Randomised withdrawal is the cleanest design in this field and the least discussed

The Journal has read all three of the relevant reports in full, and the most striking feature of the coverage they generated is how much of it treated the results as a scandal rather than as an answer. That a treatment for a chronic condition stops working when it is stopped is not a finding about the treatment. It is a finding about the condition, and it places these drugs in the same category as antihypertensives, statins and inhaled corticosteroids, none of which anybody expects to work after they are discontinued.

What a randomised withdrawal design does

A randomised withdrawal design begins with an open-label lead-in during which all participants receive the active drug and escalate to a target dose. Those who tolerate it and complete the lead-in are then randomised, usually two to one or one to one, to continue the drug or to receive matching placebo, and both arms are followed for a defined period with weight as the primary endpoint.

The design has two properties worth naming. Because randomisation occurs after the response, it isolates the effect of continuing from the effect of having lost weight, which a conventional parallel-group trial cannot do. And because the population has been selected for tolerating the drug, the withdrawal arm is not a general population — it is an enriched one, which makes the arm comparison internally valid and limits how far the absolute figures generalise.

Regulators favour the design for chronic-use products precisely because it answers the duration question. Its cost is ethical rather than statistical: participants who have achieved a substantial benefit are randomised to lose it, which is defensible only where the question is genuinely open and the follow-up is bounded. The Journal notes that all three withdrawal designs in this class published their regain data in full, which is more than can be said for several older obesity programmes.

Maintenance has a thinner evidence base than escalation, which is the opposite of where people spend their time. That imbalance is the largest gap between what has been studied and what is lived, and it is not closing quickly.

The STEP 1 extension: two-thirds back within a year

The pivotal semaglutide obesity trial ran for sixty-eight weeks with a mean weight reduction of approximately 14.9 per cent on 2.4 mg weekly against 2.4 per cent on placebo.¹ An extension followed a subset of participants for a further fifty-two weeks after both the drug and the lifestyle intervention were withdrawn, which makes it an off-treatment observation rather than a randomised withdrawal.

By week 120 — a year after stopping — participants who had received semaglutide had regained approximately two-thirds of the weight they had lost, finishing on average around 5.6 per cent below their original baseline against approximately 0.1 per cent for the former placebo group.² Improvements in glycaemic parameters, blood pressure and lipids reverted broadly in step with the weight.

Two details are consistently dropped from summaries. The residual benefit was real: a mean 5.6 per cent reduction sustained a year after stopping is not nothing, and it is more than most non-pharmacological interventions achieve while they are still being delivered. And the lifestyle support was withdrawn at the same time as the drug, so the extension describes the removal of an entire intervention package rather than of a molecule.

STEP 4: the randomised switch to placebo

STEP 4 is the cleanest test of continuation in the semaglutide programme. All participants took semaglutide through a twenty-week escalation to 2.4 mg weekly, achieving a mean reduction of approximately 10.6 per cent. They were then randomised two to one to continue semaglutide or to switch to placebo for a further forty-eight weeks, with lifestyle support maintained in both arms.³

Those who continued lost a further 7.9 per cent, reaching roughly 17.4 per cent below their original baseline at week 68. Those switched to placebo regained approximately 6.9 per cent, ending near 5 per cent below baseline. The between-group difference of about fifteen percentage points is the effect of continuing treatment for a year, measured in a population that had already demonstrated a response.

The design detail that matters most is that lifestyle support continued in the placebo arm. This is not a comparison of drug against nothing; it is a comparison of drug plus support against support alone, in people who had lost weight on the drug. The regain observed is therefore what happens with the behavioural intervention still running, which makes it a more conservative estimate of the drug contribution rather than a less one.

Extending the interval between doses is a dose reduction expressed in time. Framed that way it connects to the maintenance literature instead of sitting apart from it as an unstudied practice.

SURMOUNT-4 and the steeper curve

SURMOUNT-4 applied the same architecture to tirzepatide with a longer lead-in. Participants escalated over thirty-six weeks of open-label treatment to their maximum tolerated dose of 10 or 15 mg weekly, achieving a mean reduction of approximately 20.9 per cent, and were then randomised one to one to continue or to switch to placebo for fifty-two weeks.⁴

Continuation produced a further mean reduction of about 5.5 per cent, for a total near 25.3 per cent at week 88. Withdrawal produced a mean regain of about 14 per cent of body weight, leaving that arm approximately 9.9 per cent below original baseline. The between-arm difference of roughly fifteen percentage points is similar in magnitude to STEP 4 despite the much larger initial loss.

The steeper regain in absolute terms is the expected consequence of a larger loss rather than evidence of anything peculiar to the agent. It is nonetheless the figure most often quoted without its denominator, and a fourteen-point regain from a twenty-one-point loss is a materially different statement from a fourteen-point regain from a ten-point loss. Both arms in this trial ended below where they began, and the arm that stopped ended roughly where the continued arm of the semaglutide programme did.

What the three designs agree on

The parent programmes establish the losses from which the withdrawal arms fall: approximately 14.9 per cent at sixty-eight weeks for semaglutide 2.4 mg in adults without diabetes, approximately 20.9 per cent at seventy-two weeks for tirzepatide 15 mg, and — the only continuous two-year comparator anybody has — approximately 15.2 per cent sustained at week 104 with treatment maintained throughout.⁵⁶

Read together, the withdrawal evidence supports four statements and does not support a fifth. Regain begins promptly after cessation, within weeks rather than months. It proceeds at a decelerating rate, with the steepest portion in the first three to six months. It does not, within twelve months of follow-up, return participants fully to their original baseline; residual reductions of roughly five to ten per cent persist at one year in all three datasets. And continued treatment maintains and usually extends the loss, with the extension diminishing as the plateau is approached.

The statement not supported is that the drugs cause weight regain, or that stopping leaves a person worse off than never having started. Nothing in these datasets shows overshoot above the original baseline at a group level. Every arm that stopped remained below where it began at the end of follow-up.

The Journal makes this point repeatedly because the contrary claim circulates widely and is often accompanied by a mechanistic story about metabolic damage. The withdrawal trials are the direct test of that claim and they do not support it. What they do support is the unremarkable proposition that a treatment for a chronic condition works while it is being taken.

The correspondence this department receives on stopping divides almost evenly between people frightened by regain figures they have seen quoted without denominators and people who stopped without difficulty and cannot understand the alarm. Both groups are reading the same trials. The difference is almost entirely a matter of which number was quoted to them and whether anybody explained what it was a proportion of.