Federal Spending and Congressional Representation
Project Information
Charting federal spending alongside congressional representation comprised my first project for the Fall 2015 NYC Data Science Academy Data Science Bootcamp, as presented on October 3, 2015. (Check out the video below for a recording of my presentation.)
The pedagogical purpose of this project was to practice using dplyr for data munging and ggplot2 for visualizations. In addition, I practiced using the XML and stringr packages to scrape and parse semi-structured text data into a form appropriate for analysis.
Introduction
This project creates some visualizations that relate federal government contract spending to the status of Congressional members representing those districts in which the contract work is performed. The key exhibits are plots that I like to think of as visual contingency tables. In particular, these visualizations juxtapose Representatives’ seniority with the extent to which members serve on committees related to the department or agency issuing a contract or contracts to their districts. I look at Defense spending, for instance, in terms of membership on the Armed Services Committee. I also identify members of the Appropriations Committee: this is the committee responsible for the spending side the federal budget, and none of its members serve on any of the committees related to specific policy areas such as Armed Services or Natural Resources.
Gathering Data
All contract data are from FY 2015 and information on Congress is for the current (114th) Congress.
I realized rather late in the process that the timing of my data sources is a bit out of line, because the FY 2015 budget was approved in Fall 2014 while the 114th Congress began in January 2015. For the most part, I don’t think is a huge problem, because committee assignments are fairly stable from one term to another. There are 67 first-term Representatives in the 114th Congress, and clearly they were not in office when the FY 2015 budget was approved, but isolating this group in the analysis actually leads to an interesting finding.
I downloaded the contract data manually from USAspending.gov. I limited my downloads to the 24 agencies listed under “CFO” in the “Agency” drop-down menu. Because of the volume of the Defense Department data, I had to download data for each individual state one-by-one. This resulted in a total of 73 files containing 4.9 gbs of data.
In order to create a manageable file to use in the project, I've used the following function to open each downloaded file, extract desired fields, and place the pared-down data in a single file.
library(dplyr)
store.compact <- function(infilename, outfilename){
temp <- read.table(infilename, header = TRUE, sep = ",", stringsAsFactors = FALSE)
temp <- select(temp,
unique_transaction_id,
dollarsobligated,
maj_agency_cat,
vendorcountrycode,
state,
congressionaldistrict,
placeofperformancecongressionaldistrict,
placeofperformancezipcode)
write.table(temp, outfilename,
append = TRUE,
row.names = FALSE,
col.names = FALSE,
sep = ",")
rm(temp)
}
The following script calls the function. Here are the first several calls.
library(stringr)
library(dplyr)
source('~/nycdsa/proj_1_bootcamp/StoreCompactData.R')
system("rm ContractSpending.csv")
store.compact("contract_data/Data_Feed.csv","ContractSpending.csv")
store.compact("contract_data/Data_Feed-2.csv","ContractSpending.csv")
store.compact("contract_data/Data_Feed-3.csv","ContractSpending.csv")
store.compact("contract_data/Data_Feed-4.csv","ContractSpending.csv")
store.compact("contract_data/Data_Feed-5.csv","ContractSpending.csv")
store.compact("contract_data/Data_Feed-6.csv","ContractSpending.csv")
...
temp <- read.csv("ContractSpending.csv", header = FALSE, stringsAsFactors = FALSE)
saveRDS(temp,"contracts_raw.rds")
The following code downloads Congressional Committee data from clerk.house.gov and seniority data from pressgallery.house.gov.
library(stringr)
library(XML)
library(dplyr,warn.conflicts=FALSE)
url <- "http://clerk.house.gov/committee_info/oal.aspx"
RepData <- readHTMLTable(url, header = TRUE, stringsAsFactors = FALSE) %>%
data.frame() %>% tbl_df()
colnames(RepData) <- c("Member", "Committees")
url <- "http://pressgallery.house.gov/member-data/seniority"
SeniorData <- readHTMLTable(url, header = TRUE, stringsAsFactors = FALSE)[[1]] %>%
select(., 2:4)
colnames(SeniorData) <- c('Member','PartyAndState','Terms')
Processing Data
1. Mapping Agency Codes
A first step in preparing the contract data for analysis is to map the 4-digit codes onto agency names. These names include both the long agency names from the source files and shorter, more convenient names created with the following function.
Short.Agency.Names <- function(code){
if(code == 1200) return('Agriculture')
else if(code == 1300) return('Commerce')
else if(code == 1400) return('Interior')
else if(code == 1500) return('Justice')
else if(code == 1600) return('Labor')
else if(code == 1900) return('State')
else if(code == 2000) return('Treasury')
else if(code == 2400) return('Personnel Management')
else if(code == 2800) return('Social Security')
else if(code == 3100) return('Nuclear Reg Comm')
else if(code == 3600) return('Veterans Affairs')
else if(code == 4700) return('GSA')
else if(code == 4900) return('NSF')
else if(code == 6800) return('EPA')
else if(code == 6900) return('Transportation')
else if(code == 7000) return('Homeland Security')
else if(code == 7200) return('USAID')
else if(code == 7300) return('Small Business Admin')
else if(code == 7500) return('HHS')
else if(code == 8000) return('NASA')
else if(code == 8600) return('HUD')
else if(code == 8900) return('Energy')
else if(code == 9100) return('Education')
else if(code == 9700) return('Defense')
}
The following code creates the actual mapping.
cstest <- readRDS("contracts_raw.rds")
# Create a mapping between agency codes and agency names, as implied by the contract data
# Each name appears twice in the raw files, so we pick the version that includes lowercase letters
agency_codes <- unique(cstest$V3)
lower <- agency_codes[str_detect(agency_codes,'[:lower:]+')]
lower <- array(lower)
agency_codes <- data.frame(id = str_sub(lower,1,4), Agency = str_sub(lower,7))
for(code in agency_codes$id){
agency_codes$ShortName[agency_codes$id==code] <- Short.Agency.Names(code)
}
agency_codes$id <- as.character(agency_codes$id)
saveRDS(agency_codes,"agency_codes.rds")
rm(lower, agency_codes, code, Short.Agency.Names)
2. Cleaning contract data
Cleaning the contract data entails isolating agency codes into a separate field, adding field names (when putting data into a single file they have to be deleted), and eliminating obviously incorrect districts.
# Various cleaning tasks for contract data
# Create new field containing agency code
cstest <- mutate(cstest, Agency.Code = str_sub(V3,1,4))
# Add field names that were omitted during storage
colnames(cstest) <- c('ID', 'Dollars.Obligated', 'oldagnc', 'country','State',
'Congress.District', 'pop_cd','pop_zip','Agency.Code')
# Filter out non-state juridictions, non-existent district codes
cstest <- filter(cstest, !(str_sub(pop_cd,1,2) %in% c('PR','VI','GU','DC','AS','MP'))) %>%
filter(., !(pop_cd == 'CA00' | pop_cd == 'WV00'))
# Select fields needed downstream, and shorten zip to 5 digits (as used in county mapping)
cstest <-select(cstest, ID, Agency.Code, Dollars.Obligated, State, pop_cd, pop_zip)
cstest$Agency.Code <- as.character(cstest$Agency.Code )
cstest$pop_zip <- sapply(cstest$pop_zip, function(x) str_sub(x,1,5))
saveRDS(cstest,"contracts.rds")
rm(cstest)
3. Parsing data on members of congress
The next processing task is to parse the data on committee membership. A key task was producing codes for congressional districts that match those used in the contract data. Additionally, extracting members’ names will be crucial for joining the two data sources later.
# Get last name (required later for merge)
get.last.name <- function(member_string){
temp <- str_split(member_string, ",")
return(temp[[1]][1])
}
# Get first name (required later for merge)
get.first.name <- function(member_string){
temp <- str_split(member_string, ",")
temp <- temp[[1]][2]
temp <- str_split(str_sub(temp, start = 2)," ")
return(temp[[1]][1])
return(temp)
}
# Get state abbreviation
get.state <- function(member_string){
temp <- str_split(member_string, " ")
return(temp[[1]][length(temp[[1]])])
}
# Get district number (0 if from a single-district state such as Delaware)
get.district <- function(member_string){
temp <- str_split(member_string, " ")
text <- temp[[1]][length(temp[[1]])-1]
if(str_detect(text,'[:digit:]+')){
return(str_extract(text,'[:digit:]+'))
}else{
return('0')
}
}
# Combine state and district number into a 4-character code, e.g. NJ11
get.code <- function(state, district){
if(str_length(district) == 1)
return(str_c(state,'0',district))
else
return(str_c(state,district))
}
RepData$FirstName <- sapply(RepData$Member, get.first.name)
RepData$LastName <- sapply(RepData$Member, get.last.name)
RepData$State <- sapply(RepData$Member, get.state)
RepData$District <- sapply(RepData$Member, get.district)
RepData$DistrictCode <- mapply(get.code, RepData$State, RepData$District) %>%
as.character(.)
RepData <- filter(RepData, !(State %in% c('PR','VI','GU','DC','AS','MP'))) #Filter out non-voting members
The seniority data requires different functions. Note that there are no district numbers in this table.
get.state.senior <- function(ps){
temp <- str_split(ps,",")
return(str_sub(temp[[1]][2],2,3))
}
get.last.name.senior <- function(memb){
temp <- str_split(memb," ")[[1]][-1]
temp <- ifelse(str_detect(temp,'\\*'),str_sub(temp,1,-2),temp)
return(temp[[1]])
}
get.first.name.senior <- function(memb){
temp <- str_split(memb," ")[[1]][1]
temp <- ifelse(str_detect(temp,'\\*'),str_sub(temp,1,-2),temp)
return(temp[[1]])
}
SeniorData$FirstName = sapply(SeniorData$Member,get.first.name.senior)
SeniorData$LastName = sapply(SeniorData$Member,get.last.name.senior)
SeniorData$State = sapply(SeniorData$PartyAndState,get.state.senior)
SeniorData <- select(SeniorData, FirstName, LastName, State, Terms)
4. Merging committee data and seniority data
Having parsed committee membership data, the next challenge is to merge committee data with seniority data. Since there's no district data in the seniority table, we need instead to merge by first name, last name, and state.
LeftData <- left_join(RepData, SeniorData, by = c('State','FirstName','LastName'))
Due to features like diacritical marks, abbreviations, and nicknames, the first name/last name/state merge didn't succeed in every case. Since most of the problems seem to be with first names and since matching last names within a state are not common, a good tactic is to take rows that failed to match on the first attempt, attempt a join operation via last name and state, and populate the “Terms” field in the “LeftData” data frame with successful matches.
MissingLeft <- filter(LeftData,is.na(Terms))
FixerByLast <- left_join(MissingLeft,SeniorData, by = c('State','LastName')) %>%
select(.,DistrictCode,Terms.y) %>%
filter(.,!is.na(Terms.y))
for(d in FixerByLast$DistrictCode){
LeftData$Terms[LeftData$DistrictCode == d] <- FixerByLast$Terms.y[FixerByLast$DistrictCode == d]
}
The same process can be tweaked for the 17 remaining unmatched rows, this time joining on first name and state. (The additional manual correction below is necessary because there are two representatives from NJ named “Donald.”)
MissingLeft <- filter(LeftData,is.na(Terms))
FixerByFirst <- left_join(MissingLeft,SeniorData, by = c('State','FirstName')) %>%
select(.,DistrictCode,Terms.y) %>%
filter(.,!is.na(Terms.y)) %>%
filter(., !(DistrictCode=='NJ10' & Terms.y !=3))
for(d in FixerByFirst$DistrictCode){
LeftData$Terms[LeftData$DistrictCode == d] <- FixerByFirst$Terms.y[FixerByFirst$DistrictCode == d]
}
Having populated “Terms” for the final two unmatched rows manually, all “Terms” can be converted to integers.
LeftData$Terms[LeftData$DistrictCode == 'MO01'] <- 8
LeftData$Terms[LeftData$DistrictCode == 'WA03'] <- 3
LeftData$Terms <- as.integer(LeftData$Terms)
The seniority data in LeftData can now be merged with RepData. We'll additionally want to add a categorical field based on terms. Because of the timing discrepancy noted in the "Processing Data" section earlier, first-term Representatives should get their own category. The remaining categories are to be as close in size to each other as possible.
RepData <- inner_join(RepData, LeftData, by = "DistrictCode") %>%
select(., ends_with('x'), DistrictCode, Terms) %>%
select(., Member = Member.x, Committees = Committees.x, State = State.x, DistrictCode, Terms)
RepData$Seniority <- cut(RepData$Terms,
c(0,1,3,7,30),
c('1st Term', '2 to 3 Terms', '4 to 7 Terms', '8 or More Term'),
include.lowest = FALSE)
rm(SeniorData, FixerByFirst, FixerByLast, LeftData, MissingLeft, d, url)
rm(get.code, get.district,
get.first.name, get.first.name.senior,
get.last.name, get.last.name.senior,
get.state,
get.state.senior)
5. Coding committee membership
A useful last preparatory step is to add flags for committees and members with leadership positions.
RepData <- mutate(RepData, Agriculture = str_detect(Committees,'Agriculture')) %>%
mutate(., Appropriations = str_detect(Committees,'Appropriations')) %>%
mutate(., ArmedServices = str_detect(Committees,'Armed Services')) %>%
mutate(., Education = str_detect(Committees,'Education')) %>%
mutate(., EnergyCommerce = str_detect(Committees,'Energy and')) %>%
mutate(., HomelandSecurity = str_detect(Committees,'Homeland')) %>%
mutate(., NaturalResources = str_detect(Committees,'Natural')) %>%
mutate(., ScienceSpaceTechnology = str_detect(Committees,'Science')) %>%
mutate(., SmallBusiness = str_detect(Committees,'Small Business')) %>%
mutate(., Transportation = str_detect(Committees,'Transportation')) %>%
mutate(., VeteransAffairs = str_detect(Committees,'Veterans')) %>%
mutate(., Leadership = str_detect(Member,'Boehner.|Pelosi.|McCarthy.|Hoyer.|Clyburn.|Scalise.'))
write.table(RepData, "CongressData.csv", row.names = FALSE, sep = ',')
saveRDS(RepData,"CongressData.rds")
rm(RepData)
The contract data is now fully prepared to be analyzed. Let's walk through some programming framework and potential avenues for analysis.
Analysis
Set up
First, let's load the data stored during processing, as well as any libraries we'll need.
library(stringr)
library(XML)
library(dplyr,warn.conflicts=FALSE)
library(ggplot2)
library(scales)
contracts <- readRDS("contracts.rds") %>% tbl_df()
agencies <- readRDS("agency_codes.rds") %>% tbl_df()
congress <- readRDS("CongressData.rds") %>% tbl_df()
Analysis 1. Relative impact of agencies
To warm-up, let's display a bar chart showing the largest agencies by contract value. As is widely known, the Defense Department dominates.
library(stringr) library(dplyr,warn.conflicts=FALSE) library(ggplot2) library(scales) by_agency.plot.data <- inner_join(contracts,agencies, by = c("Agency.Code" = "id")) %>% group_by(.,ShortName) %>% summarise(., Total = sum(Dollars.Obligated)/1e9) %>% filter(., Total > 5) agency.plot <- ggplot(aes(x=reorder(ShortName,Total), y=Total), data = by_agency.plot.data) + geom_bar(stat = "identity") + scale_x_discrete(name = "") + scale_y_continuous(name = "Total Contracts (billions)", labels = dollar) + ggtitle("Largest Agencies by Total Contract Value") + theme_bw() agency.plot
Analysis 2. Distribution of contract spending across districts
Let's try plotting some histograms featuring congressional districts as data points, bucketing the data according to the total value of contacts in each district.
library(ggplot2)
library(dplyr)
library(scales)
by_district <- group_by(contracts, pop_cd) %>%
summarise(., Total = sum(Dollars.Obligated)/1e6) %>%
rename(., DistrictCode = pop_cd) %>%
semi_join(.,congress, by = "DistrictCode")
by_district$Total <- sapply(by_district$Total, function(x) ifelse(x<1,1,x))
district.plot <- ggplot(aes(x=Total), data = by_district) +
geom_histogram() +
scale_x_continuous(name = "Total Contracts (millions)",
labels = dollar) +
ggtitle('Counts of Congressional Districts by Total Contract Values') +
ylab('') +
theme_bw()
district.plot
One takeaway from the above plot is that the distribution is highly skewed. We'll therefore have to use log scales in some of our plots. Further statistical analysis might have to employ non-parametric methods.
We can plot similar histograms for specific departments, starting with Defense.
by_district_agency <- inner_join(contracts,agencies, by = c("Agency.Code" = "id")) %>%
group_by(., pop_cd, ShortName) %>%
summarise(., Total = sum(Dollars.Obligated)/1e6) %>%
rename(., DistrictCode = pop_cd) %>%
semi_join(.,congress, by = 'DistrictCode')
by_district_agency$Total <- sapply(by_district_agency$Total, function(x) ifelse(x<1,1,x))
defense.plot <- ggplot(data=filter(by_district_agency,ShortName=='Defense'),
aes(x=Total)) +
geom_histogram(binwidth = 100) +
scale_x_continuous(name = "Total Contracts (millions)",
labels = dollar) +
ggtitle('Counts of Congressional Districts by Total Defense Contract Values') +
ylab('') +
theme_bw()
defense.plot
To check whether the skew in total spending results purely from the skew in Defense spending, we can make similar plots for other selected departments and agencies.
by_district_agency_other <- filter(by_district_agency,
ShortName %in% c('NASA', 'Veterans Affairs',
'HHS', 'Energy'))
other.plot <- ggplot(data = by_district_agency_other, aes(x = Total)) +
facet_grid(ShortName~.) +
geom_histogram(binwidth = 10) +
scale_x_continuous(name = "Total Contracts (millions)",
labels = dollar) +
ggtitle('Counts of Congressional Districts by Total Contract Values') +
ylab('') +
coord_cartesian(xlim = c(0,250)) +
theme_bw()
other.plot
It seems the skew in total spending does not result purely from the skew in Defense spending, as skewness is present at multiple scales.
Analysis 3. Spending in districts versus committee membership
Next, let's try plotting some graphs to show the distribution of contract values across congressional districts, according to committee memberships of those Congressperson representing contracted districts. We'll do so by department, showing in each graph the distribution of committees responsible for the given department, as well as the Appropriations committee (responsible for all departments) and all other members of congress.
In order to plot these graphs, we first create a data frame that joins contract data to congressional data, and then a function to explicitly insure uniform-looking plots.
library(ggplot2)
library(dplyr)
library(scales)
by_distr_agc_comm <- inner_join(by_district_agency, congress, by = "DistrictCode") %>%
select(., DistrictCode, Seniority, ShortName, Total, 9:20)
make.plot <- function(DeptName,CommName){
plot.data <- filter(by_distr_agc_comm, ShortName == DeptName)
comm.bool <- select(plot.data,matches(CommName))[[2]]
plot.data$Committee <- mapply(function(x,y) if(x) CommName
else if(y) 'Appropriations'
else 'Other',
comm.bool,
plot.data$Appropriations)
plot.data$Committee <- factor(plot.data$Committee,
levels = c(CommName,'Appropriations', 'Leadership','Other'))
return.plot <- ggplot(data = plot.data,
aes(x = Committee, y = Total, color = Committee), guide = TRUE) +
geom_jitter(position = position_jitter(height = 0.2, width = 0.25),
size = 2) +
coord_trans(y = 'log10') +
scale_size_discrete(range = c(2,4)) +
scale_y_continuous(name = "Total Contracts (millions)",
labels = dollar) +
ggtitle(str_c('Department: ', DeptName)) +
scale_x_discrete(labels = '') +
geom_hline(yintercept = median(plot.data$Total),
color = 'black', linetype = 'solid', width = 2) +
xlab('') +
theme_bw()
}
Below are the plots for various departments and agencies. The black line represents the median for the set of all 435 districts for the given agency.
Think of these plots as visual contingency tables. For each committee, we can easily check whether more than half the members are above the median line. This is relevant because in a random sample of Representatives we would expect the points in the sample to be split more or less evenly on the two sides of this line.
Likewise, we can visually check whether a given committee is overrepresented among the districts with the most contract spending. (Keep in mind that each committee comprises 10-15% of Congress.)
Analysis 4. Plots by seniority
Since a Representative’s seniority is a potential contributing factor to spending in his or her district, let's look at plots broken up by seniority buckets. We can also display the size of each bucket for reference.
##
## 1st Term 2 to 3 Terms 4 to 7 Terms 8 or More Term
## 59 147 114 115
Something to keep in mind when looking at the plots below is that the leftmost panel displays data for first-term members. These are members who were not involved in the spending decisions. In each case shown, however, more than have the first-term members of the supervising committee are from districts where contract spending is above the overall median.
The pattern we see among the first-term Representatives raises the possibility that the association we observed earlier between contract spending and committee representation results from Representatives being assigned to committees based on the contracts that are already in place in their districts.
Conclusion
These findings are of course preliminary. I did not do an exhaustive analysis of every department/committee, and I have not looked at federal expenditures that are made through grants rather than contracts. These are possible extensions to be made in future projects.