mirl: mirl

In summeryingliu/MIRL: Multiple Imputation Random Lasso for Variable Selection with Missing Entries

Description

This function produce the stability selection probability and the estimated coefficients

Usage

mirl(x=NULL,y,q2,im=5,E=NULL,lam=exp(seq(from=log(0.55),to=log(0.001),length.out=70)))

Arguments

x	This is the n by p design matrix with missing entries.
y	This is a n by 1 vector of the outcome, the outcome should be non missing. Please delete any sample with missing outcome.
q2	This is the number of variables to be bootstraped, recommended size is p/2.
im	Number of multiple imputation, increase of im will increase time cost. Default is 5.
E	You can use the 'mice' function in the mice package to generate this E which is a 'mids' data type, if E is entered, E will be used and x will be ignored.
lam	The vector of tunning parameter for each lasso implemented.

Value

Probabilty	This is the selection probability for each covariate. The larger the probability, the more significant the variable is related to the outcome. Notice that the probability and coef are p+1 vectors and the first coef is the intercept term, where the probability is always zero.
coef	the coefficient estimated

Author(s)

Ying Liu

References

Liu Y, Wang Y, Feng Y, Wall MM. VARIABLE SELECTION AND PREDICTION WITH INCOMPLETE HIGH-DIMENSIONAL DATA. The annals of applied statistics. 2016;10(1):418-450. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4872715/

Examples

#This example is a similar simulation setting in the reference paper.
cor=0.6
prob=0.02
p=10
n=200 #sample size
Sigma=matrix(cor,p,p)#correlation of predictors
diag(Sigma)=1
mu=numeric(p)
set.seed(3)
#C is the complete design matrix without missing
C=mvrnorm(n,mu,Sigma)
#The missing indicator matrix
A<-matrix(rbinom(n*p,size=1,prob),n,p)
A[,c(1,4,6)]=0 #columns without missing
p1=inv.logit(C[,1]+C[,6]-2)
A[,5]=rbinom(n,size=1,p1) #Missing at Random
p2=inv.logit(-C[,1]-0.5*C[,6]-2)
A[,10]=rbinom(n,size=1,p2)
p3=inv.logit(C[,4]-2)
A[,9]=rbinom(n,size=1,p3)
beta=numeric(p)
beta[1:6]=c(0.1,0.2,0.5,-0.3,-.4,-0.5)*5
ct=c(0,beta)
#generating Y
Y=C%*%beta+rnorm(n)
B=C
B[A==1]=NA

fit<-mirl(B,Y,p/2,im=5)
cbind(fit$coef,fit$Probability)

summeryingliu/MIRL documentation built on May 7, 2019, 9:39 a.m.