The calculation method - use "bootstrap" for a
resampling-based approach, "hung" for the min-test approach
of Hung (2000) and "tdistr" for interval calculations based
on the multivariate t-distribution.
test
Either "ttest" or "ztest" - the test
statistic for the inferences to be based on. Use "ztest" for
binary data applications.
alpha
Simultaneous level of the confidence intervals.
nboot
Number of resampling iterations to use.
simerror
Prespecified simulation standard error.
...
Any further arguments.
Details
The generic functions mintest and margint calculate
adjusted p-values and simultaneous confidence intervals for the test
of parametric differences between prespecified treatment groups on bi-
or trifactorial design clinical trials. If an object of class
carpet is commited, mintest will return adjusted
p-values for the min-test on combination superiority in bifactorial
clinical trial designs (Laska and Meisner, 1989). The alternative
hypothesis of this test is that the detected effect size for the
combination treatment is better than for both single component groups;
i.e. the test results in only one p-value for each combination. The
generic function margint will, when applied to carpet
objects, return simultaneous confidence intervals for the parametric
differences between each combination treatment group and its
respective components. Depending on the type of data, the calculations
can be based on Student's t-test for metric data or the Z-statistic
for binary applications.
By default, the calculations are performed by a resampling-based
approach. The desired simulation accuracy always needs to be specified
by the number nboot of bootstrap iterations to perform or an
upper bound simerror for the simulation standard error. If both
are given, the two constraints will be held simultaneously. On the
other hand, the multivariate normal approach for unbalanced designs
from Hung (2000) is available when the argument method is set
to the value "hung". For the trifactorial case, no such
approach is available and thus the calculations are based on the
bootstrap approach, performing a generalized min-test on the data, if
an object of class cube is commited. The interval calculations
are based on the multivariate t-distribution if "tdistr" is
specified.
In the classical approach to the min-test, a normality assumption for
the data is used and the desired critical values are calculated using
quantiles of the multivariate t-distribution. However, this method
fails when handling with data that are skewed or heteroscedastic over
the treatment groups. When using the bootstrap, only the empirical
distribution of the data is used and thus the results are always
valid, provided that a sufficiently large samples are available. When
handling with data from bifactorial clinical trial designs, bootstrap
methods need much less analytical framework on the distributional
properties of the tests than if the approach given by Hung (2000) is
used. In particular, the restriction to only two compounds is not
needed and binary data applications can be handled analogously. The
theory of resampling-based multiple testing has been extensively
discussed by Westfall and Young (1993).
The calculation of simultaneous confidence intervals is much easier
because the c.d.f. of the min-statistic is not needed. Hence this is
leading to an ordinary multiple contrast problem.
Value
An object of class mintest or margint with the following slots.
p
Adjusted p-values for the respective combination groups.
stat
The observed values of the min-statistics.
kiu
The lower limits of the confidence intervals.
kio
The upper limits of the confidence intervals.
alpha
One minus the nominal coverage probability of the confidence intervals.
gnames
Names of the combination groups.
cnames
The names of the contrasts for comparisons of the
combinations with their respective components.
test
Type of test statistics that the min-tests were based on.
method
The method used for calculation.
nboot
Number of bootstrap replications used.
simerror
Maximum of the simulation standard errors in the combination groups.
duration
Total computing duration in seconds.
call
Function call.
Note
Performance of the implemented methods has been evaluated and compared. The min-test performs very
conservative if the means in the marginal treatment groups are close
for a combination.
Frommolt P, Hellmich M (2009): Resampling in multiple-dose factorial
designs. Biometrical J 51(6), pp. 915-31
Hung HMJ, Chi GYH, Lipicky RJ (1993): Testing for the existence of a
desirable dose combination. Biometrics 49, pp. 85-94
Hung HMJ, Wang SJ (1997): Large-sample tests for binary outcomes in
fixed-dose combination drug studies. Biometrics 53, pp. 498-503
Hung HMJ (2000): Evaluation of a combination drug with multiple doses in
unbalanced factorial design clinical trials. Statist Med 19, pp. 2079-2087
Hellmich M, Lehmacher W (2005): Closure procedures for monotone bi-factorial
dose-response designs. Biometrics 61, pp. 269-276
Laska EM, Meisner MJ (1989): Testing whether an identified treatment is
best. Biometrics 45, pp. 1139-1151
Snapinn SM (1987): Evaluating the efficacy of a combination
therapy. Statist Med 6, pp. 657-665
Westfall PH, Young SS (1993): Resampling-based multiple testing. John
Wiley & Sons, Inc., New York
See Also
bifactorial, carpet, cube, avetest, maxtest,
Examples
#AML example from Huang et al. (2007) with data from
#Issa et al. (2004) and Petersdorf et al. (2007)
n<-c(10,31,17,100,50,50,101,50,50)
p<-c(0.00,0.45,0.65,0.30,0.71,0.70,0.59,0.64,0.75)
y<-list()
for(i in 1:9){
y[[i]]<-0
while((sum(y[[i]])!=round(n[i]*p[i]))||(length(y[[i]])==1)){
y[[i]]<-rbinom(n[i],1,p[i])
}
}
aml<-carpet(data=y,D=c(2,2))
mintest(aml,test="ztest",nboot=25000)
margint(aml,test="ztest",nboot=25000)