exhaustive: Evaluate all possible models

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

Description

Evaluate all possible models

Usage

  exhaustive(modelData,
    modelPrior = c("flat", "exponential", "independent", "dependent", "dep.linear"),
    modelConfigs = NULL, algorithm = c("2", "1"),
    computation = getComputation(), order = FALSE)

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.
modelConfigs	optional matrix of model configurations, which are then evaluated instead of all possible configurations. It is check for coherency with modelData.
algorithm	either “2” (default, fast for small dimension of spline coefficients) or “1” (fast for small number of observations), specifying the algorithm version. Only matters for normal models, for GLMs always a type 2 algorithm is used.
computation	computation options produced by getComputation, only matters for generalised response models.
order	should the models be ordered after their posterior probability? (default: FALSE for backwards compatibility)

Value

a list with the data frame “models” comprising the model configurations, (R2 for normal models) / log marginal likelihoods / log priors / posteriors; and the inclusion probabilities matrix “inclusionProbs”.

Author(s)

Daniel Sabanes Bove [e-mail]

Examples

## get some data
attach(longley)

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

## get a list of all possible models with this data
res <- exhaustive(md)

res

## now the same, but with cubic splines and algorithm 1:

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

## get a list of all possible models with this data
res <- exhaustive(md,
                  algorithm="1")

res

## now only compute for two certain model configurations:
configs <- cbind(GNP=c(1L, 3L),
                 Armed.Forces=c(2L, 3L))
res <- exhaustive(md,
                  modelConfigs=configs)

## now for generalised response:

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

## and do the exhaustive search
res <- exhaustive(md,
                  modelPrior="dependent",
                  computation=
                  getComputation(higherOrderCorrection=FALSE,
                                 debug=FALSE))

res$models <- res$models[order(res$models$post, decreasing=TRUE), ]
res

res1 <- exhaustive(md,
                  computation=
                  getComputation(higherOrderCorrection=FALSE,
                                 debug=FALSE))
res2 <- exhaustive(md,
                  computation=
                  getComputation(higherOrderCorrection=FALSE,
                                 debug=FALSE))
res3 <- exhaustive(md,
                  computation=
                  getComputation(higherOrderCorrection=FALSE,
                                 debug=TRUE))

str(res1)
identical(res1, res2)
identical(res1, res3)


## now with offsets:
set.seed(93)
offsets <- rnorm(n=length(Employed))

md <- glmModelData(y=round(Employed / 10),
                   X=cbind(GNP, Armed.Forces),
                   family=poisson,
                   offsets=offsets)

res <- exhaustive(md,
                  computation=
                  getComputation(higherOrderCorrection=TRUE,
                                 debug=TRUE))
res




res <- exhaustive(md,
                  computation=
                  getComputation(higherOrderCorrection=TRUE,
                                 debug=TRUE))
res

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