CODEWITHSUNDEEP
rregressionpredict
jeandut
jeandut

Reputation: 2524

Predict with linearRidge : Error in as.matrix(mm) %*% beta : non-conformable arguments

I am using the package ridge to do ridge regression Let's say that I am using the mtcars data and that I want to predict the variable qsec so I do: 

install.packages('ridge')
library(ridge)
library(stats)
library(dplyr) #for the select
data(mtcars)
index=1:floor(0.75*nrow(mtcars))
train=mtcars[index,]
test=select(mtcars[-index,],-qsec)
ridge.model<-linearRidge(qsec~.,data=train)

That seems to work just fine then I can try to predict the next values using predict function.

qsec.pred<-predict(ridge.model,test)

That works perfectly fine again but when I try to do it with a test of only one row:

qsec.pred.first.row<-predict(ridge.model,test[1,])

I get the following error : 

Error in as.matrix(mm) %*% beta : non-conformable arguments

There seems to be a problem with handling data when there is only one line. Do you know a way to fix that ? Or did I do something wrong ?

Upvotes: 1

Views: 727

Answers (1)

MrFlick
MrFlick

Reputation: 206197

There does appear to be a bug in the ridge:::predict.ridgeLinear code. Specifically, when they subset their model matrix, with only one row, they are loosing the proper structure of the matrix.

You can fix it by writing your own version

predict.ridgeLinear <- ridge:::predict.ridgeLinear
body(predict.ridgeLinear)[[7]][[3]] <- quote(mm <- cbind(1, X[, ll, drop=FALSE]))
body(predict.ridgeLinear)[[7]][[4]] <- quote(mm <- X[, ll, drop=FALSE])
environment(predict.ridgeLinear) <- asNamespace("ridge")

This is essentially hacking specific line numbers so it's not very robust. I've only tested in on version ridge_2.1-3. But, after we define this function, we can call

predict(ridge.model,test[1,])
# Pontiac Firebird 
#         17.65554 
predict(ridge.model,test[1:2,])
# Pontiac Firebird        Fiat X1-9 
#         17.65554         19.34306

You may wish to contact the package maintainer if you think this fix should be permanently included in the package. Feel free to refer to this answer.

The other work around would be to always predict more than one row at a time. Even if it's the same row twice.

qsec.pred.first.row <- predict(ridge.model,test[c(1,1),])

Upvotes: 1

Related Questions