Question 16 of 20 Intermediate Practical
How would you control type I error when a trial has several key endpoints?
Short answer
Pre-specify a strategy that keeps the overall chance of any false positive claim at the planned level, usually 5 percent. Options include a single primary endpoint, a fixed testing order, splitting alpha with Bonferroni or a Holm procedure, or a graphical method that passes alpha on when a test succeeds.
Why
Each test at 5 percent carries its own chance of a false positive, so several tests together carry more. With three independent endpoints, the chance of at least one false positive is 1 minus 0.95 cubed, about 14 percent. ICH E9 says multiplicity can come from several primary variables, several treatment comparisons, repeated time points or interim analyses, and that the protocol must say how it is handled.
The cleanest fix is to avoid it: pick one primary endpoint or one composite. Where several claims matter, FDA's October 2022 guidance on multiple endpoints describes the main tools. Bonferroni divides alpha, for example 0.0167 for each of three tests.
Holm tests the smallest p value first and relaxes the threshold as tests succeed. Fixed sequence testing spends the full alpha on each endpoint in a pre-specified order and stops at the first failure. Graphical methods pass alpha along weighted paths.
The order should follow clinical importance and likelihood of success. Exploratory endpoints outside the strategy can be reported but cannot support a claim.
A strong answer covers
- Family-wise error rate and why it inflates
- Sources of multiplicity listed in ICH E9
- Bonferroni, Holm, fixed sequence and graphical methods
- Ordering by clinical importance and chance of success
- Endpoints outside the strategy cannot support claims
Common mistakes
- Testing secondary endpoints at 5 percent after the primary fails.
- Choosing the testing order after seeing the data.
- Splitting alpha with Bonferroni across highly correlated endpoints without noting that it is conservative there.
The interviewer might ask next
What happens in a fixed sequence if the second endpoint fails but the third has p = 0.001?