modelSelectionGGM {mombf}

Bayesian variable selection for linear models via non-local priors.

Description

Bayesian model selection for linear, asymmetric linear, median and quantile regression under non-local or Zellner priors. p>>n can be handled.

modelSelection enumerates all models when feasible and uses a Gibbs scheme otherwise. See coef and coefByModel for estimates and posterior intervals of regression coefficients, and rnlp for posterior samples.

modelsearchBlockDiag seeks the highest posterior probability model using an iterative block search.

Usage

modelSelectionGGM(y, priorCoef=normalidprior(tau=1), 
priorModel=modelbinomprior(1/ncol(y)), 
priorDiag=exponentialprior(lambda=1), center=TRUE, scale=TRUE, 
almost_parallel= FALSE, sampler='Gibbs', niter=10^3, 
burnin= round(niter/10), pbirth=0.5, nbirth, 
Omegaini='glasso-ebic', verbose=TRUE)

Arguments

y

Data matrix

priorCoef

Prior on off-diagonal entries of the precision matrix, conditional on their not being zero (slab)

priorModel

Prior probabilities on having non-zero diagonal entries

priorDiag

Prior on diagonal entries of the precision matrix

center

If TRUE, the columns of y will be centered to zero mean

scale

If TRUE, the columns of y will be scaled to unit sample variance

almost_parallel

Use almost parallel algorithm sampling from each column independently and using an MH step

sampler

Posterior sampler. Options are "Gibbs", "birthdeath" and "zigzag"

niter

Number of posterior samples to be obtained

pbirth

Probability of a birth move. Ignored unless sampler=="birthdeath"

nbirth

Number of birth/death updates to perform for each row of the precision matrix. Defaults to ncol(y)

burnin

The first burnin samples will be discarded

Omegaini

Initial value of the precision matrix Omega. "null" sets all off-diagonal entries to 0. "glasso-bic" and "glasso-ebic" use GLASSO with regularization parameter set via BIC/EBIC, respectively. Alternatively, Omegaini can be a matrix

verbose

Set verbose==TRUE to print iteration progress

Details

Let Omega be the inverse covariance matrix. A spike-and-slab prior is used. Specifically, independent priors are set on all Omega[j,k], and then a positive-definiteness truncation is added.

The prior on diagonal entries Omega[j,j] is given by priorDiag. Off-diagonal Omega[j,k] are equal to zero with probability given by modelPrior and, when non-zero, they are

Independent spike-and-slab priors are set on the off-diagonal entries of Omega, i.e. Omega[j,k]=0 with positive probability (spike) and otherwise arises from the distribution indicated in priorCoef (slab).

Value

Posterior inference on the inverse covariance of y. Object of class msfit_ggm, which extends a list with elements

postSample

Posterior samples for the upper-diagonal entries of the precision matrix. Stored as a sparse matrix, see package Matrix to utilities to work with such matrices

p

Number of columns in y

priors

List storing the priors specified when calling modelSelectionGGM

Author(s)

David Rossell

See Also

msfit_ggm-class for further details on the output. icov for the estimated precision (inverse covariance) matrix. coef.msfit_ggm for Bayesian model averaging estimates and intervals.

Examples

#Simulate data with p=3
Th= diag(3); Th[1,2]= Th[2,1]= 0.5
sigma= solve(Th)

z= matrix(rnorm(1000*3), ncol=3)
y= z 

#Obtain posterior samples
fit= modelSelectionGGM(y, scale=FALSE)

#Parameter estimates, intervals, prob of non-zero
coef(fit)

#Estimated inverse covariance
icov(fit)

#Estimated inverse covariance, entries set to 0
icov(fit, threshold=0.95)

#Shows first posterior samples
head(fit$postSample)