Abstract
Background: Coherence notions have a long history in statistics, as rhetorical devices that support the critical examination of statistical doctrines and practices. Within the special domain of dose-finding methodology, a widely-discussed coherence criterion has been advanced as a means to guard the conceptual integrity of formal dose-finding designs from ad hoc tinkering. This is not, however, the only possible coherence criterion relevant to dose finding. Indeed, a new coherence criterion emerges naturally when the near-universal practice of cohort-wise dose escalation is examined from a clinical perspective. Methods: The practice of enrolling drug-naive patients into an escalation cohort is considered from a realistic perspective that acknowledges patients' heterogeneity with respect to pharmacokinetics and pharmacodynamics. A new coherence criterion thereby emerges, requiring that an escalation dose be tried preferentially in participants who have already tolerated a lower dose, rather than in new enrollees who are drug-naive. The logical implications of this 'precautionary coherence' (PC) criterion are worked out in the setting of a 3+3 design. A '3+3/PC' design that satisfies this criterion is described and visualized. A simulation study is performed, evaluating the long-run performance of this new design, relative to optimal 1-size-fits-all dosing. Results: Under the PC criterion, the 3+3 dose-escalation design necessarily transmutes into a dose titration design. Two simple rules suffice to enable abandonment of low starting doses, and termination of escalation. The process of conducting the 3+3/PC trial itself models the application of a dose titration algorithm (DTA) that carries over readily into clinical care. The 3+3/PC trial also yields an interval-censored 'dose-survival curve' having a semantics that should prove familiar to oncology trialists. Simulated 3+3/PC trials yield DTAs over a median of 6 dose levels, achieving 50% improved population-level efficacy compared to optimal 1-size-fits-all dosing. Conclusions: Dose individualization can be accomplished within a trial conducted along 'algorithmic' lines resembling those of the inveterate 3+3 design. The dose-survival curve arising from this '3+3/PC' design has semantics that should prove familiar and conceptually accessible to oncology trialists, and also seems capable of supporting more formal statistical treatments of the design. In the presence of sufficient heterogeneity in individualized optimal dosing, a 3+3/PC trial outperforms any conceivable 1-size-fits-all dose-finding design. This fact eliminates the rationale for the latter designs, and should put an end to the further development and promulgation of 1-size-fits-all dose finding.
