RandomFields (version 3.1.12)

RFratiotest: Likelihood ratio test

Description

The function performs an approximate $$chi^2$ test or a Monte Carlo likelihood ratio test based on fitgauss. Currently it only works for Gaussian random fields.

Usage

RFratiotest(nullmodel, alternative, x, y = NULL, z = NULL, T = NULL,
grid=NULL, data,
alpha, n = 5 / alpha, seed = 0, 
lower = NULL, upper = NULL, methods,
sub.methods, optim.control = NULL, users.guess = NULL,
distances = NULL, dim, transform = NULL, ...)

Arguments

nullmodel, alternative

See Details. The set of parameters to be estimated for nullmodel should be a subset of the parameters to be estimated for alternative if alternative is given.

alpha

value in [0,1] or missing. Significance level.

n

integer. The test is based on n-1 simulations.

seed

integer. If not NULL and not NA, the .Random.seed is set to seed. Otherwise, set.seed is set to the value

x, y, z, T, grid, data, lower, upper, methods, sub.methods, optim.control, users.guess, distances, dim, transform, ...

see RFfit

Value

• The test returns a message whether the null hypothesis, i.e. the smaller model is accepted. Invisibly, a list that also contains
  • p, the$p$-value
  • n
  • data.ratiothe log ratio for the data
  • simu.ratiothe log ratio for the simulations
  • data.fitthe models fitted to the data
  • msgthe message that is also directly returned
  It has S3 class "RFratiotest".

Details

nullmodel (and the alternative) can be

• a covariance model, seeRMmodelor typeRFgetModelNames(type="variogram")to get all options.

  Depending weather theRFoptionsratiotest_approxis TRUE the the chisq approximation is performed. Otherwise a Monte Carlo ratio test is performed.

• RFfitorRMmodelFitHere, a chisq approximative test is always performed on the already fitted models.

RFratiotest tries to detect whether nullmodel is a submodel of alternative. If it fails,

• a message is printed that says that anautomaticdetection has not been possible;
• it is not guaranteed anymore that thealternativemodel returns a (log) likelihood that is at least as large as that of thenullmodel, even ifnullmodelis a submodel ofalternative. This is due to numerical optimisation which is never perfect.

Otherwise it is guaranteed that the alternative model has a (log) likelihood that is at least as large as that of the nullmodel.

See Also

RFfit RMmodel, RandomFields, weather.

Examples

\dontrun{
RFoptions(seed=0) ## *ANY* simulation will have the random seed 0; set
##                   RFoptions(seed=NA) to make them all random again

data(soil)  ## see also ?soil
soil <- RFspatialPointsDataFrame(
 coords = soil[, c("x.coord", "y.coord")],
 data = soil[, c("moisture", "NO3.N", "Total.N",
 "NH4.N", "DOC", "N20N")],
 RFparams=list(vdim=6, n=1)
 )

model <- ~1 + RMplus(RMwhittle(scale=NA, var=NA, nu=NA), RMnugget(var=NA))
submodel <- ~1 + RMplus(RMwhittle(scale=NA, var=NA, nu=NA), RMnugget(var=0))

RFratiotest(submodel, model, data=soil["moisture"],
            modus_operandi="sloppy")
}

FinalizeExample()