predict.lm throws error when dataframe is subset - r

Question

I am trying to use the caret::train() function to create a linear model with leave-one-out cross-validation from a data frame with multiple response variables. Some of the response variables I want to log transform. Some of the other response variables have NA variables. I am getting the following error:
Error in seq_len(p) : argument must be coercible to non-negative integer
In addition: Warning messages:
1: In predict.lm(trainlm, newdata = df2, type = "response") :
calling predict.lm(<fake-lm-object>) ...
2: In seq_len(p) : first element used of 'length.out' argument
Looking through other posts, It seemed like this arose either because:
I subset the dataframe
I had NA values
I tried to remedy this by first creating a new dataframe with the appropriate columns and selecting rows with complete.cases(), but the problem persists. Below is my reproducible example:
library(caret) # for train() function
set.seed(52) # to make reproducible
##Creating Fake Dataset
X1<-runif(100, 2, 21)
X2<-runif(100, 21, 40)
X3<-runif(100, 12, 18)
errors1<-rnorm(100, 0, 1)
errors2<-rnorm(100, 0, 1)
#multiple response variables
Y1<-2.31+(0.52*X1)+(0.84*X2)+(2.2*X3)+(1.5*X1*X2)+(1.6*errors1)
Y2<-5.31+(2.1*X1)+(2.2*X3)+(1.5*X1*X3)+(0.4*errors2)
##Creating an NA Value
Y2[82]<-NA
##Dataframe with all predictors and both response variables
df<-data.frame(Y1, Y2, X1, X2, X3)
##Subsetting to get rid of NA and other
df2<-subset(df[complete.cases(df),], select=-1)
##Building the model
TrCtrl<-trainControl(method="LOOCV")
trainlm<-train(log(Y2+1)~X1+X2+X3+(X1+X2)+(X1*X3)+(X2*X3)+(X1*X2*X3), method="lmStepAIC", data=df2, trControl=TrCtrl)
##Getting Prediction##
Train.Predict<-predict.lm(trainlm, newdata = df2, type = "response")

Answer 1

trainlm isn't an lm class so predict.lm isn't the right function to call.
class(trainlm)
#> [1] "train" "train.formula"
Use predict and let S3 choose the appropriate method.
Train.Predict <- predict(trainlm, newdata = df2)