,
14 tweets,
3 min read
Read on Twitter
Trialists, statisticians, friends: I have a scenario re: “interim analyses” that I’d like to talk through…
(this thread was motivated by a real example, and I’m planning a thread about it eventually, but trying to straighten out my own thoughts first)
(not a trial that I’m working on, but one that I’ve read, and something that made me think / want to figure it out in case this ever happens)
(if you figure out what the example is, kudos to you, but please let it be for the purposes of this thread)
Suppose that you designed a trial to recruit 1,000 patients (and furthermore, suppose all of the assumptions about MCID, power, alpha, recruitment rate, etc seemed totally reasonable at the outset – just humor me).
This trial has a primary outcome that is known almost immediately for each patient (within a few days of treatment), simple yes/no event.
Suppose that you had three interim looks planned: 250 patients, 500 patients, and 750 patients.
Suppose that you used a group-sequential design with O’Brien-Fleming stopping rules; alpha-spending function suggests that you use alpha=0.00005 at the first look, 0.0039 at the second look, 0.0184 at the third look, and 0.0412 at the final analysis.
Suppose that recruitment is going slowly; much slower than planned. At the first interim look, the stopping rule has not been met, but you are now more than halfway through the planned project period (calendar time).
It is unlikely that you will recruit to the target 1,000 within the planned project period; you’ll be lucky to make it to 500.
It is decided, collectively, that the trial will recruit to the end of the planned project period (by calendar time, not enrollment) and analyze whatever data are available at that time, expected to be about 500 patients.
The trial will terminate at approximately the enrollment target that would have been the second interim analysis (but, the trial is ending early principally due to *logistical* reasons, and the decision is *not* based on knowledge of the study outcome data…)
What is the appropriate alpha level for the ‘final’ analysis?
@JasonConnorPhD @RogerJLewis @MarionKCampbell @jd_wilko @BenVanCalster @statsepi to get us started (pls feel free to tag any other DSMB / trial / stats experts)
