drug_health_plan.Rmd

---
title: "Drug and Health Plan Rating"
author: "Bowen Liu"
date: "March 3, 2016"
output:
  html_document: default
  pdf_document:
    number_sections: yes
  word_document: default
---
# Problem Description & Current State of Affairs
Medicare.gov released recent two years's data about "Drug and Health Plan" coverage around the country. The data inclues who, when, where, and how the plans were used. There are many plans with low ratings. And we want to find out how predictors such as political affiliation, income level, insurance rates, unemployment rates, hospital type, or happiness index contribute to the final rating of one specific type of plan. This is where we start to build the predicition modeling. It could be taken for reference for future plan upgrading or new plan release.

# Algorithms Used and Rationale
|  Algorithms  |                            Rationale                         |
|--------------|--------------------------------------------------------------|
|Logistic Regession, Naive Bayes, Decision Tree, ANN|Since the project is one supervised problem, we are going to apply multiple prediction models to the datasets. And we will choose the final one with the highest accuracy.|


# Drug and Health Plan with copay, deductible, coinsurance
- Origin: www.medicare.gov
- Data Points after preprocessing: 816 obs. of  5 variables
- variables: contract_id, mnthly_prm, ann_ddctble, copay_max, coin_max

```{r,cache=TRUE,warning=FALSE,message=FALSE}
drug_cost <- read.csv("./data/2015Med2000_flatfiles/vwPlanDrugsCostSharing.csv", stringsAsFactors = FALSE, na.strings = c("Information Not Available", ""))
# keep the copay, deductible, and coinsurance
colnames <- c("contract_id", "plan_id", "mnthly_prm", "ann_ddctbl", "copay_amt_inn_prefrd", "coins_amt_inn_prefrd")
drug_cost2 <- drug_cost[,colnames]
drug_cost3 <- drug_cost2
# split the range into standalone characters
drug_cost3$copay_low <- sapply(strsplit(drug_cost2$copay_amt_inn_prefrd, split="-"), '[', 1)
drug_cost3$copay_high <- sapply(strsplit(drug_cost2$copay_amt_inn_prefrd, split="-"), '[', 2)

drug_cost3$coin_low <- sapply(strsplit(drug_cost2$coins_amt_inn_prefrd, split="-"), '[', 1)
drug_cost3$coin_high <- sapply(strsplit(drug_cost2$coins_amt_inn_prefrd, split="-"), '[', 2)
# remove $ and %
library(gdata)
drug_cost3$copay_low <- gsub("^\\$", "", trim(drug_cost3$copay_low))
drug_cost3$copay_high <- gsub("^\\$", "", trim(drug_cost3$copay_high))
drug_cost3$coin_low <- gsub("%$", "", trim(drug_cost3$coin_low))
drug_cost3$coin_high <- gsub("%$", "", trim(drug_cost3$coin_high))
# drug_cost3 <- na.omit(drug_cost3)
#change NA to 0
drug_cost3$copay_low[is.na(drug_cost3$copay_low)] <- 0
drug_cost3$copay_high[is.na(drug_cost3$copay_high)] <- 0
drug_cost3$coin_low[is.na(drug_cost3$coin_low)] <- 0
drug_cost3$coin_high[is.na(drug_cost3$coin_high)] <- 0
# choose the maxium value
drug_cost3$copay_max <- apply(drug_cost3[, c("copay_low", "copay_high")], 1, max)
drug_cost3$coin_max <- apply(drug_cost3[, c("coin_low", "coin_low")], 1, max)
# subset the needed matrix
drug_cost4 <- drug_cost3[, c("contract_id", "mnthly_prm", "ann_ddctbl", "copay_max", "coin_max")]
# change the column type to numeric
drug_cost4$ann_ddctbl <- as.numeric(drug_cost4$ann_ddctbl)
drug_cost4$copay_max <- as.numeric(drug_cost4$copay_max)
drug_cost4$coin_max <- as.numeric(drug_cost4$coin_max)
drug_cost4$mnthly_prm <- as.numeric(drug_cost4$mnthly_prm)
# aggreate the drug_cost by contract id
drug_cost4_aggr1 <- aggregate(drug_cost4[,c("mnthly_prm", "ann_ddctbl")], by = list(drug_cost4$contract_id), FUN = mean)
drug_cost4_aggr2 <- aggregate(drug_cost4[,c("copay_max", "coin_max")], by = list(drug_cost4$contract_id), FUN = max)
drug_cost4_aggr <- merge(drug_cost4_aggr1, drug_cost4_aggr2, all = TRUE, by.x = "Group.1", by.y = "Group.1")
colnames(drug_cost4_aggr) <- c("contract_id", "mnthly_prm", "ann_ddctbl", "copay_max", "coin_max")
#write.csv(drug_cost4_aggr, file = "./data/internal_drugcost.csv")
str(drug_cost4_aggr)
tail(drug_cost4_aggr)
```
# Drug and Health Cost Tier Level
- Origin: www.medicare.gov
- Data Points after preprocessing: 585 obs. of  2 variables
- variables: contract_id, tier_label

```{r,cache=TRUE,warning=FALSE}
# load vwPlanDrugTierCost.csv 
cost_tier <- read.csv("./data/2015Med2000_flatfiles/vwPlanDrugTierCost.csv", stringsAsFactors = FALSE)
# keep contract_id, plan_id, languange_id, tier_lable
colnames <- c("contract_id", "language_id", "tier_label")
cost_tier2 <- cost_tier[, colnames]
cost_tier2 <- cost_tier2[cost_tier2$language_id == "1", ]
cost_tier3 <- cost_tier2[, c("contract_id", "tier_label")]
# count unique tier numbers for each contract id
cost_tier4 <- aggregate(data = cost_tier3,  tier_label ~ contract_id, function(x) length(unique(x)))
#write.csv(drug_cost4_aggr, file = "./data/internal_costtier.csv")
str(cost_tier4)
tail(cost_tier4)
```

# Durg and Health Plan Score Ratings
- Origin: www.medicare.gov
- Data Points after preprocessing: 1214 obs. of  3 variables
- variables: Contract_ID, Summary_Score, Star_Rating_Current

```{r,cache=TRUE,warning=FALSE}
# load rating data
ratingdata <- read.csv("./data/2015StarRatings_flatfiles/vwStarRating_SummaryScores.csv",na.strings = c("Plan too new to be measured", "Not enough data available"), stringsAsFactors = FALSE)
# remove Spanish Version
ratingdata2 <- ratingdata[ratingdata$lang_dscrptn=="English",]
# keep the columns of  "Contract_ID", "Summary_Score", "Star_Rating_Current"
colnames <- c("Contract_ID", "Summary_Score", "Star_Rating_Current")
ratingdata2 <- ratingdata2[, colnames]
# split the string to get the rating number
ratingdata2$Star_Rating_Current <- sapply(strsplit(ratingdata2$Star_Rating_Current, split=" "), '[', 1)
#ratingdata2$Star_Rating_Previous <- sapply(strsplit(ratingdata2$Star_Rating_Previous, split=" "), '[', 1)
# remove NAs of the current 2015 year
ratingdata2_nocurrna <- ratingdata2[(!is.na(ratingdata2$Star_Rating_Current)), ]
# subset the rating data
ratingdata3 <- ratingdata2_nocurrna[ratingdata2_nocurrna$Summary_Score=="Overall Star Rating", 1:3]
#write.csv(ratingdata3, file = "./data/ratingdata.csv")
str(ratingdata2_nocurrna)
tail(ratingdata2_nocurrna)
#head(ratingdata3)
```

# Intermediate Data to Connect Contract and Zipcode
- Origin: www.medicare.gov
- Data Points after preprocessing: 364,847 obs. of  2 variables
- variables: Zip_Code, Contract_ID

```{r,cache=TRUE,warning=FALSE}
zip_contract <- read.csv("./data/2015Med2000_flatfiles/vwLocalContractServiceAreas.csv", stringsAsFactors = FALSE)
zip_contract2 <- unique(zip_contract[, c("Zip_Code", "Contract_ID")])
#write.csv(zip_contract2, file = "./data/internal_geography.csv")
str(zip_contract2)
tail(zip_contract2)
```

# Internal Features Combination: by contract_id
- Data Points after preprocessing: 227,432 obs. of  9 variables
- variables: contract_id, tier_label, mnthly_prm, ann_ddctbl, copay_max, coin_max, Summary_Score, Star_Rating_Current, Zip_Code

```{r,cache=TRUE,warning=FALSE}
cost <- merge(cost_tier4, drug_cost4_aggr, all = TRUE, by = "contract_id")
cost_rating <- merge(cost, ratingdata3, all = TRUE, by.x = "contract_id", by.y = "Contract_ID")
cost_rating_zip <- merge(cost_rating, zip_contract2, all = TRUE, by.x = "contract_id", by.y = "Contract_ID")
cost_rating_zip2 <- na.omit(cost_rating_zip)
str(cost_rating_zip2)
tail(cost_rating_zip)
#write.csv(cost_rating_zip2, file = "./data/internal_data.csv")
```

# Unemployment Rate
- Origin: http://zipatlas.com
- Data Points after preprocessing: 29,252 obs. of  3 variables
- variables: zipcode, polulation, unemployment_rate

```{r,cache=TRUE,warning=FALSE}
unemployment_data <- read.table("./data/unemployment.txt", stringsAsFactors = FALSE)
unemployment_data2 <- unemployment_data[, c(2, 6, 7)]
colnames(unemployment_data2) <- c("zipcode", "polulation", "unemployment_rate")
unemployment_data2$polulation <- as.numeric(gsub("[[:punct:]]", "", unemployment_data2$polulation))
unemployment_data2$unemployment_rate <- as.numeric(unemployment_data2$unemployment_rate)
#write.csv(cost_rating_zip2, file = "./data/external_unemploy.csv")
str(unemployment_data2)
tail(unemployment_data2)
```
# Total Income
- Origin: https://www.irs.gov/uac/SOI-Tax-Stats-Individual-Income-Tax-Statistics-2013-ZIP-Code-Data-(SOI)
- Data Points after preprocessing: 27,790 obs. of  2 variables
- variables: ZIPCODE, total_income

```{r,cache=TRUE,warning=FALSE}
income_data <- read.csv("./data/zipcode2013/zipcodenoagi13.csv", stringsAsFactors = FALSE)
income_data2 <- income_data[, c("ZIPCODE", "A02650")]
colnames(income_data2)[2] <- "total_income"
#write.csv(income_data2, file = "./data/external_income.csv")
str(income_data2)
tail(income_data2)
```

# Physician
- Origin: https://www.irs.gov/uac/SOI-Tax-Stats-Individual-Income-Tax-Statistics-2013-ZIP-Code-Data-(SOI)
- author: Yang Gu
- Data Points after preprocessing: 41,184. of  4 variables
- variables: count, experience, spec_count, zip_code

```{r,cache=TRUE,warning=FALSE,message=FALSE}
# creates number of doctors, average years of experience, and number of specializations BY COUNTY
# indexed by each zip code (e.g. 5 zip codes in one county would all have same info)
#read physician data set, select subset, remove non-complete rows
doctors <- read.csv('./data/physician.csv', stringsAsFactors = FALSE)
str(doctors)
myVars <- c("NPI", "Graduation.year","Primary.specialty","City", "State", "Zip.Code")
doctors <- doctors[myVars]
doctors <- doctors[complete.cases(doctors),]
str(doctors)
head(doctors)
#deal with zip code format 
doctors$Zip.Code<- as.numeric(doctors$Zip.Code)
doctors$set<-ifelse(doctors$Zip.Code > 99999,1, 2)

#separate data by 9-digit and 5-digit formats
data <- split(doctors, doctors$set)
temp1<-data[[1]] #data with 9-digit zip Codes

#add leading zeros so all data has 9 digits: 1234567 --> 001234567
# take first 5 digits, then remove 0s: 00123 --> 123
library(stringr)
temp1$Zip.Code <-str_pad(temp1$Zip.Code, 9, pad = "0")  
temp1$Zip.Code <-substr(temp1$Zip.Code,0,5)
temp1$Zip.Code <-substr(temp1$Zip.Code,regexpr("[^0]",temp1$Zip.Code),nchar(temp1$Zip.Code))

#recombine dataset separated by zip code
doctors <- rbind(temp1, data[[2]]) 
doctors$set <- NULL #remove dummy variable

#data cleaning & renaming 
doctors[2]<-2016 - doctors$Graduation.year
names(doctors)[2] <- "experience"
names(doctors)[6] <- "zip_code"
head(doctors)
#combine geography information. Fill in zips w/o doctors w/ 0
geography <-read.csv('./data/geographyGood.csv', stringsAsFactors = FALSE)
#myVars <- c("CountyFIPSCode","zip_code")
#geography <- geography[myVars]
names(geography)[2] <- "CountyFIPSCode"

doctors <- merge(x=doctors, y=geography, by="zip_code")
doctors[is.na(doctors)] <- 0

doctors<- unique(doctors)

library(plyr)
#create dataset for number of doctors in each county
myVars <- c("NPI","CountyFIPSCode")
docNum <- doctors[myVars]
docNum <- ddply(docNum, .(CountyFIPSCode), mutate, count = length(unique(NPI)))
docNum$NPI<- NULL
docNum <- unique((docNum))
docNum <- docNum[complete.cases(docNum),]

#create dataset for average experience of doctors in each county
myVars <- c("experience","CountyFIPSCode")
yearDoc <- doctors[myVars]
yearDoc<- aggregate(yearDoc,by=list(yearDoc$CountyFIPSCode), FUN=mean, na.rm=TRUE)
yearDoc$Group.1 <-NULL

#join on county code
docNum <- merge(x=docNum, y=yearDoc, by="CountyFIPSCode", all.x = TRUE )

#create dataset for number of specialties in each county
myVars <- c("Primary.specialty","CountyFIPSCode")
docSpec <- doctors[myVars]
docSpec <- ddply(docSpec, .(CountyFIPSCode), mutate, count = length(unique(Primary.specialty)))
docSpec$Primary.specialty <- NULL
docSpec <- unique((docSpec))
names(docSpec)[2]<- "spec_count"

docNum <- merge(x=docNum, y=docSpec, by="CountyFIPSCode", all.x = TRUE )
#docNum$count<-ifelse(docNum$experience == 0,0, docNum$count)
#docNum$spec_count<-ifelse(docNum$experience == 0,0, docNum$spec_count)

geography <- merge(x = docNum, y = geography, all.y = TRUE)
geography[is.na(geography)] <- 0

docNum <- geography
docNum$StateCode <- docNum$StateName <- docNum$CountyName <- NULL

rm(docSpec, doctors, yearDoc, data)

temp <- docNum$zip_code
temp <- unique(temp)

n_occur <- data.frame(table(docNum$zip_code))
n_occur[n_occur$Freq > 1,]
dupes <- docNum[docNum$zip_code %in% n_occur$Var1[n_occur$Freq > 1],]

#write.csv(docNum, file = "./data/docByZip2.csv",row.names=FALSE)
physician <- read.csv(file = "./data/docByZip2.csv", stringsAsFactors = FALSE)
str(physician)
```

# External Features Combination: by zip code
- Origin: 
- Data Points after preprocessing: 36,755 obs. of  6 variables
- variables: zip_code, experience, spec_count, unemployment_rate, average_income, avgerage_physician

```{r,cache=TRUE,warning=FALSE}
unemploy_Income <- merge(unemployment_data2, income_data2, all = TRUE, by.x = "zipcode", by.y = "ZIPCODE")
unemploy_Income_0na <- na.omit(unemploy_Income)
unemploy_Income_0na$average_income <- unemploy_Income_0na$total_income / unemploy_Income_0na$polulation
phy_unemploy_income <- merge(physician, unemploy_Income_0na, all = TRUE, by.x = "zip_code", by.y = "zipcode")
phy_unemploy_income_0na <- na.omit(phy_unemploy_income)
phy_unemploy_income_0na$average_physician = round(phy_unemploy_income_0na$polulation / (phy_unemploy_income_0na$count + 1))
str(phy_unemploy_income_0na)
#write.csv(phy_unemploy_income_0na, file = "./external_data_v2.csv")
```

# Combine Internal and External Data
- Data Points after preprocessing: 187,488 obs. of  14 variables
- variables: zip_code, experience, spec_count, unemployment_rate, average_income, avgerage_physician, contract_id, tier_label, mnthly_prm, ann_ddctbl, copay_max, coin_max, Star_Rating_Current

```{r,cache=TRUE,warning=FALSE}
final_data <- merge(phy_unemploy_income_0na, cost_rating_zip2, all = TRUE, by.x = "zip_code", by.y = "Zip_Code")
final_data <- na.omit(final_data)
colnames <- c('zip_code', 'experience', 'spec_count', 'unemployment_rate', 'average_income', 'average_physician', 'contract_id', 'tier_label', 'mnthly_prm', 'ann_ddctbl', 'copay_max', 'coin_max', 'Star_Rating_Current')
final_data2 <- final_data[, colnames]
str(final_data2)
tail(final_data2)
#write.csv(final_data2, file = "./data/final_data_v2.csv", row.names = FALSE)
```

# Combine Geographic Info
- Origin: 
- Data Points after preprocessing: 132,083 obs. of  18 variables
- variables:

```{r,cache=TRUE,warning=FALSE}
geo <- read.csv(file = "./data/zip_codes_states.csv", stringsAsFactors = FALSE)
final_data3 <- merge(final_data2, geo, all.x = TRUE, by.x = "zip_code", by.y = "zip_code")
str(final_data3)
#final_data3 <- unique(final_data3)
#write.csv(final_data3, file = "./data/final_data_v3.csv", row.names = FALSE)
```

# Exploratory Plotting
```{r,cache=TRUE}
library(ggplot2)
library(gridExtra)
summary(final_data3)

plotdata <- final_data3
# external feature plotting
ggplot(data = plotdata, aes(factor(Star_Rating_Current), experience)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Average Experience Years") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), spec_count)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Specialty Counts") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))
    
ggplot(data = plotdata, aes(factor(Star_Rating_Current), unemployment_rate)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Unemployment Rate(%)") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), average_income)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Average Income(,000$)") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), average_physician)) +
    #theme(legend.position="bottom", legend.title=element_blank()) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Average Citizens per Physician") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

# external feature plotting
ggplot(data = plotdata, aes(factor(Star_Rating_Current), tier_label)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Tiers") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), mnthly_prm)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Monthly Premium") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), ann_ddctbl)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Annualy Deductible") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), copay_max)) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Maxium Copayment") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current), coin_max)) +
    #theme(legend.position="bottom", legend.title=element_blank()) +
    guides(fill=FALSE) +
    labs(x="Rating Scores", y="Maxium Coinsurance") +
    geom_boxplot(aes(fill = factor(Star_Rating_Current)))

ggplot(data = plotdata, aes(factor(Star_Rating_Current))) +
    guides(fill=FALSE) +
    geom_bar(aes(fill = factor(Star_Rating_Current)))
# explore the clustering of
```

# Split the dataset for modeling
```{r,cache=TRUE,warning=FALSE}
final_data3 <- read.csv(file = './data/final_data_v3.csv')
# remove neutral 4 points, and change the label into 1 and 0
final_data3_rm4 <- final_data3[final_data3$Star_Rating_Current != 4, ]
# transfer the label into binary value
final_data3_rm4$sentiment <- as.factor(ifelse(final_data3_rm4$Star_Rating_Current > 4,1,0))
# subset the need columns for modeling
colnames <- c('experience', 'spec_count', 'unemployment_rate', 'average_income', 'average_physician', 'tier_label', 'mnthly_prm', 'ann_ddctbl', 'copay_max', 'coin_max', 'sentiment')
final_data3_rm4 <- final_data3_rm4[, colnames]
str(final_data3_rm4)
# subset the train and test data sets
set.seed(510)
sample_size <- floor(0.8 * nrow(final_data3_rm4))
training_index <- sample(seq_len(nrow(final_data3_rm4)), size = sample_size)
train <- final_data3_rm4[training_index,]
test <- final_data3_rm4[-training_index,]
```

# Logistic Regression

```{r,cache=TRUE,warning=FALSE}
# train the model
logsitic_regression_model <- glm(formula = sentiment ~ ., family = binomial(link = 'logit'), data = train)
summary(logsitic_regression_model)
# evaluate the model
library(caret)
lr_pred <- predict(logsitic_regression_model, newdata = test, type = "response")
lr_pred <- ifelse(lr_pred >= 0.5, 1, 0)
# evaluation <- cbind(test, pred)
# evaluation$correct <- ifelse(evaluation$sentiment == evaluation$pred,1,0)
# sum(evaluation$correct) / nrow(evaluation)
confusionMatrix(data = lr_pred, test[, 'sentiment'])
```

# Neural Network

```{r,cache=TRUE,warning=FALSE}
library (nnet)
# train the model
neural_network_model <- nnet(formula =  sentiment ~ ., data =train, decay = 0.5, size = 6)
summary(neural_network_model)
# evaluate the model
nn_pred <- predict(neural_network_model, newdata = test)
nn_pred <- ifelse(nn_pred >= 0.5, 1, 0)
# confusion matrix
confusionMatrix(data = nn_pred, test[, 'sentiment'])
```


```{r,cache=TRUE,warning=FALSE,include=FALSE}
# Decision Tree
library(C50)
# train the model
decision_tree_model <- C5.0.default(x = train[, -11], y = train$sentiment)
dt_pred <- predict(decision_tree_model, newdata = test[,colnames])
summary(decision_tree_model)
# evaluate the model
# confusion matrix
confusionMatrix(data = dt_pred, test[, 'sentiment'])
```