qle (version 0.18)

prefitCV: Covariance parameter estimation for cross-validation

Description

The function constructs a list of covariance models of statistics in order to estimate the prediction error variances by a cross-validation approach at unsampled points.

Usage

prefitCV(qsd, reduce = TRUE, type = c("cv", "max"), control = list(),
  cl = NULL, verbose = FALSE)

Arguments

qsd

object of class QLmodel

reduce

if TRUE (default), reduce the number of covariance models to refit

type

type of prediction variances, "cv" (default) and "max", see crossValTx

control

control arguments for REML estimation, passed to nloptr

cl

cluster object, NULL (default), of class MPIcluster, SOCKcluster, cluster

verbose

if TRUE, print intermediate output

Value

A list of certain length depending on the current sample size (number of evaluated points). Each list element corresponds to a (possibly reduced) number of sample points with at most $k$ points (see details) left-out for fitting the corresponding covariance models.

Details

Using the cross-validation approach (see vignette) for estimating the prediction variances might require a refit of covariance parameters of each statistic based on the remaining sample points. The covariance models are refitted if `fit` equals TRUE and otherwise simply updated without fitting which saves some computational resources. The number of points left-out, if applicable, is dynamically adjusted depending on the number of sample points in order to prevent the main estimation algorithm to fit as many models as there are points already evaluated.

The number $n_c$ of covariance models still to fit, that is, the number of partitioning sets of sample points, is limited by $n_c\le n$, with maximum $k$ sampling points deleted from the full sample set with overall $n$ sample points such that $n=n_{c}k$ (see also the vignette for further details).

See Also

QLmodel

Examples

# NOT RUN {
  data(normal)
  
  # without re-estimation of covariance parameters, default is TRUE
  qsd$cv.fit <- FALSE  
  cvm <- prefitCV(qsd)
  
# }