gammSlice (version 2.0-2)

gSlc.control: Controlling generalized additive mixed model fitting via slice sampling

Description

Function for optional use in calls to gSlc() to control Markov chain Monte Carlo sample sizes values and other specifications for slice sampling-based fitting of generalized additive mixed models.

Usage

gSlc.control(nBurn=5000,nKept=5000,nThin=5,fixedEffPriorVar=1e10,
             sdPriorScale=1e5,numBasis=NULL,preTransfData=TRUE,msgCode=1)

Arguments

nBurn

The length of the Markov chain Monte Carlo burnin. The first nBurnin Markov chain Monte Carlo samples are discarded. The default value of nBurnin is 5000.

nKept

The number of kept Markov chain Monte Carlo samples after the burnin period. The default value of nKept is 5000.

nThin

Thinning factor applied to the retained Markov chain Monte Carlo samples. Setting nThin to be an intege greater than 1 results in every nThinth value in the post-burnin samples being retained. The final Markov chain Monte Carlo sample size is an integer close to nIter divided by nIter. The default value of nThin is 5.

fixedEffPriorVar

The variance in the independent zero mean Normal priors of the fixed effect parameters after the data of each predictor have been transformed to the interval [0,1]. The default value of fixedEffPriorVar is 1e10.

sdPriorScale

The scale parameter in the Half Cauchy priors on standard deviation parameters after the data of each predictor have been transformed to the interval [0,1]. The default value of sdPriorScale is 1e5.

numBasis

Vector of positive integers specifying the number of spline basis functions to be used for each smooth function component.

preTransfData

Boolean flag: TRUE = pre-transform each of the predictors to unit interval for Bayesian analysis with the priors specified by fixedEffPriorVar and sdPriorScale (the default), FALSE = do not perform any pre-transformation of the predictors.

msgCode

A code for specification of the nature of messages printed concerning progress of the Markov chain Monte Carlo sampling: 0 = no messages printed, 1 = percentages 1,2,...,10 and then 20,30,...,100 (the default), 2 = percentages 1,2,...,100, 3 = percentages 10,20,...,100.

References

Pham, T. and Wand, M.P. (2018). Generalized additive mixed model analysis via gammSlice. Australian and New Zealand Journal of Statistics, 60, 279-300.

Zhao, Y., Staudenmayer, J., Coull, B.A. and Wand, M.P. (2003). General design Bayesian generalized linear mixed models. Statistical Science, 21, 35-51.

See Also

gSlc

Examples

# NOT RUN {
library(gammSlice)
set.seed(39402) ; m <- 100 ; n <- 2
beta0True <- 0.5 ; betaxTrue <- 1.7 ; sigsqTrue <- 0.8
idnum <- rep(1:m,each=n) ; x <- runif(m*n)
U <- rep(rnorm(m,0,sqrt(sigsqTrue)),each=n)
mu <- 1/(1+exp(-(beta0True+betaxTrue*x+U)))
y <- rbinom((m*n),1,mu)
fit <- gSlc(y ~ x,random = list(idnum = ~1),family = "binomial")
summary(fit)

# Illustration of user-specified priors:

fitMyPriors <- gSlc(y ~ x,random = list(idnum = ~1), 
                    family = "binomial", 
                    control = gSlc.control(fixedEffPriorVar=1e13, 
                                           sdPriorScale=1e3))
summary(fitMyPriors)

# Illustration of specification of larger Markov chain Monte Carlo samples:

fitBigMCMC <- gSlc(y ~ x,random = list(idnum = ~1),family = "binomial",
                   control = gSlc.control(nBurn=10000,nKept=8000,nThin=10))
summary(fitBigMCMC)
# }