setQLdata: Set quasi-likelihood (QL) data
In mbaaske/qle: Simulation-Based Quasi-Likelihood Estimation

Description Usage Arguments Details Value Author(s) Examples

Aggregate the data for quasi-likelihood estimation

1 2	setQLdata(runs, X = NULL, var.type = "cholMean", Nb = 0, na.rm = TRUE, verbose = FALSE)

`runs`	list or matrix of simulation results from `simQLdata`
`X`	list or matrix of model parameters
`var.type`	character, "`cholMean`" (default), whether to Cholesky decompose variance matrices either for sample average variance approximation or kriging variance matrices
`Nb`	numeric, number of bootstrap samples for kriging the variance matrix, only if '`var.type`'='`kriging`', default is zero which uses no bootstrapping
`na.rm`	if `TRUE` (default), remove 'NA' values from simulation results
`verbose`	if `TRUE`, print intermediate output

The function aggregates all neccessary data for quasi-likelihood estimation storing the sample locations and the corresponding simulation results of the statistics. If 'X' equals NULL, then the sample points are taken from the object 'runs'.

The most critical part is the decomposition of variance matrices for each sample location unless 'var.type' equals "const" in which case a constant variance matrix approximation is expected later by function qle. The Cholesky decompositions are used for average approximations of the variance matrix of the statistics when calculating the quasi-score vector or any type of function criterion based on the Mahalanobis distance or quasi-deviance. If these fail for any reason we try to ignore, if possible, the corresponding sample points and exclude them from all following calculations. Unless a constant estimate of the variance matrix, the default is to approximate the variance at any model parameter by either a kriging interpolation of the Cholesky terms or as an average over all sampled variance matrices also based on the decomposed Cholesky terms (see vignette).

An object of class QLdata as a data frame with columns:

`X`	Model parameters (`n=1`,...,`q`)
`mean`	Results of simulations (`m=1`,...,`p`)
`var`	Simulation variances of statistics (`m=1`,...,`p`)
`L`	if applicable, Cholesky decomposed terms of variance matrices of statistics (k=1,...,(m*(m+1)/2))

where 'p' denotes the number of user defined statistics and 'q' the problem dimension, that is, the number of parameters to be estimated.

The following items are stored as attributes:

`type`	see above
`nsim`	number of simulations at each point
`xdim`	parameter dimension
`nWarnings`	Number of warnings during simulations
`nErrors`	Number of errors during simulations
`nIgnored`	List of parameters ignored (because of failures)

M. Baaske

# simulate model statistics at LHS design
sim <- simQLdata(sim =
         function(x,cond) {
           X <- rlnorm(cond$n,x[1],x[2])
           c("MED"=median(X),"MAD"=mad(X))
         },
         cond=list("n"=10),
         nsim=10, N=10, method="maximinLHS",
         lb=c("mu"=-1.5,"sd"=0), ub=c("mu"=2,"sd"=1))

# setup the QL data model using defaults
qldata <- setQLdata(sim,verbose=TRUE)