Factorial clinical trial designs can be used to test for the efficacy of combination drugs with two or more components, where
inference on the question if a combination therapy is more efficacious
than both of its components is based on the min-test proposed by Laska and Meisner (1989). This is
due to regulatoric demands requiring a contribution of all
compounds in a combination drug. The AVE- and MAX-approaches
proposed by Hung, Chi and Lipicky (1993) test for the existence of any desirable
combination.
Bootstrap-based methods are implemented as well as classical approaches available from
literature to obtain p-values and confidence intervals in
such designs. For the min-test, analytical methods use a normality and homoscedasticity
assumption on the data (Hung, Chi and Lipicky, 1993 and Hung, 2000). Critical
values needed for determination of confidence intervals are calculated using quantiles
of the multivariate t-distribution (Bretz, Genz and Hothorn
2001). These methods fail when handling with data that are skewed or
heteroscedastic over the treatment groups. Furthermore, no analytical
approach is available for the trifactorial case and the AVE- and
MAX-tests on binary data.
In the bootstrap approach, only the empirical distribution of the data is
used and thus the results are valid for any distributional shape,
provided that sufficiently large samples are available. Less
analytical framework is needed to handle with the distributional
properties of the tests. Further information on resampling-based
methods and theoretical backgrounds are given in Westfall and Young
(1993).
Anyway, the problem of the extremely decreasing power
for small values of the so-called nuisance parameters indicating
the response differences between the marginal treatment groups cannot
be resolved by the bootstrap approach. Any
algorithm based on estimates for the nuisance parameters other than
the assumption that they are infinite will exceed the given significance level (Snapinn, 1987).
The package contains the generic functions mintest and margint to test for mean differences of given numeric data
vectors and differences in event rates for binary data
applications. Method dispatch is available for objects of class
carpet or cube, which will lead to min-test
results on a bi- or trifactorial design and corresponding confidence
intervals comparing combination treatments with their respective
component therapies. Implementations for global tests are also
available by the generic functions avetest and maxtest.