How to Write a Statistical Analysis Plan
Fix the analysis before unblinding, in terms that name what is being estimated.
What a how-to is not
A how-to is SPEQ’s practitioner method, not a procedure. It does not replace your own SOP, it is not a validated approach, and the judgement calls in it belong to your quality unit.
The analysis plan exists to remove discretion at the point where discretion is most dangerous — after the data exist. Its modern centre of gravity is the estimand: a precise statement of the treatment effect to be estimated, including what happens when participants stop treatment, take rescue medication, or die. A plan that specifies methods without specifying what is being estimated has automated a calculation without defining the question.
- 1
Define the estimand before choosing a method
Specify the treatment condition, the population, the endpoint, the strategy for each intercurrent event, and the population-level summary. These attributes together state what is to be estimated; the analysis method is chosen afterwards to estimate it, never the other way round.
- 2
Enumerate the intercurrent events and pick a strategy for each
Treatment discontinuation, rescue medication, death before the outcome is assessed — each needs a stated strategy, and different events in one trial may warrant different strategies. Leaving them unaddressed does not avoid the question; it answers it implicitly through whatever the software does with the missing values.
- 3
Define the analysis populations unambiguously
State how each population is derived and who decides membership. Population assignment made after unblinding, or by a rule that requires judgement, reintroduces exactly the discretion the plan exists to remove.
- 4
Specify the primary analysis, then the sensitivity analyses that probe its assumptions
Sensitivity analyses test whether the conclusion survives departures from the assumptions underlying the primary analysis. They are not a set of alternative analyses to choose among after seeing results, and describing them that way in the plan is worse than omitting them.
- 5
State the multiplicity strategy and the testing order
Where multiple endpoints, doses or interim looks exist, define how the error rate is controlled and in what order hypotheses are tested. A hierarchy invented after the primary endpoint misses is not a hierarchy.
- 6
Finalise and date the plan before unblinding, and version any change afterwards
The plan must be final before access to unblinded comparative data. Changes after that point are possible but must be documented with their timing and rationale, and disclosed in the study report — an undated plan cannot demonstrate it preceded the data.
- !Methods specified without an estimand, so the plan describes a calculation rather than a question.
- !Intercurrent events unaddressed, leaving the strategy to whatever the software does with missing values.
- !Sensitivity analyses framed as alternatives to select among after seeing the primary result.
- !The plan finalised without a date that can be shown to precede access to unblinded data.
How to Write a Statistical Analysis Plan: frequently asked questions
Common questions on write a statistical analysis plan.
What is an estimand?
A precise description of the treatment effect the trial sets out to estimate. It is defined by the treatment condition, the population, the endpoint, how each intercurrent event is handled, and the population-level summary measure. Two trials with identical protocols and different intercurrent-event strategies are estimating different things.
Why do intercurrent events need a strategy rather than a rule for missing data?
Because they are not missing data. An event such as starting rescue medication happens, is observed, and changes the meaning of anything measured afterwards. The strategy states what the estimated effect should mean given that it occurred — a question that arises before any value goes missing.
What if the plan needs to change after data are available?
Change it, and document what changed, when, why, and what was known at the time. What is not acceptable is a silent change, or one made after seeing comparative results. The disclosure in the study report is what preserves the interpretability of the result.