Key Takeaways
- •The same trial reports different numbers under different estimands. One figure counts everyone as randomised regardless of what they did; the other estimates the effect if the regimen was followed as intended. The pivotal trials in this field routinely publish both.
- •Comparing across trials compounds the error. In the worked example below, the naive cross-trial gap is 7.5 points and the head-to-head trial measures 5.8.
- •A randomised head-to-head trial is the only comparison that settles anything. Cross-trial arithmetic produces a number, not an estimate.
- •Duration, population and comparator all move the number independently of the molecule.
- •A figure without a trial name, a duration and an estimand is not a result, it is a marketing figure that happens to be numeric.
- •None of this transfers to a research vial. Trials are run with pharmaceutical-grade material released against a specification.
Summary
The same clinical trial can publish two different headline figures from the same participants and the same data, and both are correct. The difference is the estimand: the rule for what to do with people who stopped following the assigned regimen. Count everyone as randomised and one figure results; estimate the effect if the regimen had been followed and a different figure results.
That matters because comparing two compounds across two separate trials compounds the problem. Duration, the population enrolled, the comparator, the background care and the strength studied all move a figure independently of the molecule. A cross-trial gap built by taking each compound's most favourable number from a different trial is not an estimate of anything; only a randomised head-to-head comparison measures the difference.
The practical upshot is a checklist rather than a conclusion: which trial, over how long, which estimand, in whom, against what, and was it the primary endpoint. And the figures do not transfer to a research vial in any case. They were produced with pharmaceutical-grade material released against a specification, and a research vial shares the sequence and nothing else.
A single clinical trial can publish two different headline figures for its primary endpoint. Both can be correct. Both can come from the same patients and the same data, analysed under two different pre-specified rules. A third figure, from a different trial against a different comparator, can sit between or beside them without contradicting either.
Anyone quoting one of those numbers at you has chosen it. Once you start comparing across trials, the gap manufactured by that choice can be larger than the real difference, and this article shows how, step by step.
This matters here because research compounds are often marketed on the strength of a clinical literature that describes a different product entirely. Understanding what a trial figure means is the difference between reading that literature and being marketed at with it. All Volta material is supplied for laboratory research use only.
The estimand: the choice that moves the number without touching the data

An estimand is the precise definition of what a trial is estimating. It sounds like statistical housekeeping. On a continuous primary endpoint in a long trial it can be worth one to two percentage points on exactly the headline figures people quote.
The problem it solves is real. In a long trial, people stop the assigned regimen, drop out, or start something else. What should the analysis do with them? Two defensible answers:
| Estimand | The question it answers | Effect on the number |
|---|---|---|
| Treatment policy (also: treatment regimen) | What happens to everyone assigned to this arm, regardless of whether they stayed on it? | Usually smaller. Discontinuers are counted as randomised |
| Trial product (also: efficacy) | What happens if the regimen is followed as intended? | Usually larger. Estimates the effect of the regimen as designed |
Neither is a trick. The first answers a question about a population offered an intervention; the second isolates the intervention itself. Regulators generally lead with the first. Promotional material generally leads with the second.
Both appear, explicitly and with confidence intervals, in the published reports of large Phase 3 programmes such as STEP 1 (New England Journal of Medicine, 2021) and SURMOUNT-1. In each, the two estimands for the same primary endpoint differ by one to two percentage points. Same trial, same participants, two correct numbers.
So before any comparison begins, each compound already has two legitimate headline numbers, and usually a range across the strengths studied. The single figure that reaches a product page is a selection.
What else differs between two trials
Estimand is one axis. There are at least four more, and every one of them moves a figure independently of the molecule.
| Axis | Why it changes the number |
|---|---|
| Duration | If the endpoint has not plateaued, a longer trial reports a larger change for the same intervention |
| Population | Enrolled populations differ in baseline characteristics and comorbidities, and those differences change the size of the measured effect on the same endpoint |
| Comparator | Against placebo the number is a difference from near-zero. Against an active comparator it is a difference from something that also moves the endpoint |
| Background care | Concurrent standard care differs between programmes in intensity and content |
| Strength studied | A range gets reported. A single number gets quoted |
None of this is hidden. It is in the methods section of every paper. It is simply absent from most summaries.
The arithmetic, done properly
Here is the demonstration. The figures below are illustrative, chosen to have the same structure as a real published case: two placebo-controlled trials of two compounds, each reporting both estimands, plus one randomised head-to-head trial. The endpoint is any continuous primary endpoint reported as percentage change from baseline.
| Trial | Compound | Treatment policy | Trial product |
|---|---|---|---|
| Trial A (placebo-controlled) | Compound X | 12.0% | 14.0% |
| Trial B (placebo-controlled) | Compound Y | 18.0% | 19.5% |
| Trial C (head to head) | Y against X | 17.3% against 11.5% |
The naive comparison, which is the kind usually in circulation: take Y's most favourable figure from Trial B and X's headline from Trial A.
19.5% − 12.0% = 7.5 points
That comparison mixes estimands. 19.5% is a trial product estimand; 12.0% is a treatment policy estimand. It compares one regimen followed as intended against another as people actually followed it.
Match the estimands and compare like with like:
| Comparison | Compound Y | Compound X | Gap |
|---|---|---|---|
| Naive, mixed estimands | 19.5% | 12.0% | 7.5 points |
| Both "as intended" | 19.5% | 14.0% | 5.5 points |
| Both "regardless of adherence" | 18.0% | 12.0% | 6.0 points |
Three different answers from the same four numbers, spanning two full percentage points, purely from which pairing you choose.
Now the trial that actually settles it. Trial C randomised participants to Y or X, head to head, under one protocol:
17.3% against 11.5%. Gap: 5.8 points.
The naive cross-trial comparison overstated the real gap by well over a percentage point. The like-for-like cross-trial comparisons bracketed it, one above and one below. Only the randomised head-to-head measured it. In the published case the example is modelled on, where a randomised head-to-head trial (SURMOUNT-5) was eventually run after two placebo-controlled programmes, the naive mixed-estimand comparison likewise overstated the head-to-head result, and in that case both like-for-like pairings came out below it rather than bracketing it. The direction is not something you can know before the head-to-head trial exists.
The lesson generalises: cross-trial arithmetic produces a number, and the number is not an estimate of anything in particular. It can overstate or understate. Its error is not predictable, which is precisely why it is not a substitute for a comparison that was actually run.
In Stock Now
A batch-specific Certificate of Analysis is available for every vial below. Supplied for laboratory research use only.
>99% HPLC purity standard. Read the Certificates of Analysis.
See the full Canadian catalogueA checklist for any percentage
Six questions. If a source cannot answer the first three, the figure is not usable.
- Which trial? A name or a registry number. "Studies show" is not a source.
- Over how long? Endpoints that have not plateaued make duration a lever.
- Which estimand? If unstated, assume the more favourable one was chosen, because it usually was.
- In whom? Baseline characteristics and the entry criteria.
- Against what? Placebo or an active comparator, and at what strength.
- Primary or secondary endpoint? A missed primary with an encouraging secondary is a missed trial.
The evidence grade tool scores a claim against this shape, the claim checker tests marketing wording, and the research literacy guide covers trial design in more depth.
Why none of it transfers to a vial
This is the part that cuts against a supplier's interest, so it belongs in plain sight.
Every trial figure is produced with material manufactured to pharmaceutical standards, characterised exhaustively, released against a specification, and studied under a controlled clinical protocol. The figure is a property of that whole system, not of a molecule in the abstract.
A research-grade vial shares the sequence and none of the rest. So a trial result is evidence about a molecule under specified conditions. It is not evidence about the contents of any particular vial, and a page that places a clinical figure next to a product is inviting a transfer that the data does not support.
Which is the same substitution described in what the human evidence actually shows, and the reason the identity and supplier verification questions are answered separately from the literature question.
Frequently Asked Questions
What is an estimand, in plain terms?
The exact definition of what a trial is measuring, including how it handles participants who stop following the assigned regimen. A treatment policy estimand counts everyone as randomised, so discontinuations are part of the figure. A trial product or efficacy estimand estimates the effect if the regimen was followed as intended, which gives a different figure from identical data.
Which estimand should I trust?
Neither exclusively. They answer different questions and a good paper reports both. The trial product estimand answers what a regimen does when followed as designed; the treatment policy estimand answers what happens to a population offered it. The error is not choosing one, it is comparing one compound's trial product estimand against another's treatment policy estimand.
Why can't I just compare percentages from two trials?
Because the two trials differ in duration, population, comparator, background care, strength and estimand, and each of those moves the number independently of the molecule. In the worked example the naive comparison was off by well over a percentage point against the head-to-head trial, and the direction of that error is not predictable in advance.
Does a bigger percentage mean a better compound?
Not on its own. It means a larger measured change on one endpoint, in one population, over one duration, under one analysis rule. Adverse events, discontinuation rates, what happens after the regimen stops and the size of the trial all belong in the same judgement, and none of them appear in a single headline figure.
Do these trial results apply to research-grade material?
No. Trials use pharmaceutical-grade material with controlled identity, content, impurity profile and sterility. Research-grade material has the same sequence and none of that assurance, which is why its own documentation has to be assessed separately. The trial tells you about the molecule; the certificate tells you about the vial.
How do I find the estimand in a paper?
It is in the statistical analysis section of the methods, and in well-reported trials it appears in the results tables as separate rows or columns, each with its confidence interval. If a summary quotes a number without saying which one it is, go back to the primary paper rather than trusting the summary.

































