Summary
Background: Binary composite outcome measures are increasingly used as primary endpoints in clinical
trials. Composite endpoints combine several events of interest within a single variable.
However, as the effect observed for the composite does not necessarily reflect the
effects for the individual components, it is recommended in the literature to additionally
evaluate each component separately.
Objectives: The task is to define an adequate multiple test procedure which focuses on the composite
outcome measure but allows for a confirmatory interpretation of the components in
case of large effects.
Methods: In this paper, we determine the correlation matrix for a multiple binary endpoint
problem of a composite endpoint and its components based on the normal approximation
test statistic for rates. Thereby, we assume multinomial distributed components. We
use this correlation to calculate the adjusted local significance levels. We discuss
how to use our approach for a more informative formulation of the test problem. Our
work is illustrated by two clinical trial examples.
Results: By taking into account the special correlation structure between a binary composite
outcome and its components, an adequate multiple test procedure to assess the composite
and its components can be defined based on an approximate multivariate normal distribution
without much loss in power compared to a test problem formulated exclusively for the
composite.
Conclusions: By incorporating the correlation under the null hypotheses, the global power for
the multiple test problem assessing both the composite and its components can be increased
as compared to simple Bonferroni-adjustment. Thus, a confirmatory analysis of the
composite and its components might be possible without a large increase in sample
size as compared to a single endpoint problem formulated exclusively for the composite
Keywords
Binary endpoints - clinical trials - composite endpoints - correlation matrix - multiple
testing
