Accounting for parameter uncertainty in two-stage designs for Phase II dose-response studies

TL;DR

This paper compares two adaptive dose-response study designs, one using maximum likelihood estimates and the other incorporating Bayesian parameter uncertainty, to improve phase II trial planning.

Contribution

It introduces a computationally efficient Bayesian design method using k-means clustering for dose-response studies, accounting for parameter uncertainty.

Findings

Bayesian approach better at early interim analysis

Proposed k-means clustering method improves computational efficiency

Bayesian design depends on prior specification

Abstract

In this paper we consider two-stage adaptive dose-response study designs, where the study design is changed at an interim analysis based on the information collected so far. In a simulation study, two approaches will be compared for these type of designs; (i) updating the study design by calculating the maximum likelihood estimate for the dose-response model parameters and then calculating the design for the second stage that is locally optimal for this estimate, and (ii) using the complete posterior distribution of the model parameter at interim to calculate a Bayesian optimal design (i.e. taking into account parameter uncertainty). In particular, for an early interim analysis respecting parameter uncertainty seems more adequate, on the other hand for a Bayesian approach dependency on the prior is expected and an adequately thought-through prior is required. A computationally efficient method is proposed for calculating the Bayesian design at interim based on approximating the full posterior sample using k-means clustering. The sigmoid Emax dose-response model and the D-optimality criterion will be used in this paper.