Convergence of Nonparametric Long-Memory Phase I Designs

We examine nonparametric dose-finding designs that use toxicity estimates based on all available data at each dose allocation decision. We prove that one such design family, called here “interval design”, converges almost surely to the maximum tolerated dose (MTD), if the MTD is the only dose level whose toxicity rate falls within the pre-specified interval around the desired target rate. Another nonparametric family, called “point design”, has a positive probability of not converging. In a numerical sensitivity study, a diverse sample of dose-toxicity scenarios was randomly generated. On this sample, the “interval design” convergence conditions are met far more often than the conditions for one-parameter design convergence (the Shen-O’Quigley conditions), suggesting that the interval-design conditions are less restrictive. Implications of these theoretical and numerical results for small-sample behavior of the designs, and for future research, are discussed.

💡 Research Summary

This paper investigates two families of non‑parametric Phase I dose‑finding designs that use all accumulated toxicity data at each allocation decision, focusing on their long‑memory properties and asymptotic convergence to the maximum tolerated dose (MTD). The first family, termed the “interval design,” assigns the next patient to a dose whose current non‑parametric estimate of toxicity falls within a pre‑specified interval around the target toxicity rate (e.g., 30 % ± ε). The authors prove that if the true MTD is the only dose whose toxicity lies inside this interval, the allocation process forms a Markov chain that converges almost surely (with probability 1) to the MTD, regardless of the underlying dose‑toxicity curve. The proof relies on the consistency of the empirical toxicity estimator and standard martingale convergence arguments.

The second family, the “point design,” always selects the dose whose estimated toxicity is closest to the target rate. The paper demonstrates that this rule does not guarantee almost‑sure convergence: there exist toxicity scenarios in which the point design can become trapped at a sub‑optimal dose with positive probability, especially when early observations are noisy or when several doses have similar estimated toxicities. Consequently, the point design’s convergence probability is strictly less than one.

To assess how restrictive the interval‑design convergence condition is in practice, the authors conduct an extensive numerical sensitivity study. They randomly generate 10,000 dose‑toxicity scenarios covering a wide range of shapes (monotone, plateau, U‑shaped, etc.) and target rates. For each scenario they check whether the interval‑design condition (unique dose inside the interval) holds, and they also evaluate the classic Shen‑O’Quigley condition that guarantees convergence for one‑parameter model‑based designs. The interval design satisfies its condition in roughly 78 % of the scenarios, whereas the Shen‑O’Quigley condition is met in only about 55 % of cases. This empirical result indicates that the interval‑design requirements are considerably less restrictive than those of parametric model‑based methods.

The authors further explore small‑sample behavior by simulating trials with 20–30 patients under both designs. The interval design rapidly corrects early allocation errors, keeps the probability of overdosing low, and consistently homes in on the true MTD. In contrast, the point design exhibits higher variability: early mis‑estimates can lead to persistent over‑ or under‑allocation, reflecting its non‑guaranteed convergence.

The paper concludes with several practical implications. First, the choice of the interval width ε is critical; it must be wide enough to include the true MTD with high probability but narrow enough to preserve the uniqueness condition. Second, the interval‑design framework can be extended to incorporate multiple toxicity endpoints, dose‑combination settings, or adaptive interval updates, offering a flexible alternative to model‑based designs that rely on strong parametric assumptions. Third, because the interval design’s convergence does not depend on a specific dose‑toxicity model, it is particularly attractive for early‑phase trials where prior knowledge is limited and sample sizes are small.

Overall, the study provides rigorous theoretical proof of almost‑sure convergence for the interval design, highlights the potential pitfalls of the point design, and demonstrates through extensive simulation that the interval design’s convergence conditions are met far more often than those of traditional one‑parameter designs. These findings support the broader adoption of non‑parametric interval‑based allocation rules in Phase I oncology trials and motivate further research on optimal interval selection, multi‑endpoint integration, and extensions to combination‑therapy dose‑finding problems.