wilcox_R: What the package does (short line)

regpreS <-
function(x,y,regfun=lsfit,error=absfun,nboot=100,
mval=round(5*log(length(y))),locfun=mean,pr=TRUE,
xout=FALSE,outfun=out,
plotit=TRUE,xlab="Model Number",ylab="Prediction Error",SEED=TRUE,...){
#
#   Stepwise selection of predictors based on
#   estimates of  prediction error using the regression method
#   regfun,
#   which defaults to least squares.  Prediction error
#   is estimated with .632 method.
#   (See Efron and Tibshirani, 1993, pp. 252--254)
#
#   The predictor values are assumed to be in the n by p matrix x.
#   The default number of bootstrap samples is nboot=100
#
#   Prediction error is the expected value of the function error.
#   The argument error defaults to absolute  error. To use
#   squared error, set error=sqfun.
#
#   regfun can be any R function that returns the coefficients in
#   the vector regfun$coef, the first element of which contains the
#   estimated intercept, the second element contains the estimate of
#   the first predictor, etc.
#
#   The default value for mval, the number of observations to resample
#   for each of the B bootstrap samples is based on results by
#   Shao (JASA, 1996, 655-665). (Resampling n vectors of observations,
#   model selection may not lead to the correct model as n->infinity.
#
if(SEED)set.seed(2)
q=ncol(x)
qm1=q-1
x<-as.matrix(x)
d<-ncol(x)
p1<-d+1
temp<-elimna(cbind(x,y))
x<-temp[,1:d]
y<-temp[,d+1]
x<-as.matrix(x)
if(xout){
x<-as.matrix(x)
flag<-outfun(x,SEED=FALSE,...)$keep
x<-x[flag,]
y<-y[flag]
x<-as.matrix(x)
}
adit=NULL
pval=c(1:ncol(x))
#pval=c(1:q)
allp=pval
for(ip in 1:qm1){
model=list()
for(j  in 1:length(pval))model[[j]]=c(adit,pval[j])
temp=regpre(x,y,model=model,pr=FALSE,plotit=FALSE,adz=FALSE,regfun=regfun,
SEED=SEED)$estimates
pbest=order(temp[,5])
adit=model[[pbest[1]]]
pval=allp[-adit]
}
output=model[[pbest[1]]]
output=c(output,allp[-output])
output
}