# Error using rfe in caret package in R

I am doing some exploratory data analysis in the Heritage Health Prize , and have come across a weird error using R's caret package. In the dataset, I've created a dataframe counting how many times a patient (by MemberID) has visited a specific primary care physician (denoted by a number) in year 1 insurance claims. Let's say there are 57 such primary care physicians, and so our dataframe looks like this:

> head(cbind(right.a[,1],right.a[,107:111]))
MemberID 2448 91972 20893 30870 13281
1     24027523    0     0     0     0     0
2     98321677    2     0     0     3     0
3     33896267    0     1     0     0     1
4      5481382    0     0     2     0     0
5     69908341    0     0     0     0    17
6      8169352    0     0     0     0     0


As you can see, this is a very sparse matrix.

Our outcome here is the number of days they spent in hospital the following year in year 2 (DaysInHospital). I'm using year 2 insurance claims and year 3 days in hospital as a validation set.

I'm using rfe to do some recursive feature selection, and I'm getting the following error:

> subsets <- seq(12,57,by = 3)
> lmProfile <- rfe(right.a[,c(107:164)],
log1p(right.a$DaysInHospital), right.b[,c(107:164)], log1p(right.b$DaysInHospital),
sizes=subsets,
metric="RMSE",
maximize="FALSE",
rfeControl=ctrl)
.
(snip)
.
External resampling iter:        10
Computing importance
Fitting subset size:   58
Computing importance
Computing importance
Error in y$pred :$ operator is invalid for atomic vectors


Am I missing something here?

• Do you mean rfeIter() (which expects train X and y and test X and y) instead of rfe()? Are right.a and right.b data.frame or matrix objects? – chl Jun 26 '11 at 21:12
• @chl Well this is embarrassing. Yes, I meant rfeIter(), although when I run it again I get a new error: Error in '[.data.frame'(x, , retained, drop = FALSE) : undefined columns selected. Both right.a and right.b are data.frames. – sylowtheorems Jun 27 '11 at 0:01
• @sylowtheorems that new error usually means your summaryFunction in trainControl is looking for a column (like class probabilities) that isn't there. Often this can be fixed by using classProbs =TRUE argument. – Zach Oct 6 '11 at 20:33