Estimands changed the question. Most analysis plans still answer the old one

The estimand addendum to ICH E9 has been in force long enough that every statistical analysis plan now has a section headed Estimands. Reading a stack of them, a pattern shows up quickly: the section is present, correctly structured, and frequently describes an analysis method rather than a question.

That distinction is the entire point of the addendum, and it is easy to lose because the old habit is so well established. For decades the analysis plan named a population, named a model, and named a handling rule for missing data. The treatment effect being estimated was implied by those three choices rather than stated. When the choices were conventional, everyone read the same meaning into them.

What the addendum actually asks

An estimand is a description of the treatment effect, written so that two statisticians reading it would agree on what number the trial is trying to produce, before anyone discusses how to produce it. It has five attributes: the population, the treatment conditions being compared, the variable, the summary measure, and the handling of intercurrent events.

Four of those five were already familiar. The fifth is where the work is, and it is the one that reveals disagreement.

Intercurrent events are where trials differ

An intercurrent event is something that happens after randomisation and changes what the outcome means: a patient stops treatment, starts a rescue medication, or switches arms. The addendum offers five strategies for handling them, and choosing between them is a clinical decision rather than a statistical one.

Take rescue medication in a diabetes trial. Under a treatment policy strategy, the outcome counts regardless of rescue, and the estimand answers what happens if we assign this drug in practice, rescue included. Under a hypothetical strategy, the estimand answers what the outcome would have been had rescue not been available. Both are legitimate. They are different questions, they will give different numbers, and they support different label claims.

An analysis plan that says missing data will be handled by multiple imputation under a missing at random assumption has specified a method. It has not said which of those two questions the trial is answering. The reader has to reverse engineer the question from the method, which is exactly the situation the addendum was written to end.

The test for an estimand section is simple. Cover the method and read only the estimand. If you cannot tell what would count as a treatment success, the estimand has not been specified.

Why this reaches programming, not just statistics

The practical consequence lands on the analysis datasets. Different intercurrent event strategies need different flags, different observation windows and sometimes different derived endpoints in ADaM. A treatment policy estimand needs the post rescue observations retained and identifiable. A hypothetical estimand needs the point of rescue captured precisely enough to censor or impute from it.

If the estimand is settled late, those requirements arrive after the datasets are specified, and the correction is expensive. If it is settled early, the ADaM specification can carry the flags from the start and the supplementary estimands become variations on one structure rather than three parallel builds.

Sensitivity analyses are not a substitute

A common pattern is to specify one primary analysis loosely and then add a list of sensitivity analyses. The intent is reasonable, but sensitivity analyses answer a different question: they test whether the answer to this estimand is robust to assumptions. They do not resolve which estimand was being asked about.

Where different intercurrent event strategies are genuinely of interest, they belong as supplementary estimands, each stated in full, rather than as sensitivity runs on an underspecified primary.

A practical way in

The teams that handle this well tend to do one thing: they write the estimand before the method, in a meeting that includes the clinical lead, and they write it in plain sentences before it is put into the five attribute table. The table is good discipline for completeness. It is poor at surfacing disagreement, because it is easy to fill in a cell that everyone reads differently.

The plain sentence version forces the disagreement into the open while it is still cheap to resolve.