The Estimand That Made Placebo Look Better

What an estimand is, briefly

An estimand is a precise statement of what a trial is estimating, including how it handles events that occur after randomisation - stopping treatment, starting something else, having surgery. Two analyses of the same data can produce different numbers not because one is wrong but because they answer different questions. The efficacy estimand asks what would happen if participants remained on randomised treatment without rescue. The treatment-regimen or treatment-policy estimand asks what happened, counting all data regardless of what participants did afterwards.

The usual direction, and the exception

Normally the treatment-regimen estimand gives a smaller effect, because it includes data from people who stopped the drug and regained weight. TRIUMPH-4's active arms behave exactly that way: -26.4% becomes -20.0%, and -28.7% becomes -23.7%. The placebo arm goes the other way. Its weight change moves from -2.1% under the efficacy estimand to -4.6% under the treatment-regimen estimand, more than doubling. A placebo arm that loses more weight when you count everything that happened is worth stopping over.

The most likely explanation

The arithmetic is unambiguous; the explanation is inferential, and the sponsor's topline does not spell it out. The reading that fits is that participants who left the placebo arm did not simply resume their previous trajectory. Some of them obtained effective weight-loss treatment outside the trial. Under the efficacy estimand those post-discontinuation observations are excluded or handled as if treatment had continued; under the treatment-regimen estimand they are counted as they occurred. The placebo arm therefore becomes, in part, a treated arm - and its measured weight loss rises. Other contributors are possible: differential dropout, intercurrent illness, or lifestyle intervention effects concentrated among those who left. None of these can be separated from the published topline.

Why it narrows the gap from both ends

This is the part that matters for interpretation. The drug arms shrink because discontinuers regain; the placebo arm grows because discontinuers get treated elsewhere. The measured difference between them falls twice as fast as either movement alone. At the higher dose the gap goes from 26.6 percentage points on the efficacy estimand to 19.1 on the treatment-regimen estimand - a 28% reduction in the estimated treatment difference driven entirely by the analytical question asked, not by anything that happened in the joints or the gut of any participant.

Why this gets worse every year

A placebo arm in an obesity trial is only a placebo arm if it stays untreated. That was easier to arrange when there was little effective treatment to leave for. It is now much harder: participants randomised to placebo in a 68-week trial have licensed, effective options available, and there is no ethical mechanism to prevent them from seeking those options. As effective agents spread, placebo arms in this field will drift further toward being partially treated arms, and treatment-policy estimands will drift further from efficacy estimands. This is a structural problem for the field's evidence base, not a flaw in any one trial.

What to do with this when reading a result

Three things. Check which estimand a quoted number comes from, because the same trial routinely publishes both. Never compare a number from one estimand against a number from another, even within the same trial. And when a placebo arm shows unexpected movement, treat it as information about the trial's conduct rather than as background noise - it is often the most revealing number in the table.

Quick reference