ABSTRACT
Phase I clinical trials are almost always conducted at a single investigational site. Many Phase II clinical trials are conducted at two or more investigational sites or centers. Virtually all Phase III clinical trials have to be conducted at several investigational sites in order to accrue the number of patients required by the protocol. The analyses of multicenter clinical trials pose challenges in terms of what
are the most appropriate analyses of the data collected. Some analysts believe that the proper analysis of a multicenter clinical trial is based on an analysis of variance Model (1) with fixed effects of treatment and center, and then follow this analysis with the Model (2) that contains fixed effects of treatment, center, and center-by-treatment interaction to assess the significance of the interaction term. If the interaction term is not statistically significant, then treatment effects from Model (1) are reported and considered generalizable across centers. If the interaction term is significant, some analysts report the treatment effect from Model (1) with the caveat that the treatment effect is inconsistent across centers. Other analysts aver that the proper analysis derives from Model (2), and if
there is a statistically significant treatment-by-center interaction, there should be no overall pooled analysis of the data. Further, if the treatment-by-center interaction is statistically significant, some analysts test treatment effect from Model (2) and base an inference on the difference in treatment group means adjusted for differences between centers and (treatment-by-center interaction). To get around the interaction problem, other analysts suggest that the
proper analysis derives from an analysis of variance Model (3) with fixed treatment effects and random center and treatment-by-center effects, and base the test for treatment effect on the ratio of mean square for treatment divided by the mean square for treatment-by-center interaction. The main criticism of this latter strategy is that centers or investigators do not represent a random
sample from some population of centers or investigators. Further, the power of the test for treatment effect from this strategy would be reduced (as compared to Model (1) or Model (2)), particularly in multicenter clinical trials that have few centers. The main focus of this chapter is to discuss design-based and model-based
approaches in the analysis of multicenter trials and to argue that a pooled analysis should be performed for every quality designed and conducted multicenter clinical trial.
