Information Design for Early‐Stage Dose‐Finding Trials

ABSTRACT To enhance enrollment rates in early‐stage dose‐finding clinical trials, we propose an information design approach, where the clinical investigator (CI) commits to an information releasing mechanism (IRM) based on the treatment's uncertain efficacy and toxicity to encourage patients to participate in the trial. The optimal IRM has a threshold structure, involving completely revealing/pooling the state (utility) when the realization is in the respective region. We analyze the comparative statics under the optimal IRM. Because general IRMs may be difficult to implement in practice, motivated by response‐adaptive clinical trials where information from past patients will help drive decisions for future patients, we consider a practical IRM, where the CI decides on the number of patients to recruit, whose efficacy/toxicity outcomes will be used as information for future patients. For example, the proposed IRM may be the number of patients in the first batch in dose‐escalation methods in Phase I, the expansion cohort at the end of Phase I, or the number of patients assigned to the maximum tolerated dose in Phase II. We show that this practical IRM can achieve at most 50% of the value of the optimal abstract IRM. Since patients' risk attitude toward toxicity may be private, we study the impact of belief about it on the optimal IRM and show that the structure of the optimal IRM can be in sharp contrast with that in the public information setting, depending on the distribution of risk attitude toward toxicity. To better understand the impact of the association between efficacy and toxicity on the optimal IRM, we study a bivariate Bernoulli model and show that the optimal IRM has a threshold structure, and the region in which it is a randomized recommendation, when the treatment is ineffective and toxic, shrinks when the association becomes larger.