Interval estimation for dose-finding studies

Projektbeginn: 2016
Forschungsteam: Saswati Saha, Georg Gutjahr, Werner Brannath, Luc Bijnens (Janssen-BE)
Förderung: EU
Klinischer Berater:Dr. John Warren


This project is part of EU project "Improving Design, Evaluation and Analysis of early drug development Studies" (IDEAS; MARIE SKODOWSKA-CURIE ACTIONS, H2020-MSCA-ITN-2014).

Finding the correct dose is crucial during the early phases of most drug developments. Usually, a maximum tolerated dose is estimated during a phase I study. Hereafter, phase II studies are performed, where statistical considerations are used for the selection of a dose used in a subsequent confirmatory phase III study. Statistical objectives in such a phase II study are the estimation of the minimal effective dose (the smallest dose level whose expected response is larger than the placebo response by some pre-specified margin) or estimation of the median effective dose (ED50).

Two approaches exist for the analysis of phase II dose-response studies. First, in a modelling approach, a specific form of the relationship between the dose level and the mean response is assumed. The data from the study are then used to estimate the parameters of the dose-response shape within a certain class of parametric nonlinear models. Techniques such as modelling averaging, that combine predictions from multiple models, are frequently used to avoid the dependency on modelling assumptions from a single model. Second, in a multiple-comparison procedure, pairwise comparisons between the dose levels are performed. The estimation of the quantities of interest, such as the minimal effective dose, is then based on pairwise comparisons with statistical significance testing. Depending on the size of the study, multiplicity adjustment may be applied to the comparisons. Such adjustments either increase the critical levels of the comparisons or they constrain the ordering of the comparisons and requiring that certain comparisons must give significant results before considering comparisons of lower ordering (for example, performing sequential t-test of decreasing dose levels against a placebo zero dose). An interesting combination of these two adjustment strategies is to allow the ordering of the comparisons to be partially data-dependent. Contingent on a parameter that governs the degree to which the ordering can depend on the data, this strategy includes the sequential t-test as well as the Dunnett test as special cases.

While the calculation of confidence intervals is mandatory in many areas of clinical research, confidence intervals for dose-finding studies have not yet been thoroughly investigated. The goal of this project will be the development and analysis of confidence intervals for dose-finding studies based modelling techniques on the one hand, and on multiple comparison procedures on the other hand.

Link zu