MChtest (version 1.0-3)

MCtest: Perform Monte Carlo hypothesis tests.

Description

Performs Monte Carlo hypothesis test with either a fixed number of resamples or a sequential stopping boundary on the number of resamples. Outputs p-value and confidence interval for p-value. The program is very general and different bootstrap or permutation tests may be done by defining the statistic function and the resample function.

Usage

MCtest(x, statistic, resample, bound, extreme = "geq", seed = 1234325)
MCtest.fixed(x, statistic, resample, Nmax, extreme = "geq", 
    conf.level=.99, seed = 1234325)

Arguments

x

data object

statistic

function that inputs data object, x, and outputs a scalar numeric

resample

function that inputs data object,x, and outputs a "resampling" of x, an object of the same type as x

Nmax

the number of resamples for the fixed boundary

bound

a object of class MCbound, can be created by MCbound see that help

extreme

character value either "geq" or "leq". Defines which Monte Carlo outputs from statistic (T1,T2,...) are denoted extreme with respect to the original output (T0) ; extreme values are either all Ti >= T0 ("geq") or all Ti <= T0 ("leq").

conf.level

confidence level for interval about p-value

seed

a numeric value used in set.seed

Value

A LIST of class "MCtest",with elements:

Ti

an N vector of the outputs of the resampled statistic

type

type of boundary: "fixed", "tsprt", "Bvalue" or "BC"

parms

vector of parameters that define boundary specified by type

T0

the value of the test statistic applied to the original data

p.value

p.value

pvalue.ci

confidence interval about the p-value

Details

Performs Monte Carlo hypothesis test. MCtest allows any types of Monte Carlo boundary created by MCbound, while MCtest.fixed only performs Monte Carlo tests using fixed boundaries. The only advantage of MCtest.fixed is that one can do the test without first creating the fixed stopping boundary through MCbound. The default boundary is described in MCbound.precalc1.

Let T0=statistic(x) and let T1,T2,... be statistic(resample(x)). Then in the simplest type of boundary, the "fixed" type, then the resulting p-value is p=(S+1)/(N+1), where S=(\# Ti >= T0) (if extreme is "geq") or S=(\# Ti <= T0) (if extreme is "leq"). The confidence interval on the p-value is calculated by an exact method.

There are several different types of MC designs that may be used for the MCtest. These are described in the MCbound help.

References

Fay, M.P., Kim, H-J. and Hachey, M. (2007). Using truncated sequential probability ratio test boundaries for Monte Carlo implementation of hypothesis tests. Journal of Computational and Graphical Statistics. 16(4):946-967.

See Also

MCbound,MCbound.precalc1

Examples

# NOT RUN {
x<-data.frame(y=1:100,z=rnorm(100),group=c(rep(1,50),rep(2,50)))
stat<-function(x){ cor(x[,1],x[,2]) }
### nonparametric bootstrap test on correlation between y and z
### low p-value means that such a large correlation unlikely due to chance
resamp<-function(x){ n<-dim(x)[[1]] ; x[sample(1:n,replace=TRUE),] }
out<-MCtest(x,stat,resamp,extreme="geq") 
out$p.value
out$p.value.ci
### permutation test, permuting y only within group
resamp<-function(x){ 
    ug<-unique(x[,"group"])
    y<- x[,"y"]
    for (i in 1:length(ug)){
         pick.strata<- x[,"group"]==ug[i]
         y[pick.strata]<-sample(y[pick.strata],replace=FALSE)
    }
    x[,1]<-y
    x 
}

out<-MCtest.fixed(x,stat,resamp,N=199) 
out$p.value
out$p.value.ci
# }