compositions (version 1.10-1)

gof: Compositional Goodness of fit test

Description

Goodness of fit tests for compositional data.

Usage

acompGOF.test(x,...)
acompDirichletGOF.test(x,samplesize=nrow(x)*20,R=999)
acompNormalGOF.test(x,...,method="etest")
## S3 method for class 'formula':
acompGOF.test(formula, data,...,method="etest")
## S3 method for class 'list':
acompGOF.test(x,...,method="etest")
gsi.acompUniformityGOF.test(x,samplesize=nrow(x)*20,R=999)
acompTwoSampleGOF.test(x,y,...,method="etest",data=NULL)

Arguments

x

a dataset of compositions (acomp)

y

a dataset of compositions (acomp)

samplesize

number of observations in a reference sample specifying the distribution to compare with. Typically substantially larger than the sample under investigation

R

The number of replicates to compute the distribution of the test statistic

method

Selecting a method to be used. Currently only "etest" for using an energy test is supported.

...

further arguments to the methods

formula

an anova model formula defining groups in the dataset

data

unused

Value

• A classical "htest" object
• data.name
  The name of the dataset as specified
• method
  a name for the test used
• alternative
  an empty string
• replicates
  a dataset of p-value distributions under the Null-Hypothesis got from nonparametric bootstrap
• p.value
  The p.value computed for this test

Missing Policy

Up to now the tests can not handle missings.

Details

The compositional goodness of fit testing problem is essentially a multivariate goodness of fit test. However there is a lack of standardized multivariate goodness of fit tests in R. Some can be found in the energy-package.

In principle there is only one test behind the Goodness of fit tests provided here, a two sample test with test statistic. $$\frac{\sum_{ij} k(x_i,y_i)}{\sqrt{\sum_{ij} k(x_i,x_i)\sum_{ij} k(y_i,y_i)}}$$ The idea behind that statistic is to measure the cos of an angle between the distributions in a scalar product given by $$(X,Y)=E[k(X,Y)]=E[\int K(x-X)K(x-Y) dx]$$ where k and K are Gaussian kernels with different spread. The bandwith is actually the standarddeviation of k. The other goodness of fit tests against a specific distribution are based on estimating the parameters of the distribution, simulating a large dataset of that distribution and apply the two sample goodness of fit test.

References

Aitchison, J. (1986) The Statistical Analysis of Compositional Data Monographs on Statistics and Applied Probability. Chapman & Hall Ltd., London (UK). 416p.

See Also

fitDirichlet,rDirichlet, runif.acomp, rnorm.acomp,

Examples

x <- runif.acomp(100,4)
y <- runif.acomp(100,4)

erg <- acompGOF.test(x,y)
#continue
erg
unclass(erg)
erg <- acompGOF.test(x,y)


x <- runif.acomp(100,4)
y <- runif.acomp(100,4)
dd <- replicate(1000,acompGOF.test(runif.acomp(100,4),runif.acomp(100,4))$p.value)
hist(dd)

dd <- replicate(1000,acompGOF.test(runif.acomp(20,4),runif.acomp(100,4))$p.value)
hist(dd)
dd <- replicate(1000,acompGOF.test(runif.acomp(10,4),runif.acomp(100,4))$p.value)

hist(dd)
dd <- replicate(1000,acompGOF.test(runif.acomp(10,4),runif.acomp(400,4))$p.value)
hist(dd)
dd <- replicate(1000,acompGOF.test(runif.acomp(400,4),runif.acomp(10,4),bandwidth=4)$p.value)
hist(dd)


dd <- replicate(1000,acompGOF.test(runif.acomp(20,4),runif.acomp(100,4)+acomp(c(1,2,3,1)))$p.value)

hist(dd)


x <- runif.acomp(100,4)
acompUniformityGOF.test(x)

dd <- replicate(1000,acompUniformityGOF.test(runif.acomp(10,4))$p.value)

hist(dd)