It should come as no surprise by now that placebo patients enrolled in obesity clinical trials can eventually figure out their assignment; the scale unblinds. And with trials running six to twelve months (or more), it’s a lot to ask patients to remain enrolled when they’re not losing weight. Case in point: in the trial for Lilly’s oral GLP-1, orforglipron, 6.2% of placebo patients dropped out because they were dissatisfied with their weight loss, while 2.5% started seeking other ways to lose weight, including commercial obesity medications. 

That dropout pull gets stronger with every impressive GLP-1 headline. Zepbound and Wegovy established a new efficacy baseline — 21% and 15% mean weight loss in their pivotal trials, respectively. Retatrutide's Phase 3 readout raised the efficacy ceiling yet again, with a US filing planned for early 2027. Structure's aleniglipron advanced into Phase 3 in August, and the late-phase pipeline promises more efficacy and less frequent dosing still.

That is the environment placebo patients are now comparing themselves against. When they can weigh their own experience inside a study against what they know is possible outside it, the motivation to stay in the study fades. 

Why This Matters for the Data

When trial participants infer they are on placebo and drop out, seek treatment elsewhere, or even switch to another trial (or two), the missingness in the data becomes informative. The comparison the protocol intended to make starts to diverge from the comparison the trial can actually support. The same trial gives different answers depending on how you count the people who didn’t adhere.

Statisticians manage this by asking the same question in more than one way. One analysis asks what the drug does if everyone stays on it — the efficacy estimand. Another counts everyone assigned to an arm regardless of what they actually did — the treatment-policy estimand (ATTAIN-1 calls it the treatment-regimen), the version FDA guidance generally expects as the primary. In a well-behaved trial, the two answers land close together. Placebo contamination pushes them apart: dropout and off-protocol treatment hit each analysis differently, and the per-protocol view — which regulators treat as a sensitivity analysis, secondary to the results they rely on — tells only part of the story.

And the problem is not confined to the placebo arm. In Lilly’s TRIUMPH-4 study, some participants discontinued retatrutide because of perceived excessive weight loss, particularly those starting at a lower BMI. That is the mirror image of placebo attrition. Still, it points to the same underlying reality: the scale unblinds.

The pivotal TRIUMPH-1 readout shows the size of the problem. On the 12mg dose, patients lost 28.3% of their body weight if they stayed on the drug (the efficacy estimand), but 25.0% once everyone assigned to the arm counts, including those who stopped and regained (the treatment-policy estimand). The placebo arm moved in the opposite direction: 2.2% weight loss under placebo alone, but 3.9% once off-protocol behavior counts. Placebo patients aren’t likely to lost extra weight on their own; it’s more likely they found treatment outside of the trial and stayed enrolled. Put the two shifts together and the measured treatment effect swings by five percentage points, depending on which question the analysis asks.

And five points is a fortune here. In a field where all the drugs work relatively well, what separates them in the eyes of regulators, payers, and prescribers is relative efficacy, tolerability profile, and convenience. Just a few percentage points of weight loss, or how it is measured, can affect the label, payer coverage, and competitive positioning. The market has already shown what those points are worth. Novo shed more than $90 billion in value the day CagriSema reported 22.7% weight loss against a projected 25%. Lilly lost about $100 billion the day orforglipron came in near 12% when investors had expected Wegovy-level results. Neither miss was larger than three percentage points. 

Incremental Fixes Are Not Enough

Some of the industry response is understandable. Extension phases can help with recruitment and retention messaging. Alternative comparator strategies, including active-comparator and putative-placebo approaches, are getting more attention as obesity trials start to resemble studies in other chronic cardiometabolic conditions. But these approaches come with their own operational and inferential demands, especially when inclusion criteria, endpoints, and background care differ across studies. They are part of the conversation, but not a universal fix. Sponsors can also test directly for prohibited medications, but that is expensive, and it does nothing about the upstream credibility problem once dropout or non-adherence is already substantial.

And none of them address the core analytical problem. The treatment-policy estimand counts whatever patients actually do — and what patients do depends on the world outside the trial: every new approval makes off-protocol treatment easier to get, so a placebo arm enrolled this year is more contaminated than one enrolled last year. Follow that to its uncomfortable conclusion: two drugs with identical therapeutic benefit can report different efficacy simply because their trials ran a year apart. And in a market that prices a few points of efficacy in the billions, a readout that depends on what year the trial ran is a problem.

What Needs to Change

First, dropout in obesity trials should be treated as part of the design rather than something handled during analysis. Placebo discontinuation is at least partly predictable from baseline features, early weight trajectories, tolerability, and the availability of approved drugs outside the trial, so it can be simulated during planning and incorporated into design assumptions. Teams should be testing how different randomization ratios, follow-up durations, rescue rules, and attrition patterns affect the readout before committing to a design. The same simulations can run the enrolled population through competitor regimens alongside the study drug, which makes design review a commercial question as much as a feasibility one: what will this readout look like next to the drugs it will be compared against?

Second, off-protocol treatment should be caught while the trial is running, not discovered at database lock. A model that predicts each placebo patient’s expected weight trajectory can also flag the patients whose observed trajectory no longer fits it, which can be the signature of someone who has started an approved GLP-1 on their own. That likelihood gives sponsors a reason to act, whether that means targeted blood testing for prohibited medications or a conversation with the site, and it lets the analysis account for the deviation instead of absorbing it.

Third, design can reduce the damage from dropout, but it cannot eliminate it — and once it happens, the question moves to the analysis. Sensitivity analyses that account for differential dropout should be prespecified in the protocol. The primary analysis is anchored in FDA guidance and is not the thing to change. But dropout in obesity is differential — the two arms lose patients at different rates and for different reasons — and analyses that address this directly are worth planning for up front. Several approaches already exist, including censoring-weighted analyses, reference-based imputation, and models that predict how each discontinuing patient would have progressed had they stayed on their assigned treatment. These are most useful when they are specified in the protocol from the start, rather than constructed after the data are in.

All three depend on the same thing: a model of each enrolled patient’s expected control-arm trajectory that is calibrated, validated, and used in applications that are able to hold up to regulatory review.

Placebo attrition in obesity shows that design assumptions inherited from an earlier era of drug development are collapsing under the weight of the current commercial landscape. When a single percentage point of weight loss dictates billions of dollars in market capitalization, we can no longer afford to treat clinical trial controls as passive, predictable baselines. Placebo patients are active participants in a hyper-competitive healthcare market, and their behavioral responses to trial unblinding are actively distorting clinical data. Moving forward, the sponsors who succeed won’t just be those with the most potent molecules—they will be the ones who treat protocol adherence and control-arm behavior as dynamic variables to be modeled, anticipated, and engineered directly into the trial design itself.

With the weight-loss GLP-1 market generating $55 billion in the past year — up 75% — and roughly 190 obesity medicines in the industry pipeline, every one of these challenges will intensify. When the field changes this much, trial design and analysis have to change with it.

And obesity is only the entry point. As GLP-1 franchises expand into follow-on indications, the same evidence problems compound — the sponsors who solve them in obesity carry that advantage into everything that comes next.

This is not a problem we are watching from the sidelines. Unlearn is working with obesity sponsors on these exact challenges today, generating digital twins as a simulated control group for each enrolled patient and correcting for differential dropout. We take validation seriously, and we are doing that work the hard way: testing these approaches against real, held-out trial data before we make claims on them.

If your team is navigating a placebo arm you no longer fully trust, we would welcome the conversation.