stochSearch: Stochastic search for models with high posterior probability

In hypergsplines: Bayesian model selection with penalised splines and hyper-g prior

Description

Stochastic search for models with high posterior probability

Usage

  stochSearch(modelData,
    modelPrior = c("flat", "exponential", "independent", "dependent", "dep.linear"),
    startModel = rep(0, modelData$nCovs),
    chainlength = 100000L, nCache = chainlength,
    nModels = as.integer(max(nCache/100, 1)),
    computation = getComputation())

Arguments

modelData	the data necessary for model estimation, which is the result from modelData or glmModelData
modelPrior	either “flat” (default), “exponential”, “independent”, “dependent”, or “dep.linear”, see getLogModelPrior for details.
startModel	model configuration where the MCMC chain starts. Defaults to the null model. Checked for coherency with modelData.
chainlength	length of the model sampling chain (default: 100,000)
nCache	maximum number of best models to be cached at the same time during the model sampling (by default equal to chainlength)
nModels	how many best models should be saved? (default: 1% of the total number of nCache). Must not be larger than nCache.
computation	computation options produced by getComputation, only matters for generalised response models.

Value

a list with the data frame “models” comprising the model configurations, log marginal likelihoods / priors / posteriors and hits in the MCMC run, the inclusion probabilities matrix “inclusionProbs”, the number of total visited models “numVisited” and the log normalization constant “logNormConst”.

Author(s)

Daniel Sabanes Bove [e-mail]

Examples

## get some data
attach(longley)

## get large model data
md <- modelData(y=Employed,
                X=
                cbind(GNP,
                      Armed.Forces,
                      Population,
                      Year))

## do a stochastic search over the model space
res <- stochSearch(md)
res

## now the same, but with cubic splines:

## get large model data
md <- modelData(y=Employed,
                X=
                cbind(GNP,
                      Armed.Forces,
                      Population,
                      Year),
                splineType="cubic")

## do a stochastic search over the model space,
## and choose a special start model
res <- stochSearch(md,
                   startModel=c(2, 2, 2, 2))
res

## and now for generalised response:

## get the model data
md <- glmModelData(y=as.numeric(Employed > 64),
                   X=
                   cbind(GNP,
                         Armed.Forces,
                         Population,
                         Year),
                   family=binomial)

## do a stochastic search over the model space,
## also with a special start model
res <- stochSearch(md,
                   startModel=c(0, 1, 2, 1),
                   chainlength=1000L,
                   computation=
                   getComputation(higherOrderCorrection=FALSE))
res

hypergsplines documentation built on May 2, 2019, 6:14 p.m.