five-fold cross-validation with the use of linear regression - r

Question

I would like to perform a five-fold cross validation for a regression model of degree 1
lm(y ~ poly(x, degree=1), data).
I generated 100 observations with the following code
set.seed(1)
GenData <- function(n){
x <- seq(-2,2,length.out=n)
y <- -4 - 3*x + 1.5*x^2 + 2*x^3 + rnorm(n,0,0.5)
return(cbind(x,y))
}
GenData(100)
D<-GenData(100)
and my code for this goal is
ind<-sample(1:100)
re<-NULL
k<-20
teams<- 5
t<-NULL
for (i in 1:teams) {
te<- ind[ ((i-1)*k+1):(i*k)]
train <- D[-te,1:2]
test <- D[te,1:2]
cl <- D[-te,2]
lm1 <- lm(cl ~train[,1] , data=train)
pred <- predict(lm1,test)
t<- c(t, sum(D[te,2] == pred) /dim(test)[1])
}
re<-c(re,mean(t))
where I split my data into training and test.With the training data I run a regression with purpose to make a prediction and comperate it with my test data.But I have the following error
"Error in predict(mult, test)$class :
$ operator is invalid for atomic vectors
In addition: Warning message:
'newdata' had 20 rows but variables found have 80 rows "
So I understand that I have to change something on the line
pred<-predict(lm1,test)
but I dont know what .
Thanks in advance!

Answer 1

lm requires a data frame as input data. Also trying to validate the model by just verifying if the result matches the expected value will not work. You are simulating the irreducible error using normal error.
Here is the updated code:
ind<-sample(1:100)
re<-NULL
k<-20
teams<- 5
t<-NULL
for (i in 1:teams) {
te<- ind[ ((i-1)*k+1):(i*k)]
train <- data.frame(D[-te,1:2])
test <- data.frame(D[te,1:2])
lm1 <- lm(y~x , data=train)
pred <- predict(lm1,test)
t<- c(t, sum(abs(D[te,2] - pred)) /dim(test)[1])
}
re<-c(re,mean(t))

Answer 2

In the lm() function, your y variable is cl, a vector not included in the data = argument:
cl <- D[-te,2]
lm1 <- lm(cl ~train[,1] , data=train)
No need to include the cl at all. Rather, simply specify x and y by their names in the dataset train, in this case the names are x and y:
names(train)
[1] "x" "y"
So your for loop would then look like:
for (i in 1:teams) {
te<- ind[ ((i-1)*k+1):(i*k)]
train <- D[-te,1:2]
test <- D[te,1:2]
lm1 <- lm(y ~x , data=train)
pred <- predict(lm1,test)
t[i]<- sum(D[te,2] == pred)/dim(test)[1]
}
Also, note that I have added the for loop index i so that values can be added to the object. Lastly, I had to make the D object a dataframe in order for the code to work:
D<-as.data.frame(GenData(100))
Your re object ends up being 0 because your model does not predict any numbers correctly. I would suggest using RMSE as a performance measure for continuous data.