twostep.mgp: Two-step Markov chain Monte Carlo algorithm for multivariate...

In lbelzile/mgp: Multivariate generalized Pareto models

Two-step Markov chain Monte Carlo algorithm for multivariate generalized Pareto models

Description

The algorithm estimates marginal parameters in a preliminary run, then fix the latter to the posterior median and samples from the dependence parameters in a second time.

Usage

twostep.mgp(dat, mthresh, thresh, lambdau = 1, model = c("br", "xstud",
  "lgm"), coord, start, numiter = 40000L, burnin = 5000L, thin = 1L,
  verbose = 100L, filename, censor = TRUE, keepburnin = TRUE,
  geoaniso = TRUE, blockupsize = ncol(dat), transform = FALSE,
  likt = c("mgp", "pois", "binom"), saveinterm = 500L, ...)

Arguments

dat	n by D matrix of observations
mthresh	vector of marginal thresholds under which data are censored
thresh	functional max threshold determining the risk region
lambdau	probability of exceedance of the threshold for censored observations
model	dependence model, either of "br", "xstud" or "lgm", in which case the model uses the independence likelihood with generalized Pareto margins
coord	matrix of coordinates, with longitude and latitude in the first two columns and additional covariates for the latent Gaussian model
start	named list with starting values for the parameters, with arguments: scale: a D vector of scale parameters, strictly positive. shape: a scale containing the shape and satisfying support constraints for all sites. marg.pcov: initial proposal covariance for the marginal parameters (scale and shape). dep: initial values for the dependence function. dep: dependence function, with distance matrix as first argument and dep as second argument. aniso: initial values for the anisotropy parameters, scale and angle. df: degrees of freedom (if model == "xstud"). dep.lb: lower bounds for the dependence parameters dep.ub: upper bounds for the dependence parameters If any of scale, shape or marg.pcov are missing, the function will attempt to find starting values.
numiter	number of iterations to be returned
burnin	number of initial parameters for adaptation and discarded values.
thin	thining parameter; only every thin iteration is saved
verbose	report current values via print every verbose iterations.
filename	name of file for save.
censor	logical; should censored likelihood be used? Default to TRUE
keepburnin	logical; should initial runs during burnin be kept for diagnostic. Default to TRUE.
geoaniso	logical; should geometric anisotropy be included? Default to TRUE.
blockupsize	size of block for updates of the scale parameter; ncol(dat) yields individual updates
transform	logical; should parameters be sampled on an unconstrained space if they are bounded. Default is FALSE, in which case the Metropolis-Hastings proposals are performed by sampling from multivariate truncated Gaussian
likt	string indicating the type of likelihood, with an additional contribution for the non-exceeding components: one of "mgp", "binom" and "pois".
saveinterm	integer indicating when to save results. Default to 500L.
...	Arguments passed on to clikmgp mthresh vector of individuals thresholds under which observations are censored dat matrix of observations thresh functional threshold for the maximum loc vector of location parameter for the marginal generalized Pareto distribution scale vector of scale parameter for the marginal generalized Pareto distribution shape vector of shape parameter for the marginal generalized Pareto distribution par list of parameters: alpha for the logistic model, Lambda for the Brown–Resnick model or else Sigma and df for the extremal Student. model string indicating the model family, one of "log", "br" or "xstud" likt string indicating the type of likelihood, with an additional contribution for the non-exceeding components: one of "mgp", "binom" and "pois". lambdau vector of marginal rate of marginal threshold exceedance.

Value

a list with res containing the results of the chain

lbelzile/mgp documentation built on Aug. 5, 2023, 2:34 a.m.