A trial simulation example to support the design and model-based analysis of a new dose and regimen finding study
D Renard, CH Hsu, M Dolker, V Yu, M Looby
Objectives: Use trial simulation as a general tool to make informed recommendations on the following aspects of a new dose and regimen finding study protocol:
- trial design (Stage I),
- sample size (Stage II),
- analysis methodology (Stage III).
Methods: Trial simulations were performed in three stages. In Stage I, the primary objective was to facilitate trial design decisions regarding the use of a modeling approach instead of a traditional method (ANCOVA) and the choice of some design features (e.g. parallel or crossover, dose levels, study visits). The median absolute deviation from true value for quantities of interest (comparisons of active doses or placebo-corrected responses) was utilized as a measure to compare efficiencies of different design attributes. In Stage II simulations were conducted to determine sample size. In Stage III simulations were conducted to evaluate the proposed model-based analysis approach which was articulated along 4 key principles:
- The dose-response relationship is of Emax type.
- The totality of data is included in the analysis, not just the end-point.
- Several candidate models, deemed a priori reasonable to describe the data, are considered.
- Model averaging  is employed to achieve more robust inference.
Simulations were performed to reflect current knowledge and uncertainty based on available data. Key metrics were bias as well as length and coverage of confidence intervals to compare model-based and ANCOVA methods.
Results: Stage I: Simulations revealed that considerable improvements in precision can be expected from using the model-based approach over the traditional endpoint analysis. The choice of analysis method (model-based vs. ANCOVA) was the most discriminative feature among those investigated. Other design attributes such as parallel groups vs. crossover resulted in net efficiency gains one order of magnitude lower.
Stage II: Not further discussed here.
Stage III: Model averaging revealed good properties, with a favorable trade-off between bias and precision resulting in less variability overall. In particular, significant gains in efficiency remained over ANCOVA with greater benefits seen when comparing active doses. The procedure was found to be slightly conservative when examining coverage probability.
Conclusions: Stage I and II of trial simulations allowed to determine key design features for the new study. The choice of a model-based method over ANCOVA was the primary factor to be considered in order to improve the overall study efficiency. The proposed analysis methodology was shown to be robust and efficient, with a favorable trade-off between bias and precision resulting in less variability overall. Greater benefits should be anticipated in comparisons of active doses rather placebo-corrected responses, which is of particular importance when it comes to contrasting doses at the dose selection stage. The procedure was found to be on the conservative side which should not be regarded negatively, at least within a regulatory context.
 Hoeting JA, Madigan D, Raftery A, et al (1999) Bayesian model averaging: a tutorial. Statist Science; 14: 382-417.