gfmR (version 1.1-0)

GFMR.cv: Tuning parameter selection using validation likelihood for GFMR.

Description

This routine implements K fold cross validation for group fused multinomial regression.

Usage

GFMR.cv(Y,X,lamb,sampID,H,n.cores=1,rho=10^-8,...)

Arguments

Y

A matrix of response category counts where the columns represent the categories and rows represent the observations. Currently supported for n=1.

X

A matrix of predictor variables. The columns represent predictors and rows represent observations.

lamb

tuning parameter for fusion penalty

sampID

An identified or the sampleID for the cross validation routine. Should take values 1:k and be user supplied.

H

An indicator matrix representing the edge set of the penalty set. The matrix is square and symmetric with dimension number of response categories, and if two categories are in the penalty set a 1 should be in the row column combination.

n.cores

The number of cores for the mclapply function

rho

Step Size parameter of ADMM

...

Other arguments for Group fused multinomial regression

Value

A list is returned with elements

vl

The validation likelihood for all tuning parameters

vl.sd

The standard deviation of the validation likelihood

lambda

tuning parameter used

vl.mat

Validation likelihood for each

Details

This routine implements the validation likelihood approach proposed by Price et. al. to obtain the tuning parameter for Group Fused Multinomial Regression which automatically combines response categories in multinomial regression. Show to do well with regard to predicting the true category probabilities.

References

Price, B.S, Geyer, C.J. and Rothman, A.J. "Automatic Response Category Combination in Multinomial Logistic Regression." https://arxiv.org/abs/1705.03594.

Examples

# NOT RUN {
# }
# NOT RUN {
data(nes96)
attach(nes96)
Response=matrix(0,944,7)
for(i in 1:944){
  if(PID[i]=="strRep"){Response[i,1]=1}
  if(PID[i]=="weakRep"){Response[i,2]=1}
  if(PID[i]=="indRep"){Response[i,3]=1}
  if(PID[i]=="indind"){Response[i,4]=1}
  if(PID[i]=="indDem"){Response[i,5]=1}
  if(PID[i]=="weakDem"){Response[i,6]=1}
  if(PID[i]=="strDem"){Response[i,7]=1}
}

Hmat=matrix(1,dim(Response)[2],dim(Response)[2])
diag(Hmat)=0
ModMat<-lm(popul~age,x=TRUE)$x

X=cbind(ModMat[,1],apply(ModMat[,-1],2,scale))

set.seed(1010)
n=dim(Response)[1]
sampID=rep(5,n)
samps=sample(1:n)
mine=floor(n/5)
for(j in 1:4){
  sampID[samps[((j-1)*mine+1):(j*mine)]]=j
}

o1<-GFMR.cv(Response,X,lamb = 2^seq(4.2,4.3,.1),H=Hmat2,sampID = sampID,n.cores =5)

which(o1$vl==max(o1$vl))
# }