choosethin {systemicrisk}

Calibrate Thinning

Description

Attempts to automatically choose a thinning paramter to achieve an overall relative effective sample size (defined as the effective sample size divided by the number of samples) for all parameters in the model (that do not seem to be constant). This function provides no guarantees that the desired relative effective sample size (rESS) will actually be achieved - it is best treated as a rough guide for this.

Usage

choosethin(
  l,
  a,
  L_fixed = NA,
  model,
  relESStarget = 0.3,
  burnin = 100,
  matrpertheta = length(l)^2,
  silent = TRUE,
  maxthin = 10000
)

Arguments

l

observed row sum

a

observerd column sum

L_fixed

Matrix containing known values of L, where NA signifies that an element is not known. If L_fixed equates to NA (the default) then no values are assumed to be known.

model

Underlying model for p and lambda.

relESStarget

Target for the relative effective sample size, must be in (0,1). Default 0.3.

burnin

number of iterations for the burnin. Defaults to 5 of the steps in the sampling part.

matrpertheta

number of matrix updates per update of theta.

silent

(default FALSE) suppress all output (including progress bars).

maxthin

Upper bound on thinning to consider. Default 10000.

Details

The approach used involves a pilot run of the sampler, followed by a computation of the acf (autocorrelation function) for each component. The acf is used only up to (and excluding) the point used where it becomes negative for the first time. This part of the acf is then used to approximate the rESS and to determine the amount of thinning needed. The reported result is the thinning needed to achieve the rESS for all components (the matrix as well as the parameter theta). The initial pilot run may not be sufficient and further pilot runs may have to be started.

Value

An integer describing the amount of thinning required.

Examples

set.seed(12689)
n <- 10
m <- Model.Indep.p.lambda(Model.p.BetaPrior(n),
                          Model.lambda.GammaPrior(n,scale=1e-1))
x <- genL(m)
l <- rowSums(x$L)
a <- colSums(x$L)
choosethin(l,a,model=m)
choosethin(l,a,model=m,relESStarget=0.7)