, rather than letting the biggest number define the story.
First, identify the question
A randomized trial assigns treatments by chance, which helps reduce systematic differences between groups. It still applies to a defined population, follow-up period, and treatment schedule. Check whether the comparison is with placebo or another active medicine, and whether participants and researchers knew which treatment was given. The FDA trial guide sets out the role of protocols, controls, and participant criteria.
For example, a trial in adults with obesity without diabetes does not automatically answer a question about adolescents, pregnancy, or people with diabetes. A placebo comparison does not rank a drug against every other treatment. An open-label comparison may provide valuable direct evidence while remaining vulnerable to effects of knowing the assignment.
Separate a change from a difference
Consider a made-up trial: one group loses 12% of its starting weight and a control group loses 3%. The treatment group’s average change is 12%; the difference between groups is 9 percentage points. It is misleading to call those two numbers interchangeable. These figures are an arithmetic example, not a result from a real medicine trial.
Also distinguish percentages from kilograms and pounds. The same percentage corresponds to different absolute weights for different starting weights. A group average cannot establish a promised personal result. If a summary gives only the treated group’s number, look for the control group before drawing a conclusion.
Ask what happens to people who stop treatment
An estimand describes the treatment effect a study aims to estimate. Treatment-policy approaches can ask about effects regardless of events such as stopping the assigned treatment. Hypothetical or efficacy approaches can ask about a scenario in which specified interruptions did not occur. The exact definitions vary by study; the label is not enough. ICH/FDA estimand guidance calls for a clear account of the question and how these events are handled.
In REDEFINE 1, the peer-reviewed treatment-policy result for cagrilintide–semaglutide was a 20.4% mean weight reduction at week 68, compared with 3.0% for placebo. A sponsor communication also emphasized an estimate assuming continued treatment. Those are different questions, not a reason to silently choose the larger number. REDEFINE 1 paper.
Read the safety and uncertainty alongside the benefit
Look for adverse events, withdrawals, missing data, and confidence intervals. “No statistically significant difference” does not prove two options are identical. A statistically significant result can be too small or too narrow to resolve the decision you care about. Check whether a claimed subgroup finding was specified in advance or explored after the study.
A useful reading note has five lines: who joined; what was compared; how long it lasted; what analysis was used; and what important harms or unresolved questions remain. Add funding and conflicts of interest. Funding does not automatically invalidate a trial, but readers deserve to know who supported it.
Use the protocol to keep the story honest
The protocol describes the planned study question and methods. A registry helps identify the study and its intended outcomes. Compare them with the published report when a claim depends on a specific endpoint or subgroup. A new analysis can be useful, but it should be recognizable as new rather than quietly replacing the original question.
CONSORT 2025 is a reporting guideline, with a checklist and participant-flow diagram to make trial reports more transparent. It is not a certificate that every reported trial was designed well. CONSORT explanation. Complete reporting gives you more to assess; it does not eliminate the need to assess it.
Random assignment and blinding solve different problems
Random assignment addresses how participants enter the comparison groups. Blinding concerns who knows the assignment after it occurs. Allocation concealment concerns whether the next assignment can be known during enrollment. These are related safeguards, but one term should not stand in for all three. Their distinction is explained in the CONSORT guideline.
When a comparison is open-label, ask which outcomes might be influenced by knowing the treatment. That does not mean the result is worthless. It means the limitation belongs in the interpretation. A direct comparison can answer a valuable question while still leaving uncertainty about how the setting affects what was measured.
Missing outcomes are different from stopped treatment
A participant may stop the assigned medicine but still attend follow-up and provide outcome measurements. Another may stop attending entirely. Those events are different. An analysis that estimates a treatment-policy effect still needs a method for missing observations, and the method rests on assumptions. Check the reasons and the sensitivity analyses.
For an invented example, suppose ten people stop treatment but eight continue to be measured. It would be wrong to describe all ten as missing follow-up, and also wrong to assume their later outcomes are automatically identical to those who continue treatment. The study report should make the flow and analytical treatment visible.
Read a confidence interval as part of the estimate
A point estimate is one summary. A confidence interval expresses uncertainty using the study’s statistical method and assumptions. Read its width and what values it includes; do not treat the point estimate as an exact population truth. “Not significant” is not the same as a carefully designed demonstration of equivalence.
Imagine a fictional difference of two units with a broad interval that includes no difference. That does not establish that the interventions are identical. It indicates that this analysis has not precisely resolved the comparison. The example intentionally avoids a treatment recommendation or a numerical confidence interpretation for a real medicine.
Ask how harms were looked for
A table of adverse events depends on how the trial collected them, when it asked, what counted as an event, and which participants were included. CONSORT Harms calls for those methods and denominators to be clear. Harms reporting guideline. A statement that there were no new safety signals is less informative without the relevant follow-up and methods.
Common events, serious events, and treatment withdrawals answer different questions. An event occurring during treatment is not automatically proven to be caused by it. Conversely, a trial that observes few uncommon events may not be able to rule out every rare risk. Keep the claim as narrow as the design and the actual report permit.
Create a reading note before sharing the headline
- Identity: exact title, year, trial name, and identifier.
- Population: relevant eligibility, including diabetes status and age.
- Intervention: exact product and comparator, with shared background support.
- Outcome: what was measured and at which time point.
- Analysis: estimand, missing-data method, and any sensitivity checks.
- Benefit and harm: group results, uncertainty, and tolerability.
- Context: funding, publication status, and questions the study did not answer.
You do not need to become a statistician to use this note. Its value is that it slows down a conclusion when essential information is missing. If a source contains only a topline release, mark it that way and wait for the paper before claiming details that have not been made available.
A reusable evidence sentence
Try this template in your own notes: “In [population], [treatment] compared with [comparator] changed [outcome] over [duration], using [analysis]; the main limitations were [limitations].” If you cannot fill in those blanks from the paper or its protocol, keep the conclusion provisional. That is more useful than collecting percentages from unrelated studies.
