Primary endpoint hierarchies and why order matters — one year on posts 31–60

Composite endpoints should be read component by component. A composite driven entirely by its softest component is a different finding from one where the components move together.

A qualification I should have led with rather than closed on.

Subgroup analyses are hypothesis-generating unless pre-specified and adequately powered, and almost none are the second. The interaction test matters more than the subgroup point estimate.

Reporting the observation and leaving the explanation open deliberately.

Practical answer on Primary endpoint hierarchies, since the theoretical one is upthread: do the simplest check first, write down the result, and only then decide whether the complicated explanation is needed. It usually is not.

Post #45 put the caveat in the right place and I want to underline it.

Genuine question rather than a rhetorical one: has anyone here actually observed Primary endpoint hierarchies, as opposed to read about it? The thread is long and I cannot tell.

Risk of bias: structured appraisal of internal validity. Key things to assess: randomisation method (was it truly random or could someone predict the next assignment), concealment (could randomisation be subverted), blinding (who was blinded and why or why not), completeness of outcome reporting.

Open-label design: unblinded trials admit expectation effects. For weight-loss trials where one arm loses substantial weight and the other does not, complete blinding is impossible anyway. The unblinded nature is a limitation worth noting.

Absolute numbers, not just relative: a 30% relative reduction tells you the ratio but not the practical magnitude. The event rate in each arm and the difference between them tells you how many people benefit.

If it helps: the failure mode here is usually boring rather than dramatic.

The arithmetic in post #51 is right; the assumption feeding it is the part to check.

A note on scope: what I am saying about Primary endpoint hierarchies applies to the case in the first post and I would not extend it further without checking.

Post #51 put the caveat in the right place and I want to underline it.

Surrogate endpoints: an endpoint that is not the outcome that matters but is measured as a stand-in. HbA1c is a surrogate for long-term glucose control and the short-term complications it prevents. Weight loss is a surrogate for metabolic health and long-term outcomes. Surrogates are useful but not identical to the endpoint that matters.

That is what I would do. It may not be what is correct.

I think the Primary endpoint hierarchies question is answerable and has not been answered, which is a more optimistic position than most of this thread.

On Primary endpoint hierarchies, the part that usually goes wrong is that the question is asked as though it has one answer. It has a range, and the width of the range is the interesting bit.

If you can post the two or three numbers you are working from, several people here will check the arithmetic rather than argue about the conclusion.

Where I part company with post #56, and it is a narrow parting.

A treatment-policy estimand asks what happens to people assigned to a strategy, including those who abandon it. A hypothetical estimand asks what would have happened had everyone continued. Both are legitimate and they give different numbers.

That has been true for the cases I have seen and I have not seen many.

Post #56 is the version of this I will quote in future. One addition.

Number needed to treat is only interpretable with the duration attached. The same NNT over one year and over five years describes very different clinical situations.

It is the sort of thing that seems obvious in retrospect and was not at the time.

Open-label design: unblinded trials admit expectation effects. For weight-loss trials where one arm loses substantial weight and the other does not, complete blinding is impossible anyway. The unblinded nature is a limitation worth noting.

Anyone who has looked at this more carefully, please correct the record.

Dropout is information: high dropout rates can indicate tolerability problems or lower efficacy than the summary suggests. Where the analysis handled dropouts matters. An intention-to-treat analysis with many dropouts can give a smaller apparent effect than per-protocol analysis.