BGPT: Paper Review: A Three-Outcome Design for Phase II Clinical Trials

Explore by Goal

Quick Explanation Copied

Core idea: replace the usual 2-outcome Phase II decision (reject H0 vs reject Ha) with a 3-outcome rule that also allows an inconclusive result when the observed response count falls in an uncertainty region—then extend the construction to 2-stage designs and study operating characteristics and optimal/minimax sample sizes.

Long Explanation

Paper Review (visual, critical, evidence-based): “A Three-Outcome Design for Phase II Clinical Trials”

Paper DOI: 10.1016/s0197-2456(00)00115-x • Type: methodological statistical design paper • Key contribution: formalizes “inconclusive” regions in Phase II decision-making and analyzes operating characteristics + optimal/minimax criteria

Figure 1. Sample-size comparison (Table 1 values)

The paper provides example cutoff/designs and required sample sizes N for both the standard 2-outcome design and the proposed 3-outcome design across several (p0, pa) scenarios (using exact binomial calculations with constraints).

Note on figure construction: Values are taken directly from the paper’s Table 1 sample-size entries as provided in the supplied full text.

Figure 2. Fractional sample-size reduction (from Table 1)

For each scenario above, compute (Nstandard − Nthree)/Nstandard using Table 1 N values.

Design concept (three outcomes)

Standard Phase II (two outcomes): with a binary endpoint, the design uses a cutoff rule on observed successes Y among N—either reject the null H0 (typically implying activity is sufficient) or reject the alternative Ha (implying insufficient activity), with boundary defined by a single cutoff.
Three-outcome design: introduce two cutoff points r and s such that when Y ≤ r one decision is made, when r < Y < s the result is labeled inconclusive (reject neither), and when Y ≥ s the other decision is made.
Error-logic changes: because an inconclusive region exists, operating characteristics are described not only by probabilities of wrong rejection under p0 and pa, but also by probabilities of concluding inconclusive under each hypothesis; the paper explicitly distinguishes these quantities.
Rationale: standard designs can force a “clear” decision even when the data fall near the boundary, producing ambiguity in how clinicians interpret the study; the three-outcome structure is meant to make that ambiguity explicit.

Single-stage theory: what is derived

The paper first gives the standard one-stage binomial/based hypothesis test review, then defines the three-outcome single-stage design and—for analytic tractability—uses the normal approximation to the binomial to derive closed-form formulas for required N and cutoff points r and s (with r and s computed from p0, pa and target error probabilities).

It then emphasizes that exact binomial probabilities are used for actual design comparison and sample-size optimization.

Two-stage extension + optimal/minimax designs

Rule structure: define stage-wise cutoffs so that in Stage 1 the trial can stop only to conclude lack of activity (reject Ha) and continue otherwise; in Stage 2, the same three-outcome logic applies using total cumulative successes Y.
Operating characteristics: it writes expressions for probabilities of rejecting Ha and H0 in terms of binomial distributions for stage-wise successes.
Optimal vs minimax: “optimal” minimizes expected sample size under H0 subject to constraints on alpha/beta and inconclusive-region probabilities; “minimax” minimizes the maximum total sample size across admissible designs, with an “optimal minimax” tie-breaker selecting the design with smaller expected sample size.
Computational note: it states a Fortran program is available on request for finding optimal/minimax designs.

Skeptical critique (what is strong vs what is uncertain)

What’s strong

Explicit handling of inconclusive outcomes with formally defined probabilities under each hypothesis (rather than informal “borderline” talk).
Exact-binomial comparison (not purely relying on asymptotic normal approximations) for the sample-size comparison table.

Key uncertainties / limitations

Pre-specification sensitivity: the approach depends on choosing p0, pa, and target error constraints, and also requires setting inconclusive-region probabilities; if those design inputs are mis-specified relative to reality, operating characteristics may shift.
Normal approximation used in derivations: while the paper compares designs using exact binomial probabilities, the closed-form derivations for N,r,s rely on normal approximations.
Clinically, “inconclusive” still requires interpretation: although the design makes the uncertainty formal, subsequent decision-making is not automatic; it still depends on other factors outside the primary binary success endpoint (toxicity/cost/etc. are explicitly mentioned as possible drivers).
Computational transparency: it indicates Fortran code is available on request rather than providing it publicly in the manuscript, which reduces straightforward reproducibility without contacting the authors.

What would disprove/force modification?

Empirically, if in real trials clinicians still cannot reach consistent downstream decisions from an “inconclusive” region (or if the inconclusive region introduces operational complexity without improving alignment with clinical intuition), the practical value could be challenged.

Author reviews (bespoke BGPT pages)

Feedback:

Updated: April 04, 2026