Weight-Loss Pills Compared: Why 12.4% vs 13.6% Isn't a Fair Fight

Oral weight-loss drugs are the medical story of the moment, and so is a tempting comparison. Lilly reported that the highest dose of orforglipron produced a mean weight loss of 12.4% at 72 weeks in its ATTAIN-1 trial, while a summary of the OASIS 4 trial put oral semaglutide at 13.6% after 64 weeks. So is semaglutide better? Statistics says: not so fast.

Same number, different questions

Even within ATTAIN-1, the headline depends on definition. The 12.4% figure uses an efficacy estimand, which counts participants while on treatment, whereas the treatment-regimen estimand, which counts everyone randomized, gave about 11.2% (Lilly summary, Diabetes on the Net). Follow-up length differs too, at 72 weeks against 64, and so do populations and protocols.

Why a naive test misleads

Suppose, purely for illustration, each trial had 500 patients per arm and a standard deviation of 8 points. The independent samples t-test calculator on 12.4% against 13.6% gives p ≈ 0.018, and the difference of means confidence interval runs from 0.2 to 2.2 points. It looks like a win for the second drug.

But that test treats the groups as if patients were randomized between the two trials. They were not.

Compare each drug with its own control

A better approach is an adjusted indirect comparison, introduced by Bucher and colleagues in 1997. It compares each drug with the placebo arm of its own trial, then compares those two differences, and their uncertainties add. In our illustration the placebo groups lost 2.1% and 3.4%. The effects against placebo become 10.3 and 10.2 points, a difference of 0.1 with a 95% interval from about −1.5 to +1.3. The apparent advantage disappears.

Randomization is preserved inside each trial, and that protection is what the method borrows. Even so, it assumes the trials are similar enough, a strong assumption when populations, doses and follow-up differ.

The most reliable answer is a head-to-head trial. Until one exists, treat cross-trial gaps of a point or two as hypotheses, not rankings. Dose groups within one trial can be compared directly, for example with the one-way ANOVA calculator.

None of this makes cross-trial numbers useless. They help generate hypotheses and plan future trials, and a large gap that survives adjustment deserves attention. The error is treating a casual comparison as a verdict. Ask whether populations, dose escalation and endpoint definitions match before trusting any ranking.

Beyond the average

Averages hide the spread of response. Lilly's summary reported the share of participants losing at least 10% of body weight as 54.6% on the highest dose against 12.9% on placebo, a gap of nearly 42 points (Lilly results). The difference of two proportions calculator turns such counts into an interval.

The same release reported that overall treatment discontinuation on the highest dose was 24.4%, roughly one participant in four. That matters for interpretation: a drug's effect among those who stay on it differs from its effect across everyone who started, which is exactly the distinction between the two estimands above.

Weight loss is also only one outcome. Tolerability, discontinuation, cost and how patients actually take a pill matter as much as a decimal. Our guide to reading a clinical trial lists the questions to ask. The numbers above are illustrative, not clinical advice.