Regression Tree using random samples

regress_with_random.R

# Transform days spent to log(1 + x)
y1_data$DaysInHospital_Y2 <- log1p(y1_data$DaysInHospital_Y2)

# Divide the set, 80% for train and 20% for test
indexes <- sample(1:nrow(y1_data), size=0.2*nrow(y1_data))
test <- y1_data[indexes,]
train <- y1_data[-indexes,]

# Remove unwanted features from both the sets
train <- subset(train, select=-c(MemberID_t, YEAR_t, DaysInHospital, trainset,  DaysInHospital_Y3, age_05, PayDelay_max, PayDelay_min, PayDelay_stdev, LOS_max, LOS_min,
  LOS_stdev, LOS_TOT_UNKNOWN, LOS_TOT_SUPRESSED, LOS_TOT_KNOWN, dsfs_max, dsfs_min, dsfs_range, dsfs_stdev, CharlsonIndexI_max, CharlsonIndexI_min, CharlsonIndexI_range, CharlsonIndexI_stdev,drugCount_max, drugCount_min, memberID_lc, YEAR_lc, labCount_max, labCount_min, labNull, drugNull))

test <- subset(test, select=-c(MemberID_t, YEAR_t, DaysInHospital, trainset,  DaysInHospital_Y2, DaysInHospital_Y3, age_05, PayDelay_max, PayDelay_min, PayDelay_stdev, LOS_max, LOS_min,
  LOS_stdev, LOS_TOT_UNKNOWN, LOS_TOT_SUPRESSED, LOS_TOT_KNOWN, dsfs_max, dsfs_min, dsfs_range, dsfs_stdev, CharlsonIndexI_max, CharlsonIndexI_min, CharlsonIndexI_range, CharlsonIndexI_stdev,drugCount_max, drugCount_min, memberID_lc, YEAR_lc, labCount_max, labCount_min, labNull, drugNull))

# Model a regression tree using the training data
tr <- tree(DaysInHospital_Y2 ~ ., train)

# Get the visual for the tree
plot(tr,type="uniform"); text(tr,pretty=0)

# Use the tree to predict the test set values and 
# append the predicted values to the test set
result <- predict(tr, test, type="vector")
test$predicted <- result

# Calculate the RMSE without truncation
sqrt(mean((y1_data[indexes,]$DaysInHospital_Y2 - test$predicted)^2))

# Truncate the values and then calculate the RMSE again
test$new_predicted <- log1p(trunc(exp(test$predicted) - 1))
sqrt(mean((y1_data[indexes,]$DaysInHospital_Y2 - test$new_predicted)^2))