R help - Apr 2010 - Caret package and lasso

Dear all,

I have used following code but everytime I encounter a problem of not having
coefficients for all the variables in the predictor set.

# code

rm(list=ls())
library(caret)
# generating response and design matrix

X<-matrix(rnorm(50*100),nrow=50)
y<-rnorm(50*1)

# Applying caret package

con<-trainControl(method="cv",number=10)

data<-NULL
data<- train(X,y, "lasso", metric="RMSE",tuneLength = 10,
trControl = con)

coefs<-predict(data$finalModel,s=data$bestTune$.fraction, type
="coefficients", mode ="fraction")$coef
coefs

*This is the output which I got :*

you can see some of the predictors are missing  like V4, V6, V7

         V1          V2          V3          V5          V8
V9         V10         V11         V13         V14         V15
V17         V19         V22         V24         V26
 0.00000000  0.00000000  0.00000000  0.00000000  0.00000000  0.00000000
0.06165530  0.02693335  0.00000000  0.00000000  0.00000000 -0.15699831
0.00000000  0.00000000  0.00000000  0.00000000
        V27         V28         V33         V35         V36
V37         V39         V41         V42         V43         V45
V46         V47         V48         V49         V50
 0.00000000  0.00000000  0.00000000  0.00000000  0.00000000  0.00000000
0.00000000  0.00000000  0.00000000  0.00000000  0.00000000 -0.01881011
0.00000000  0.00000000  0.00000000  0.00000000
        V51         V52         V54         V55         V56
V57         V58         V60         V61         V64         V65
V66         V67         V72         V74         V75
 0.00000000  0.00000000  0.00000000  0.00000000  0.00000000  0.00000000
0.00000000  0.00000000 -0.02772797  0.01659148  0.00000000  0.00000000
0.00000000  0.00000000  0.00000000  0.21293642
        V77         V78         V79         V81         V84
V85         V86         V88         V91         V94         V95
V99        V100
 0.00000000  0.00000000  0.00000000  0.04849013  0.04563922  0.00000000
0.00000000  0.00000000  0.00000000  0.06291593  0.00000000  0.00000000
0.00000000

Thanks in advance

-- 
Linda Garcia

	[[alternative HTML version deleted]]

Linda,

Thanks for the example.

I did this to make it more reproducible:

  set.seed(1)
  X<-matrix(rnorm(50*100),nrow=50)
  y<-rnorm(50*1)

  dimnames(X)

  colnames(X) <- paste("V", 1:nrow(X))

  # Applying caret package

  set.seed(2)
  con<-trainControl(method="cv",number=10)

  data<-NULL
  data<- train(X,y, "lasso", metric="RMSE",tuneLength =
10, trControl = con)

I see your point here, but this code gives the same results:

  fit2 <- enet(X, y, lambda = 0)
  predict(fit2, mode = "fraction", s = data$bestTune$.fraction, type
"coefficient")$coef

(at least train() names the predictors).

To me, it looks like enet is doing some filtering:

   > dim(X)
   [1]  50 100
   > length(fit2$meanx)
   [1] 56

This appears to be independent of caret. I would contact the package
maintainer off-list and ask.

Max