What a placebo arm is for

A placebo arm is the control group in a randomised trial. Those people get the ritual of treatment — the visits, the pen or the injection, the weighing, often the same diet and activity counselling — and they do not get the drug. The point is subtraction. Whatever happens in both groups is the trial itself. Whatever happens only in the drug group is the thing the drug added. In the big obesity trials that number is not zero on the placebo side. In SURMOUNT-1 the placebo mean at 72 weeks was −3.1 percent. In STEP-1, at 68 weeks, it was −2.4 percent. The drug means sit next to those figures, or they are not means of a trial.

In short. The placebo group does everything except take the drug. They still lose some weight. The drug's number only means something next to that.

Nobody in these trials is sent home to do nothing. SURMOUNT-1, Jastreboff, Aronne, Ahmad and colleagues, New England Journal of Medicine, 2022, assigned 2,539 adults with obesity, and without diabetes, in a 1:1:1:1 ratio to tirzepatide 5, 10 or 15 mg, or placebo, for 72 weeks. STEP-1, Wilding and colleagues, New England Journal of Medicine, 2021, assigned 1,961 adults to semaglutide 2.4 mg or placebo for 68 weeks. Both programmes included a lifestyle intervention for every arm. People know they are in a study. They are weighed on a schedule. They talk to staff. Some of them would have lost weight that year anyway, and some who were unusually heavy on the morning of the baseline visit drift back toward their own average. That last one is regression to the mean, and it does not require a molecule.

In short. Both groups get the diet advice and the visits. Some of the loss would have happened anyway. The placebo number is that shared background.

Read the placebo percent as the background, then read the drug percent as background plus drug. In SURMOUNT-1 the treatment-regimen estimand — the average for everyone who was randomised, whether or not they stayed on the injections — was −15.0 percent at 5 mg (95 percent confidence interval −15.9 to −14.2), −19.5 percent at 10 mg (−20.4 to −18.5), and −20.9 percent at 15 mg (−21.8 to −19.9), against −3.1 percent on placebo (−4.3 to −1.9). About a third of the placebo group, 34.5 percent, still lost at least 5 percent of their body weight. A person can truthfully say the injections changed nothing and they lost weight. That person is already inside the placebo bar.

In short. On placebo, the average loss was about 3 percent, and about one in three people lost 5 percent or more. A success story can come from the group that did not get the drug.

The same trial prints a second, larger set of means if you ask a different question. The treatment-regimen, or treatment-policy, estimand asks what happened to everyone you randomised, including people who stopped. The efficacy estimand asks what the difference would have been if everyone had stayed on the assigned treatment. In the SURMOUNT-1 figure the efficacy means at week 72 are about −16.0, −21.4 and −22.5 percent at 5, 10 and 15 mg, with the placebo line near −2.4 percent, beside the regimen numbers of −15.0, −19.5, −20.9 and −3.1. Neither set is a lie. They are answers to two questions. A screenshot that picks the larger one and never says which question it answered has changed the experiment.

In short. One number includes people who stopped the drug. The other imagines that everyone stayed on it. The second looks bigger. Say which one you are quoting.

STEP-1's treatment-policy mean at week 68 was −14.9 percent with semaglutide 2.4 mg and −2.4 percent with placebo, a difference of −12.4 percentage points (95 percent confidence interval −13.4 to −11.5). That is not a head-to-head against tirzepatide. Different years, different people, a different drug, a different clock. Putting −14.9 beside −20.9 and calling it a contest is how two good papers become one bad sentence. The GIP page walks the chemistry. This page is only the comparison that makes the percent mean something.

In short. Semaglutide's trial and tirzepatide's trial were not the same experiment. You can put the papers next to each other. You cannot pretend they were one race.

n is the count of people behind the sentence. SURMOUNT-1's n is 2,539 randomised, split four ways, so each arm is on the order of six hundred. STEP-1 started with 1,306 assigned to semaglutide and 655 to placebo. The retatrutide Phase 2, Jastreboff and colleagues, New England Journal of Medicine, 2023, randomised 338 adults. The 12 mg mean at 48 weeks was −24.2 percent, and the interval around it ran from −26.6 to −21.8, with 62 people on that dose and the placebo mean at −2.1 percent. Look at the width. The Phase 3 interval around −20.9 percent is about two points from end to end. The Phase 2 interval around −24.2 percent is about eight. A mean can be dramatic and still be standing on a small platform. n is the platform.

In short. n is how many people were counted. Hundreds of people make a narrow band around the average. Sixty people make a wide one. The average can be large either way.

A p-value is not that platform, and it is not the weight change. P less than 0.001 means the difference was unlikely if the drug had done nothing, under the model the statisticians wrote down. It does not mean the effect was large. A tiny difference becomes 'significant' if you recruit enough people, and a large difference in forty people can fail a test because the bar is wide. Quote the percent, the interval, the n, and the placebo. If the sentence still needs a p-value after that, it is a footnote, not the result.

In short. Significant does not mean large. The size is the percent, next to placebo, with the range and the headcount.

Double-blind means the participant and the study staff are not told the assignment. It does not mean nobody can guess. Nausea, a slower stomach, and a belt that needs a new hole are information. In these trials a lot of people work out which arm they are in, and some of the staff can guess too. That is functional unblinding. It is a real crack in the design, and it is why symptoms that depend on belief — energy, mood, how hard the diet felt — are softer than the number on the scale. The scale is still the better object, because a kilogram is a kilogram. The placebo arm does not repair the crack. It keeps the visits and the counselling matched, so the crack is not the whole story.

In short. People often guess which injection they got, because they feel sick or they lose more weight. The placebo group still matters, because they had the same appointments.

A before-and-after is n of one, with no coin toss, no concealed assignment, and no pre-written rule for which photograph gets posted. The weeks when nothing happened are missing. The placebo rate you just read says that roughly one in three people in a careful trial lose 5 percent without the drug, and then go back to their life without a paper. A trial mean is a centre of a spread. The SURMOUNT-1 interval around −20.9 percent runs from −21.8 to −19.9, which is the uncertainty of the average, not the range of the people. Individuals land all over the place. Someone will lose less than the placebo average while taking the drug. Someone on placebo will lose more than a friend on the top dose. Neither anecdote moves the mean.

In short. One story is one person. The average is the centre of a crowd, and the crowd includes people who did well with no drug at all.

The comparison belongs to the product that was randomised: a formulated injection, on a protocol, with inclusion rules and a sponsor. A lyophilised research vial of a published sequence is a chemical with a chromatogram. It was not the arm in SURMOUNT-1, STEP-1, or the retatrutide Phase 2. Reading those papers is the right use of them. Treating the vial as if it had inherited the placebo-controlled result is a different claim, and it is not the one on the label. Laboratory research only. Not for human consumption.

In short. The placebo comparison belongs to the medicine in that trial. A research vial of the published chain was not in the draw. The paper is public. The vial is a reagent.