Starting Values and Fine Tuning of the MCEM algorithm

Description

Function latpos.start is used to construct “good” starting values, while function latpos.control provides settings for the numerical aspects of the MCEM algorithm, with reasonable defaults.

Usage

latpos.start(resp,latent.dims,manifest,start,
             unfold.method,restrictions=standard.restrictions,
             maxiter,...)
latpos.control(maxiter=200,initial.size=101,
                          Lambda.alpha=.05,
                          Lambda.eps=1e-7,
                          diff.logLik.eps=1e-7,
                          abs.diff.psi.eps=0,
                          rel.diff.psi.eps=0,
                          max.size=Inf,
                          min.final.size=1000,
                          force.increase=TRUE,
                          Q.linesearch=TRUE,
                          ...)

Arguments

resp

an internal representation of the observed data.

latent.dims

a character vector with the names of the axes of the latent space.

manifest

a character vector with the names of the observed variables, i.e. emphasis counts of policy objectives.

start

an optional list with starting values for the model parameters

unfold.method

the unfolding method to be used to generate reasonable starting values.

restrictions

an object representing restrictions on the model parameters, see restrictor.

maxiter

the maximum number of iterations to use, in latpos.start to get initial values for the posterior modes, in latpos.control to set the maximum number of MCEM iterations.

initial.size

a positive number, the simulation sample size to use in the first MCEM iteration.

Lambda.alpha

a “significance level” for the increase of the Q-function. If the increase is not “statistically significant” at this level, the sample size is automatically increased.

Lambda.eps

a non-negative number as convergence critierion. If the increase of the Q-function is smaller than this value, convergence of the MCEM is declared.

diff.logLik.eps

a non-negative number as convergence critierion. If the increase of the marginal log-likelihood is smaller than this value, convergence of the MCEM is declared.

abs.diff.psi.eps

a non-negative number as an alternative convergence critierion. if the absolute change of the model parameters is smaller than this value, convergence of the MCEM is declared.

rel.diff.psi.eps

a non-negative number as an alternative convergence critierion. if the absolute change of the model parameters is smaller than this value, convergence of the MCEM is declared.

max.size

a positive number, the maximum simulation sample size to be used.

min.final.size

a positive number, the minimal simulation sample size to be used in the final iterations of the MCEM algorithm.

force.increase

logical; if TRUE and the likelihood or the Q-function cannot be increased then conduct a line search for the optimal step size.

Q.linesearch

logical; if TRUE, force.increase==TRUE and the likelihood or the Q-function cannot be increased then conduct a line search for the optimal step size; if FALSE, but force.increase==TRUE and the likelihood or the Q-function cannot be increased then step back to the values of the previous iteration.

...

other arguments, ignored.