gamselBayes (version 2.0-1)

gamselBayes.control: Controlling Bayesian generalized additive model selection

Description

Function for optional use in calls to gamselBayes() to control spline basis dimension, hyperparameter choice and other specifications for generalized additive model selection via using Bayesian inference.

Usage

gamselBayes.control(numIntKnots = 25,truncateBasis = TRUE,numBasis = 12,
                    sigmabeta0 = 100000,sbeta = 1000,sepsilon = 1000,
                    su= 1000,rhoBeta = 0.5,rhoU = 0.5,nWarm = 1000,
                    nKept = 1000,nThin = 1,maxIter = 1000,toler = 1e-8,
                    msgCode = 1)

Value

A list containing values of each of the seventeen control parameters, packaged to supply the control argument to gamselBayes. The values for gamselBayes.control can be specified in the call to gamselBayes.

Arguments

numIntKnots

The number of interior knots used in construction of the Demmler-Reinsch spline basis functions. The value of numIntKnots must be an integer between 8 and 50.

truncateBasis

Boolean flag:
TRUE = truncate the Demmler-Reinsch spline basis to the value specified by numBasis (the default)
FALSE = produce a black-and-white version of the graphical check.

numBasis

The number of spline basis functions retained after truncation when truncateBasis is TRUE. The value of numBasis cannot exceed numIntKnots+2. The default value of numBasis is 12.

sigmabeta0

The standard deviation hyperparameter for the Normal prior distribution on the intercept parameter for the standardized version of the data used in Bayesian inference. The default is 100000.

sbeta

The standard deviation hyperparameter for the Normal prior distribution on the linear component coefficients parameters for the standardized version of the data used in Bayesian inference. The default is 1000.

sepsilon

The scale hyperparameter for the Half-Cauchy prior distribution on the error standard deviation parameter for the standardized version of the data used in Bayesian inference. The default is 1000.

su

The scale hyperparameter for the Half-Cauchy prior distribution on the spline basis coefficients standard deviation parameter for the standardized version of the data used in Bayesian inference. The default is 1000.

rhoBeta

The probability parameter for the Bernoulli prior distribution on the linear component coefficients spike-and-slab auxiliary indicator variables probability parameter. The default is 0.5.

rhoU

The probability parameter for the Bernoulli prior distribution on the spline basis coefficients spike-and-slab auxiliary indicator variables probability parameter. The default is 0.5.

nWarm

The size of the Markov chain Monte Carlo warmup, a positive integer, when method is Markov chain Monte Carlo. The default is 1000.

nKept

The size of the kept Markov chain Monte Carlo samples, a positive integer, when the method is Markov chain Monte Carlo. The default is 1000.

nThin

The thinning factor for the kept Markov chain Monte Carlo samples, a positive integer, when the method is Markov chain Monte Carlo. The default is 1.

maxIter

The maximum number of mean field variational Bayes iterations, a positive integer, when the method is mean field variational Bayes. The default is 1000.

toler

The convergence tolerance for mean field variational Bayes iterations, a positive number less than 0.5, when the method is mean field variational Bayes. Convergence is deemed to have occurred if the relative change in the logarithm of the approximate marginal likelihood falls below toler. The default is 1e-8.

msgCode

Code for the nature of progress messages printed to the screen. If msgCode=0 then no progress messages are printed. If msgCode=1 then a messages printed each time approximately each 10% of the sampling is completed. If msgCode=2 then a messages printed each time approximately each 10% of the sampling is completed. The default is 1.

Author

Virginia X. He and Matt P. Wand

References

He, V.X. and Wand, M.P. (2021). Generalized additive model selection via Bayesian inference. Submitted.

Examples

library(gamselBayes) 

# Generate some simple regression-type data:

set.seed(1) ; n <- 1000 ; x1 <- rbinom(n,1,0.5)  
x2 <- runif(n) ; x3 <- runif(n) ; x4 <- runif(n)
y <- x1 + sin(2*pi*x2) - x3 + rnorm(n)
Xlinear <- data.frame(x1) ; Xgeneral <- data.frame(x2,x3,x4)

# Obtain a gamselBayes() fit for the data:

fit <- gamselBayes(y,Xlinear,Xgeneral)
summary(fit) ; plot(fit) ; checkChains(fit)

# Now modify some of the control values:

fitControlled <- gamselBayes(y,Xlinear,Xgeneral,control = gamselBayes.control(
                             numIntKnots = 35,truncateBasis = FALSE,
                             sbeta = 10000,su = 10000,nWarm = 2000,nKept = 1500))
summary(fitControlled) ; plot(fitControlled) ; checkChains(fitControlled)