bestglm (version 0.13)

bestglm-package: bestglm: Best Subset GLM

Description

Provides new information criterion BICq as well as AIC, BIC and EBIC for selecting the best model. Additionally, various CV algorithms are also provided.

Details

ll{ Package: bestglm Type: Package Version: 0.13 Date: 2009-09-09 License: GLP 2.0 or greater LazyLoad: yes } bestglm is the main function. All other functions are utility functions and are not normally invoked. Many examples are provided in the vignettes accompanying this package. The vignettes are produced using the R package Sweave and so R scripts can easily be extracted. The R package xtable is needed for the vignette in SimExperimentBICq.Rnw.

References

Xu, C. and McLeod, A.I. (2009). Improved Extended Bayesian Information Criterion. Submitted for publication.

See Also

leaps

Examples

data(zprostate)
train<-(zprostate[zprostate[,10],])[,-10]
#Best subset using AIC
bestglm(train, IC="AIC")
#Best subset using BIC
bestglm(train, IC="BIC")
#Best subset using EBIC
bestglm(train, IC="BICg")
#Best subset using BICg with g=0.5 (tuning parameter)
bestglm(train, IC="BICg", t=0.5)
#Best subset using BICq. Note BICq with q=0.25 is default.
bestglm(train, IC="BICq")
#Best subset using BICq with q=0.5 (equivalent to BIC)
bestglm(train, IC="BICq", t=0.5)
#Remark: set seed since CV depends on it
set.seed(123321123)
#Best subset using 10-fold CV. Default when IC="CV"
bestglm(train, IC="CV")
#Best subset using 5-fold adjusted CV (Davison & Hartigan algorithm)
bestglm(train, IC="CV", CVArgs=list(Method="DH", K=5, REP=1))
#Best subset using delete-d CV, with d=5 and 100 replications
bestglm(train, IC="CV", CVArgs=list(Method="d", K=5, REP=100))
#Best subset, logistic regression
data(SAheart)
bestglm(SAheart, IC="BIC", family=binomial)
#Best subset, factor variables with more than 2 levels
data(AirQuality)
#full regression
bestglm(AirQuality, IC="BICq", t=1)
#subset
bestglm(AirQuality, IC="BICq")