Skip to content
PeptideHormone

Two numbers, one trial

Retatrutide’s pivotal trial reported 25.0% weight loss at the top dose. It also reported 28.3%, for the same people at the same week. Neither is an error: one counts what happened to everyone assigned the drug, the other models what would have happened had nobody stopped. Every modern obesity trial reports both, the label carries the smaller and the headline the larger, and the gap is wide enough to reorder drugs when the two kinds are mixed. What an estimand is, why the gap varies, and which number to use for what.

8 min read · reviewed November 2026

The same arm, the same day, two results

In the pivotal trial of retatrutide, the top-dose group lost 25.0% of body weight at 80 weeks. In the same trial, the same group, at the same 80 weeks, lost 28.3%. The first figure is in the journal paper (Jastreboff et al., NEJM, 2026). The second is in the company’s announcement (Lilly, 2026). Neither is a mistake and neither is spin in the usual sense. They are answers to two different questions, and the trial was designed in advance to answer both.

This happens in every modern obesity trial, and it is the single most common reason two articles about the same drug quote different numbers. Once you know which question each number answers, you can tell at a glance which one you are being shown, and whether a comparison between two drugs is fair.

What goes wrong in the middle of a trial

A trial randomizes people to a drug or a placebo and weighs them at the end. The complication is what happens in between. Over a year and a half some participants stop taking the drug: side effects, a move, a pregnancy, loss of interest. With a weight-loss drug this matters a great deal, because people who stop regain weight.

So at the final weigh-in, the drug group contains two kinds of people: those still on the drug, and those who stopped months ago and have partly rebounded. The statistician has to decide what to do with the second kind. There are two defensible choices, and each one defines a different quantity. Trial statisticians call that quantity the estimand: the precise thing the trial is trying to estimate.

Question one: what happened to everyone we assigned the drug?

The first choice is to use the weights as they really were. Someone who stopped at month four and regained most of the loss is counted at their real final weight. The result is the average outcome of a policy: prescribe this drug to this kind of person and see what follows, stopping included.

This is the treatment-policy estimand. Lilly’s trials use a closely related version it calls the treatment-regimen estimand. It is the smaller number, it is the primary analysis in the published papers, and it is the figure that appears on United States drug labels. Regulators prefer it for a plain reason: it keeps the protection of randomization intact and it does not depend on a model of what might have been.

Question two: what does the drug do if you keep taking it?

The second choice is to ask a hypothetical. Suppose nobody had stopped. Weights measured after a participant quit the drug are set aside, and a statistical model fills in the path they would have followed had they continued, based on the people who did.

This is the efficacy estimand; Novo Nordisk calls it the trial-product estimand. It is the larger number, and it is the one press releases tend to lead with. It is a real analysis with a real use: it is the closest the trial gets to the effect of the molecule itself, separated from the question of whether people tolerate it.

What it is not

The efficacy figure is often described as “on-treatment”, which makes it sound like a count of the people who finished. It is not. Dropping everyone who stopped would compare the hardiest drug-takers with the whole placebo group, and that is no longer a randomized comparison. The efficacy estimand keeps everyone in and changes only how their later data are treated. The distinction is technical, but it is why the number is respectable rather than a trick.

The gap, across four trials

Here are both figures for the top dose of four pivotal obesity trials. In every row the two marks describe the same people.

What happened (treatment-policy)If everyone stayed on drug (efficacy)15%20%25%30%Semaglutide (STEP 1, 68 weeks): 14.9% treatment-policy, 16.9% efficacy, gap 2.0 pointsSemaglutideSTEP 1 · 68 wk14.916.9+2.0Tirzepatide (SURMOUNT-1, 72 weeks): 20.9% treatment-policy, 22.5% efficacy, gap 1.6 pointsTirzepatideSURMOUNT-1 · 72 wk20.922.5+1.6CagriSema (REDEFINE-1, 68 weeks): 20.4% treatment-policy, 22.7% efficacy, gap 2.3 pointsCagriSemaREDEFINE-1 · 68 wk20.422.7+2.3Retatrutide (TRIUMPH-1, 80 weeks): 25.0% treatment-policy, 28.3% efficacy, gap 3.3 pointsRetatrutideTRIUMPH-1 · 80 wk25.028.3+3.3
Mean % body-weight loss at the top dose, both analyses of the same participants. Different trials, populations and durations: read each row’s gap, not the order of the rows.

Two things are worth reading off the chart. First, the gap is not a constant: it runs from under two points to over three. It grows with the share of people who stop and with how long the trial runs, because both widen the distance between what happened and what would have. A drug with a large gap is not a worse drug, but the gap is telling you something about how many people did not stay on it.

Second, the gap is large enough to reorder the drugs if you mix the two kinds. Tirzepatide’s efficacy figure (22.5%) is higher than CagriSema’s treatment-policy figure (20.4%) and lower than its efficacy figure (22.7%). Which drug “wins” depends entirely on which pair of numbers a writer happened to pick up (SURMOUNT-1, REDEFINE-1, STEP 1).

Why trials now state both

Until recently a trial could report a result without saying which of these questions it had answered, and readers were left to dig through the methods. In 2019 the international body that harmonizes drug regulation added a requirement that trials name their estimand in advance, including how they will treat events such as stopping the drug (ICH E9(R1)). The obesity trials of the 2020s were designed under it, which is why they report two clearly labelled numbers where older trials often reported one ambiguous one.

That is progress, and it has a side effect. Two labelled numbers mean a sponsor can lead with the larger one and a critic with the smaller one, and both can say truthfully that the figure is from the trial.

Three more ways one trial gives several numbers

The estimand is the biggest source of mismatched figures, but not the only one. Before comparing any two numbers, check these as well.

  • Raw or placebo-subtracted. People on placebo lose some weight too, usually two to four percent, because everyone in the trial gets diet and activity counselling. A drug that produced 25% against a placebo's 4% has a placebo-subtracted effect of 21%. Both are quoted
  • Which dose. Trials test several doses. Headlines use the top one. The dose most people end up on in practice may be lower
  • Which week. Weight is still falling at the end of most of these trials. A result at 80 weeks is not comparable with one at 68, and an extension figure at 104 weeks is a different measurement again

Which number to use

Neither is the true one. They are for different jobs.

  • For what to expect across a group of patients. use the treatment-policy figure. It includes the reality that some people stop
  • For what the molecule does when taken. use the efficacy figure, and remember it assumes a level of persistence the trial itself did not achieve
  • For comparing two drugs. use the same kind for both, and treat the result as context. Different trials enrol different people for different lengths of time. Only a head-to-head trial is a ranking

This site quotes the published treatment-policy analysis as the main figure and names the other when it is the one in circulation, as in the retatrutide explainer and the GLP-3 bull case. When a page anywhere gives a single weight-loss percentage with no label on it, the safe assumption is that it is the larger one.

The short version

  • Two numbers, two questions. What happened to everyone assigned the drug, and what the drug does if everyone stays on it.
  • The smaller one is on the label. The treatment-policy (or treatment-regimen) analysis is the published primary result and the regulator's reference.
  • The larger one is in the headline. The efficacy (or trial-product) analysis is real, but it is a modelled hypothetical.
  • The gap is two to three points. Enough to reorder drugs if the two kinds are mixed.
  • Check dose, week and placebo too. Each can move a figure as much as the estimand does.

Educational reference on how clinical trial results are analysed and reported, summarized and simplified from the public record. Not medical advice. Trial figures are quoted to explain the method, not to rank or endorse any drug. Investigational, not approved: retatrutide and CagriSema. Approved in the United States: semaglutide and tirzepatide. Status checked November 2026.