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

In gfmR: Implements Group Fused Multinomial Regression

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

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.

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

Author(s)

Brad Price, [e-mail].

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: 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))

## End(Not run)

gfmR documentation built on May 1, 2019, 8:41 p.m.