CS112 14.1 lalonde breakout

data(lalonde); library(randomForest)

# create lalonde2 (just control units); delete orig data to avoid mistakes
lalonde2 <- lalonde[which(lalonde$treat == 0),]; rm(lalonde)

# eliminate the treatment indicator variable (they are all control units)
# remove additional columns--we are going to predict "u75" (unemployed in '75)
elimin.cols <- which( names(lalonde2) == "treat" |  
                      names(lalonde2) == "re75"| 
                      names(lalonde2) == "re78")

lalonde2 <- lalonde2[, -elimin.cols]

# make the dependent variable (what we are trying to predict)
# a factor, because random forest will perform classification
# for factors, not for 'numeric' variables.
# notice that I set the levels when I define this factor...
# otherwise, by default, R would set the first factor value
# of the first value appearing in the dataset as "0"
# even though it is actually a "1"--this could be very confusing.
lalonde2$u75 <- factor(lalonde2$u75, levels = c(1,0))

# data has 260 rows and 9 columns
dim(lalonde2)

# set the random seedset.seed(1234)

# create a test set and training set 
test <- sample(1:nrow(lalonde2), 60)
train <- c(1:260)[-test]
lalonde2.train <- lalonde2[train,]
lalonde2.test <- lalonde2[test,]

cat("There are", length(which(lalonde2.train$u75 == 1)),
    "unemployed observations in the TRAINING set.")

## create a bagged forest, allowing it to try all 8 features
bag.lalonde2 <- randomForest( u75~., data=lalonde2, subset=train, 
                              mtry=8, ntree = 5000, importance =TRUE)

# how well does it perform on the training set?
# what is the OOB test error?
print(bag.lalonde2)

# how well does this perform on the test set?
cat("There are", length(which(lalonde2.test$u75 == 1)),
    "unemployed observations in the TEST set.")

yhat.bag <- predict(bag.lalonde2, newdata=lalonde2[test,])
table(yhat.bag, lalonde2.test$u75)

## check the importance of the variables
importance(bag.lalonde2)
varImpPlot(bag.lalonde2)

# # # # # #
# Now re-run for a random forest by setting the# mtry parameter lower--
# which 'mtry' minimizes training OOB error rate?