Trial data explorer

The single commonest failure in coverage of this drug class is a percentage quoted without the design that produced it. A weight-reduction figure is a function of the population randomised, the comparator, the trial duration, the background intervention, and — decisively — which of two conventional analyses the number comes from. Change any of those and the figure changes, sometimes by more than the difference between two molecules.

This table exists so that a reader can check a number against its design in one place. Every row carries the primary publication. Where a trial reported both a treatment-policy and a trial-product analysis, both are given; where only one was published, the table says which.

The estimand column is the point of this page

Two analyses are conventionally reported in the obesity programmes and they answer different questions. The treatment-policy estimand includes every randomised participant regardless of whether they kept taking the drug — it estimates the effect of a policy of offering the treatment. The trial-product estimand estimates the effect had every participant taken it as intended. The second is larger, in STEP 1 by two percentage points.¹

Neither is wrong and neither is more honest. But a source quoting 16.9% and a source quoting 14.9% for the same trial are not disagreeing about the data; they are quoting different analyses, and a reader who does not know that will conclude that one of them is lying. The Journal states the estimand every time, which is why the column exists.

What this table cannot be used for

It cannot be used to compare molecules. SURMOUNT-1 randomised a population with a different baseline weight from STEP 1, ran four weeks longer, and used a different background intervention; the difference between −20.9% and −14.9% is not the difference between tirzepatide and semaglutide. STEP 8, which randomised participants to semaglutide or liraglutide directly, is the only row in this table that supports a head-to-head claim, and it supports it only for those two molecules at those two doses.

It also cannot be used to predict an individual outcome. Every weight figure in this table is a mean, and the distribution around every one of those means is wide. Reported non-response — failure to reach 5% reduction — runs from roughly one participant in eleven to nearly one in five depending on the arm.²

The withdrawal rows

STEP 4 and SURMOUNT-4 are the two most important rows in this table and the two least quoted. Both randomised participants who had already responded to treatment either to continue or to stop, and both found rapid and substantial regain in the group that stopped: in SURMOUNT-4, participants who switched to placebo regained 14.0% of body weight over 52 weeks while those who continued lost a further 5.5%.³

These are the rows that determine whether this drug class is a treatment or a course, and therefore the rows that determine every health-economic model built on it. A reimbursement decision that assumes two years of treatment and a reimbursement decision that assumes twenty are different decisions about the same drug, and the withdrawal data is the reason.