Holger Kraiczi (1), Marianne Frisén (2)
(1) Pfizer Consumer Healthcare, Lund, Sweden, and Department of Clinical Pharmacology, Sahlgrenska University Hospital, Gothenburg, Sweden; (2) Statistical Research Unit, Gothenburg University, Gothenburg, Sweden
In clinical trial simulation (CTS), uncertainty about population parameters determining the shape of the simulation model will influence the degree of uncertainty about the variable predicted with the simulation, e.g., trial power. However, the impact of uncertainty about single model parameters on the CTS prediction may vary considerably. In this sense, ‘important’ parameters, required to be entered into simulations with a high degree of precision to allow reliable CTS outcomes can be distinguished from ‘unimportant’ parameters, uncertainty about which has no serious consequences for the predicted variable. If such discrimination is done prior to trial simulations, research resources expended to inform CTS can be focused on the ‘important’ model parameters.
Using a worked example, we illustrate how uncertainty about population parameters may be incorporated into CTS by simulating full Bayesian predictive distributions of CTS outcome variables. We then suggest a method of sensitivity analysis, based on 2^k factorial simulation experiments, to rank input parameters with respect to their influence on uncertainty about the predicted CTS variable.
The analysis used to exemplify our approach is applied to the simulation of a completely randomized, placebo-controlled parallel-groups efficacy trial, in which the effect of the study drug is measured as a continuous outcome variable during steady-state conditions. The structural trial model links dose, concentration, effect (assuming a sigmoidal Emax model) and trial power. Trial power is the predicted variable and defined as the probability of an observed treatment effect that significantly exceeds a pre-specified clinically worthwhile difference. Analyses were performed for 3 different doses and 4 different settings of hyperparameters for 10 population parameters (means and variances for each of 5 structural model parameters).
The example illustrates that Bayesian predictive simulation, combined with a suitable experimental design of simulation studies, may be used to estimate the amount of information required to enable reliable predictions in CTS and, thus, to guide the process of learning in early drug development.
Reference: PAGE 13 () Abstr 472 [www.page-meeting.org/?abstract=472]
Poster: poster