Adaptive designs are easy to propose and hard to defend. Sponsors hear "fewer patients, faster decisions," and reviewers hear "more ways to inflate type-I error." Both groups are correct, which is why every adaptive protocol we write at ACRO is, in some sense, an exercise in pre-empting an interview neither side has had yet.
This is the story of one such trial — a Phase II, single-agent dose-finding study in a solid tumor indication, conducted across six Korean academic sites between 2023 and 2025. The sponsor came to us with a traditional 3+3 design and a timeline that didn't quite work. We came back with a Bayesian Logistic Regression Model (BLRM) and a set of decisions we now look back on as roughly 70% correct.
The trade-off we set out to quantify
The headline claim for adaptive dose-finding is operational: smaller expected sample size, shorter trial. The actual claim, when you write the operating characteristics carefully, is more nuanced — better dose selection probability at a given sample size, with the option (not the obligation) to stop earlier when the data permit it.
For this program, the meaningful question wasn't "how few patients can we use?" It was "how confident can we be in the recommended Phase 2 dose (RP2D), given the operational constraints we'll actually face?"
The protocols that survive review aren't the ones with the cleverest design. They're the ones where the cleverness is in service of a decision the reviewer can recognize.
What the BLRM bought us
The BLRM with escalation-with-overdose-control (EWOC) gave us three things that mattered:
- A continuous posterior estimate of DLT probability at each dose, rather than the 3+3's discrete cohort logic.
- A principled mechanism for skipping dose levels when the data justified it — important when the predicted therapeutic window was narrow.
- An interim decision framework the DSMB could review without needing to relitigate the original design assumptions.
Where the design earned its complexity
The interim look at 18 patients was the moment everything paid off — or would have, if we'd designed it slightly better. The DSMB had clear stopping rules: futility if the posterior probability of efficacy at the candidate dose fell below 15%, escalation if the safety boundary was clear. The data fell in between. The committee asked, reasonably, what we'd do if this exact data pattern repeated.
That question wasn't in our pre-specified decision matrix. We added it.
Pre-specifying adaptive rules isn't a one-time exercise. The DSMB's first interim review is where you learn which decision boundaries you forgot to spell out.
What we'd do differently
- Wider posterior intervals at the interim. Our 90% credible intervals were narrower than the actual sampling variability suggested. Next time: 95%, with explicit sensitivity to prior choice.
- Pre-registered DSMB question library. We treated the DSMB as a single decision point. They aren't — they're a series of conversations, and the protocol should anticipate that.
- Earlier biostatistics-DSMB alignment. Two of our most useful interim slides were drafted the night before. They should have been built into the design package.
The operational headline
The trial closed enrollment six months ahead of the original projection, with an RP2D supported by 90% posterior probability of an acceptable safety-efficacy profile. The manuscript is in revision at a major oncology journal. The sponsor's Phase III protocol — currently in development with our team — inherits the dose, the population definition, and the safety run-in design.
The longer version of this story is in our forthcoming methods paper. But the short version is this: adaptive designs work when you treat them as a discipline, not a feature. Every shortcut is a future conversation with a reviewer; every pre-specification is a conversation you've already had.