The Impact of Demographics on Campaign Spending and Voting

Summer 2025 Data Science Project

Tyshon Brown, David Li, Zile Liu, and Samuel Opoku-Agyemang

Contributions:

• Tyshon Brown:
  • Population Density vs Voter Turnout Section
  • Extracted County Info for ML (median income, age, total pop)
  • ML Training and Testing
  • Machine Learning Performance Analysis
• David Li:
  • Project idea
  • Hypothesis tests and charts
  • Final report creation and deployment
• Zile Liu:
  • Extracted data of funding of the 2020 compaign
  • Performed Chi squared test
• Samuel Opoku-Agyemang:
  • Data Curation
  • EDA Hypothesis 1
  • Extracted County Info for ML (total individual contributions)
  • Machine Learning Data Scaling

Table of Contents:

1. Introduction
2. Data Curation
3. Exploratory Data and Analysis
4. Machine Learning Analysis
5. Final Insights

1. Introduction

Contributions from constituents and voters play a very important role in any political campaign, and we wish to explore the relationship between demographics such as age, income, and population (at the county and state level) and political campaign contributions, as well as political election outcomes.

We ask and attempt to answer questions such as

• Which demographics spend the most?
• Which demographics spend the least?
• How much does funding correlate with election outcomes?
• Is there a strong enough relationship between these demographics where we can train a machine learning model to predict outcomes?

We do this using tools from the Python Data Analysis Library and scikit-learn to

• Clean our data and merge into one, well maintained database.
• Explore the relationship between the aformentioned variables.
• Train a predictive machine learning model and test its performance.
• Visualize any trends in a very presentable way.

2. Data Curation

We collected and merged the following datasets:

• Population, Income, and Age from the 2020 ACS via the census Python API.
• Population density from Census Historical Density Tables.
• Campaign finance and expenditures from the FEC’s 2020 data archive.
• Popular vote totals by state from the Federal Elections 2020 report.

We chose these because these datasets are of very high volume due to them coming from the federal government.

We used pandas, scipy, and seaborn for preprocessing and visualization.

import json

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from census import Census
from scipy.stats import chisquare, f_oneway, pearsonr, spearmanr, tukey_hsd
from us import states

sns.set_theme()
c = Census("23c580dcbb047d99e93502f60681fc6430f2fc9a", year=2020)  # Sam's API key

/Users/davidli/cmsc320_project/env/lib/python3.9/site-packages/urllib3/__init__.py:35: NotOpenSSLWarning: urllib3 v2 only supports OpenSSL 1.1.1+, currently the 'ssl' module is compiled with 'LibreSSL 2.8.3'. See: https://github.com/urllib3/urllib3/issues/3020
  warnings.warn(

State Population

population_data = c.acs5.state(("NAME", "B01003_001E"), Census.ALL)
population = pd.DataFrame(population_data)
population.rename(columns={"NAME": "Name", "B01003_001E": "Population"}, inplace=True)
population["Population"] = population["Population"].astype(int)

population = population.drop("state", axis=1)
population.head()

	Name	Population
0	Pennsylvania	12794885
1	California	39346023
2	West Virginia	1807426
3	Utah	3151239
4	New York	19514849

Population Density by State

# https://www.census.gov/data/tables/time-series/dec/density-data-text.html
# renamed from "apportionment.csv"
pop_density = pd.read_csv("population_density.csv")
pop_density = pop_density[
    ["Name", "Geography Type", "Year", "Resident Population Density"]
]
pop_density = pop_density[
    (pop_density["Year"] == 2020) & (pop_density["Geography Type"] == "State")
]
pop_density = pop_density[["Name", "Resident Population Density"]].reset_index(
    drop=True
)

# 52 datapoints because DC and Puerto Rico count as states
pop_density.head()

	Name	Resident Population Density
0	Alabama	99.2
1	Alaska	1.3
2	Arizona	62.9
3	Arkansas	57.9
4	California	253.7

Median Age by State

median_age_dictionary = c.acs5.state(("NAME", "B01002_001E"), Census.ALL)
median_age = pd.DataFrame.from_dict(median_age_dictionary).rename(
    columns={"NAME": "Name", "B01002_001E": "Median Age"}
)
median_age = (
    median_age[["Name", "Median Age"]].sort_values(by=["Name"]).reset_index(drop=True)
)
median_age.head()

	Name	Median Age
0	Alabama	39.2
1	Alaska	34.6
2	Arizona	37.9
3	Arkansas	38.3
4	California	36.7

Median Income by State

median_income_dictionary = c.acs5.state(("NAME", "B19013_001E"), Census.ALL)
median_income = pd.DataFrame.from_dict(median_income_dictionary).rename(
    columns={"NAME": "Name", "B19013_001E": "Median Household Income"}
)
median_income = (
    median_income[["Name", "Median Household Income"]]
    .sort_values(by=["Name"])
    .reset_index(drop=True)
)
median_income.head()

	Name	Median Household Income
0	Alabama	52035.0
1	Alaska	77790.0
2	Arizona	61529.0
3	Arkansas	49475.0
4	California	78672.0

Popular Vote by State (Biden v Trump)

# https://www.fec.gov/introduction-campaign-finance/election-results-and-voting-information/federal-elections-2020/
popular_vote = pd.read_csv("federalelections2020.csv")
popular_vote.head()

	Name	Biden Votes	Trump Votes	Total Votes
0	Alabama	849624	1441170	2323282
1	Alaska	153778	189951	359530
2	Arizona	1672143	1661686	3387326
3	Arkansas	423932	760647	1219069
4	California	11110639	6006518	17501380

2020 Election Candidate Funding

## Source: https://www.fec.gov/resources/campaign-finance-statistics/2020/tables/ie/IE2_2020_24m.pdf
independent_expenditure = pd.read_csv("candidates_funding.csv")
independent_expenditure = independent_expenditure.sort_values(
    by=["Total IEs For"], ascending=False
)
independent_expenditure.index = range(0, len(independent_expenditure))
independent_expenditure.head()

	Candidate	Total IEs For	Total IEs Against
0	Biden, Joseph R Jr	383204094	299388312
1	Trump, Donald J.	53500635	308135887
2	Warren, Elizabeth	14996979	18965
3	Buttigieg, Pete	3668912	30862
4	Klobuchar, Amy J.	2711204	101

State Contributions

state_contributions = pd.read_csv("state_spending.csv")
state_contributions = state_contributions.rename(
    columns={"State (or District)": "Name"}
).drop(columns=["Rank"])
state_contributions["Total Contributions"] = (
    state_contributions["Total Contributions"].str.replace("$", "").astype(int)
)
state_contributions["Percent to Democrats"] = (
    state_contributions["Percent to Democrats"].str.replace("%", "").astype(float) / 100
)
state_contributions["Percent to Republicans"] = (
    state_contributions["Percent to Republicans"].str.replace("%", "").astype(float)
    / 100
)
state_contributions["Contributions to Democrats"] = (
    state_contributions["Total Contributions"]
    * state_contributions["Percent to Democrats"]
)
state_contributions["Contributions to Republicans"] = (
    state_contributions["Total Contributions"]
    * state_contributions["Percent to Republicans"]
)
state_contributions.head()

	Name	Total Contributions	Percent to Democrats	Percent to Republicans	Contributions to Democrats	Contributions to Republicans
0	California	1666572970	0.6931	0.2842	1.155102e+09	4.736400e+08
1	New York	978988661	0.7462	0.2318	7.305213e+08	2.269296e+08
2	District of Columbia	810155426	0.6701	0.3148	5.428852e+08	2.550369e+08
3	Texas	687622660	0.3488	0.6315	2.398428e+08	4.342337e+08
4	Florida	646468276	0.3692	0.6108	2.386761e+08	3.948628e+08

3. Exploratory Data Analysis

We conducted exploratory analysis to test the following three hypotheses:

1. Does median age correlate with political funding outcomes?
   Specifically, do older states tend to contribute more to Republican or Democratic candidates?

2. Does population density correlate with voter turnout?
   We explored whether people in more densely populated states are more likely to vote.

3. Does campaign spending correlate with election success?
   In particular, do top-funded candidates consistently secure more votes, and is the relationship statistically significant?

To answer these, we:

• Merged and cleaned datasets from the U.S. Census Bureau and the FEC.
• Performed statistical tests including ANOVA, Tukey's HSD, Pearson and Spearman correlation, and Chi-Square tests.
• Created scatterplots, regression lines, and bar graphs to visualize relationships.
• Used state-level data for the 2020 U.S. presidential election to ensure consistency and completeness.

Each method helped us characterize the data, identify patterns, and evaluate the strength of relationships between variables. See the subsections below for details, plots, and conclusions.

Hypothesis 1: Does median age correlate with political funding outcomes?

We performed a one-way ANOVA and a Tukey's honest significant difference test. Our null hypothesis is that the proportion of contributions to Democrats/Republicans from a state is independent of its median age.

joint_df = pd.merge(left=median_age, right=state_contributions, how="inner")
joint_df = pd.merge(left=joint_df, right=population, how="inner")
joint_df.insert(
    2,
    "Age Group",
    pd.qcut(joint_df["Median Age"], q=3, labels=["Lower", "Medium", "Higher"]),
)

lower = joint_df[joint_df["Age Group"] == "Lower"]["Percent to Republicans"]
medium = joint_df[joint_df["Age Group"] == "Medium"]["Percent to Republicans"]
higher = joint_df[joint_df["Age Group"] == "Higher"]["Percent to Republicans"]

statistic, p_value = f_oneway(lower, medium, higher)
print(f"ANOVA statistic: {statistic:.4f}, ANOVA p-value: {p_value:.4f}")
# significant, so we proceed with a post-hoc Tukey's HSD
print(tukey_hsd(lower, medium, higher))

ANOVA statistic: 5.3387, ANOVA p-value: 0.0081
Tukey's HSD Pairwise Group Comparisons (95.0% Confidence Interval)
Comparison  Statistic  p-value  Lower CI  Upper CI
 (0 - 1)      0.064     0.481    -0.069     0.197
 (0 - 2)      0.178     0.006     0.045     0.311
 (1 - 0)     -0.064     0.481    -0.197     0.069
 (1 - 2)      0.114     0.109    -0.020     0.247
 (2 - 0)     -0.178     0.006    -0.311    -0.045
 (2 - 1)     -0.114     0.109    -0.247     0.020

Visual: Median Age and Percent to Republican Candidates (ANOVA, Tukey's HSD, Regresison)

plt.figure(figsize=(10, 6))
sns.set_theme(style="whitegrid")

# Regression Line
sns.regplot(
    data=joint_df,
    x="Median Age",
    y="Percent to Republicans",
    scatter=False,
    color="red",
    line_kws={"linewidth": 1.5},
)

# Scatter Plot
plt.scatter(
    x=joint_df["Median Age"],
    y=joint_df["Percent to Republicans"],
    s=joint_df["Population"] / 25000,
    alpha=0.7,
    edgecolor="k",
)
plt.tight_layout()
plt.show()

Visual: Median Age and Percent to Democratic Candidates (Plot & Regression)

plt.figure(figsize=(10, 6))
sns.set_theme(style="whitegrid")

# Regression Line
sns.regplot(
    data=joint_df,
    x="Median Age",
    y="Percent to Democrats",
    scatter=False,
    color="red",
    line_kws={"linewidth": 1.5},
)

# Scatter Plot
plt.scatter(
    x=joint_df["Median Age"],
    y=joint_df["Percent to Democrats"],
    s=joint_df["Population"] / 25000,
    alpha=0.7,
    edgecolor="k",
)
plt.tight_layout()
plt.show()

Conclusion for Hypothesis 1:

There is statistically significant evidence that median age is associated with differences in political contribution patterns across states.

• The one-way ANOVA results in a statistic of 5.34 and a p-value of 0.0081, indicating we can reject the null hypothesis at the 1% level. This means that the proportion of contributions to political parties is not independent of median age.
• The Tukey's HSD post-hoc test shows that the largest difference in contribution patterns exists between group 0 (younger age group) and group 2 (older age group), with a statistically significant difference (p = 0.006, 95% CI = [0.045, 0.311]).
• The scatter plots show:
  • A negative trend between median age and percent of contributions to Republicans, suggesting that states with older populations may tend to contribute less to Republican candidates.
  • A positive trend between median age and percent of contributions to Democrats, suggesting that states with older populations may tend to contribute more to Democratic candidates.

Hypothesis 2: Does population density correlate to voter turnout? (Pearsons Coefficient, Spearman Coefficient, Regression)

We performed a simple linear regression. We use Pearson and Spearman correlation analyses to test the null hypothesis that a state's voter turnout is independent of its population density.

Turnout Rate:

merged_pop_df = popular_vote.merge(population, on="Name")
merged_pop_df["Turnout"] = merged_pop_df["Total Votes"] / merged_pop_df["Population"]

# popular_vote['Turnout'] = popular_vote['Total Votes'] / population['Population']
merged_pop_df.head(5)

	Name	Biden Votes	Trump Votes	Total Votes	Population	Turnout
0	Alabama	849624	1441170	2323282	4893186	0.474799
1	Alaska	153778	189951	359530	736990	0.487836
2	Arizona	1672143	1661686	3387326	7174064	0.472163
3	Arkansas	423932	760647	1219069	3011873	0.404754
4	California	11110639	6006518	17501380	39346023	0.444807

Merging the dataframes for visualization:

# merging Popular Vote and Population Density and Population
density_turnout = merged_pop_df.merge(pop_density, on="Name")

# Calculating turnout and adding it to the df
density_turnout["Turnout"] = (
    density_turnout["Total Votes"] / density_turnout["Population"] * 100
)
density_turnout["Resident Population Density"] = pd.to_numeric(
    density_turnout["Resident Population Density"], errors="coerce"
)

# Dropping Na's
density_turnout = density_turnout.dropna(
    subset=["Resident Population Density", "Turnout"]
)

density_turnout.head()

	Name	Biden Votes	Trump Votes	Total Votes	Population	Turnout	Resident Population Density
0	Alabama	849624	1441170	2323282	4893186	47.479945	99.2
1	Alaska	153778	189951	359530	736990	48.783566	1.3
2	Arizona	1672143	1661686	3387326	7174064	47.216278	62.9
3	Arkansas	423932	760647	1219069	3011873	40.475445	57.9
4	California	11110639	6006518	17501380	39346023	44.480684	253.7

The Plot:

• x-axis: Population Density
• y-axis: Voter Turnout
• blue dots/points: scaled to the population size
• red regression line

plt.figure(figsize=(10, 6))
sns.set_theme(style="whitegrid")

# Regression Line
sns.regplot(
    data=density_turnout,
    x="Resident Population Density",
    y="Turnout",
    scatter=False,
    color="red",
    line_kws={"linewidth": 1.5},
)


# Scatter Plot
plt.scatter(
    x=density_turnout["Resident Population Density"],
    y=density_turnout["Turnout"],
    s=density_turnout["Population"] / 25000,
    alpha=0.7,
    edgecolor="k",
)

plt.xlabel("Population Density (people per sq. mile)")
plt.ylabel("Voter Turnout (%)")
plt.title("Population Density vs Voter Turnout (2020)")
plt.tight_layout()
plt.show()

# Pearson Correlation
p_corr, p_p = pearsonr(
    density_turnout["Resident Population Density"], density_turnout["Turnout"]
)
print(f"Pearson correlation: {p_corr:.3f}, p-value: {p_p:.3g}")

# Spearman Correlation
s_corr, s_p = spearmanr(
    density_turnout["Resident Population Density"], density_turnout["Turnout"]
)
print(f"Spearman correlation: {s_corr:.3f}, p-value: {s_p:.3g}")

Pearson correlation: 0.070, p-value: 0.636
Spearman correlation: 0.051, p-value: 0.733

Conclusion for Hypothesis 2:

There is no significant correlation between population density and voter turnout across states.

• The Pearson correlation coefficient is 0.070 and the Spearman correlation coefficient is 0.051. Both indicating a very weak positive relationship.
• The p-values for both tests exceed 0.05, meaning the correlations are not statistically significant.
• We therefore fail to reject the null hypothesis that population density and voter turnout are unrelated.
• This finding is also supported by the scatter plot, which shows no clear trend or pattern between the two variables.

Hypothesis 3: Are all candidates equally funded? (Pearson's Chi-square)

For simplicity, we use Total IEs For.

x2, p = chisquare(independent_expenditure["Total IEs For"])
print(f"Chi-square: {x2}, p-value: {p}")

ax = independent_expenditure["Total IEs For"].hist(bins=16, xrot=45)
ax.set_xlabel("Total IEs For")
ax.set_ylabel("Candidates")
ax.set_title("Histogram of Candidate Independent Expenditures")
ax.ticklabel_format(style="plain")

Chi-square: 12085429117.471506, p-value: 0.0

Conclusion for Hypothesis 3:

• The chi-square test yields a p-value of 0, well below the standard alpha level of 0.05. Thus, we reject the null hypothesis, indicating not all candidates are the same size. We must not make this assumption in later analysis.
• The histogram visually reinforces this conclusion: the distribution is skewed right, with two candidates, Biden and Trump, receiving far more independent expenditures than all other candidates.

4. Machine Learning Analysis

We will use a Model to predict total campaign contributions using

• Median Household Income
• Median Age
• Total Population

This model helps us understand how demographic factors influence financial engagement in U.S. elections. Initially we collected this data based only on states. But we realized that 50 state sample size was too small for reliable machine learning. To improve our model, we expanded to county-level data.

This model would offer value to campaigns:

• Candidates can identify high-potential fundraising areas based on local demographics
• Instead of targeting entire states, campaigns can prioritize specific counties likely to yield more contributions

Overall, this approach transforms demographics into a predictive fundraising map, helping campaigns allocate resources more effectively based on data.

County Median Age, Income, and Total Population

all_counties = []
# Variables:
# B19013_001E → Median Household Income
# B01003_001E → Total Population
# B01002_001E → Median Age
variables = ("NAME", "B19013_001E", "B01003_001E", "B01002_001E")

for state in states.STATES:
    try:
        county_data = c.acs5.get(
            variables, {"for": "county:*", "in": f"state:{state.fips}"}
        )
        for row in county_data:
            row["state_name"] = state.name  # state name is included
        all_counties.extend(county_data)
    except Exception as e:
        print(f"Error retrieving data for {state.name}: {e}")

county_df = pd.DataFrame(all_counties)
county_df.rename(
    columns={
        "NAME": "County Name",
        "B19013_001E": "Median Income",
        "B01003_001E": "Population",
        "B01002_001E": "Median Age",
        "state": "State FIPS",
        "county": "County FIPS",
        "state_name": "State",
    },
    inplace=True,
)

# Convert to numeric
for col in ["Median Income", "Population", "Median Age"]:
    county_df[col] = pd.to_numeric(county_df[col], errors="coerce")

print(county_df.shape)
county_df.head(100)

(3142, 7)

	County Name	Median Income	Population	Median Age	State FIPS	County FIPS	State
0	Autauga County, Alabama	57982.0	55639.0	38.6	01	001	Alabama
1	Baldwin County, Alabama	61756.0	218289.0	43.2	01	003	Alabama
2	Barbour County, Alabama	34990.0	25026.0	40.1	01	005	Alabama
3	Bibb County, Alabama	51721.0	22374.0	39.9	01	007	Alabama
4	Blount County, Alabama	48922.0	57755.0	41.0	01	009	Alabama
...	...	...	...	...	...	...	...
95	Prince of Wales-Hyder Census Area, Alaska	54018.0	6338.0	42.1	02	198	Alaska
96	Southeast Fairbanks Census Area, Alaska	66941.0	6911.0	36.8	02	240	Alaska
97	Apache County, Arizona	33967.0	71714.0	35.4	04	001	Arizona
98	Cochise County, Arizona	51505.0	126442.0	41.0	04	003	Arizona
99	Coconino County, Arizona	59000.0	142254.0	31.0	04	005	Arizona

100 rows × 7 columns

Individual Contributions

import re

import zipcodes

# https://code.activestate.com/recipes/577305-python-dictionary-of-us-states-and-territories/
state_names = {
    "AL": "Alabama",
    "AK": "Alaska",
    "AZ": "Arizona",
    "AR": "Arkansas",
    "CA": "California",
    "CO": "Colorado",
    "CT": "Connecticut",
    "DE": "Delaware",
    "FL": "Florida",
    "GA": "Georgia",
    "HI": "Hawaii",
    "ID": "Idaho",
    "IL": "Illinois",
    "IN": "Indiana",
    "IA": "Iowa",
    "KS": "Kansas",
    "KY": "Kentucky",
    "LA": "Louisiana",
    "ME": "Maine",
    "MD": "Maryland",
    "MA": "Massachusetts",
    "MI": "Michigan",
    "MN": "Minnesota",
    "MS": "Mississippi",
    "MO": "Missouri",
    "MT": "Montana",
    "NE": "Nebraska",
    "NV": "Nevada",
    "NH": "New Hampshire",
    "NJ": "New Jersey",
    "NM": "New Mexico",
    "NY": "New York",
    "NC": "North Carolina",
    "ND": "North Dakota",
    "OH": "Ohio",
    "OK": "Oklahoma",
    "OR": "Oregon",
    "PA": "Pennsylvania",
    "RI": "Rhode Island",
    "SC": "South Carolina",
    "SD": "South Dakota",
    "TN": "Tennessee",
    "TX": "Texas",
    "UT": "Utah",
    "VT": "Vermont",
    "VA": "Virginia",
    "WA": "Washington",
    "WV": "West Virginia",
    "WI": "Wisconsin",
    "WY": "Wyoming",
    "AS": "American Samoa",
    "DC": "District of Columbia",
    "GU": "Guam",
    "MP": "Northern Mariana Islands",
    "PR": "Puerto Rico",
    "VI": "U.S. Virgin Islands",
    "AA": "Armed Forces Americas",
    "AE": "Armed Forces Europe",
    "AP": "Armed Forces Pacific",
}


def zip_to_state(zip_code):
    if zip_code:
        matches = zipcodes.matching(zip_code)
        return matches[0]["state"] if matches else None
    else:
        return None


# County Name, State
def zip_to_county(zip_code):
    matches = zipcodes.matching(zip_code)
    if matches:
        if matches[0]["county"] and (matches[0]["state"] in state_names):
            return matches[0]["county"] + ", " + state_names[matches[0]["state"]]
        else:
            return None
    else:
        return None


def clean_zip(zip_code):
    if zip_code is None or pd.isna(zip_code):
        return None

    zip_code = str(zip_code).strip()

    if not re.fullmatch(r"\d+", zip_code):
        return None

    return zip_code[:5].zfill(5)

Sampling Individual Contributions

There were over 69 million campaign total campaign contributions for the 2020 election cycle, and to cope our current technical limitations, we use only a portion of these. Our data is restricted to the time period of 2020-09-02 to 2020-11-26, the time closest to the elections where we believe the most important political activity would occur.

header = [
    "CMTE_ID",
    "AMNDT_IND",
    "RPT_TP",
    "TRANSACTION_PGI",
    "IMAGE_NUM",
    "TRANSACTION_TP",
    "ENTITY_TP",
    "NAME",
    "CITY",
    "STATE",
    "ZIP_CODE",
    "EMPLOYER",
    "OCCUPATION",
    "TRANSACTION_DT",
    "TRANSACTION_AMT",
    "OTHER_ID",
    "TRAN_ID",
    "FILE_NUM",
    "MEMO_CD",
    "MEMO_TEXT",
    "SUB_ID",
]

# data from https://www.fec.gov/data/browse-data/?tab=bulk-data, contributions by individuals
files = [
    "by_date/itcont_2020_20200902_20200911.txt",
    "by_date/itcont_2020_20200912_20200919.txt",
    "by_date/itcont_2020_20200920_20200925.txt",
    "by_date/itcont_2020_20200926_20200930.txt",
    "by_date/itcont_2020_20200930_20201004.txt",
    "by_date/itcont_2020_20201005_20201010.txt",
    "by_date/itcont_2020_20201011_20201014.txt",
    "by_date/itcont_2020_20201015_20201018.txt",
    "by_date/itcont_2020_20201019_20201022.txt",
    "by_date/itcont_2020_20201023_20201026.txt",
    "by_date/itcont_2020_20201027_20201030.txt",
    "by_date/itcont_2020_20201030_20201101.txt",
    "by_date/itcont_2020_20201102_20201106.txt",
    "by_date/itcont_2020_20201107_20201115.txt",
    "by_date/itcont_2020_20201116_20201126.txt",
]

# chunks = []
# i = 1
# for file in files:
#     chunk = pd.read_csv(file, sep = '|', names = header, on_bad_lines = 'skip', usecols=['ZIP_CODE', 'TRANSACTION_AMT'])
#     chunk = chunk.dropna(how = 'any')

#     # cleaning zip codes
#     chunk = chunk[~chunk['ZIP_CODE'].isin(["00000", "000.0"])]
#     chunk.loc[:, 'ZIP_CODE'] = chunk['ZIP_CODE'].apply(clean_zip)

#     # cleaning transaction amounts
#     chunk.loc[:, 'TRANSACTION_AMT'] = pd.to_numeric(chunk['TRANSACTION_AMT'], errors = 'coerce', downcast = 'integer')
#     chunk = chunk.dropna(subset = ['TRANSACTION_AMT'])

#     # grouping and summing by zip code
#     grouped_df = chunk.groupby('ZIP_CODE').agg({'TRANSACTION_AMT': 'sum'}).reset_index()

#     # adding county column
#     grouped_df.loc[:, 'COUNTY'] = grouped_df['ZIP_CODE'].apply(zip_to_county)

#     # final processing and exporting
#     grouped_by_county_df = grouped_df.groupby('COUNTY').agg({'TRANSACTION_AMT' : 'sum'})

#     chunks.append(grouped_by_county_df)
#     print(f"{(i / 15):.4%}")
#     i += 1

# sept_to_nov_df = pd.concat(chunks).groupby('COUNTY').agg({'TRANSACTION_AMT' : 'sum'})
# sept_to_nov_df = sept_to_nov_df.rename(columns={'COUNTY' : 'County Name', 'TRANSACTION_AMT' : 'Contributions'})

# contributions_df = pd.read_csv("individual_contributions.csv")
# print(contributions_df.shape)

contributions_df = pd.read_csv("sept_to_nov.csv").rename(
    columns={"COUNTY": "County Name", "TRANSACTION_AMT": "Contributions"}
)
contributions_df.head()

	County Name	Contributions
0	Abbeville County, South Carolina	31868.0
1	Acadia Parish, Louisiana	52030.0
2	Accomack County, Virginia	139421.0
3	Ada County, Idaho	5564755.0
4	Adair County, Iowa	13036.0

MLPRegressor

We opted to train our data on a MLPRegressor model over a Linear Regression as we expect for the relationship between the inputs and outputs to be slightly non-linear. The increased complexity gained from choosing MLPRegressor also allows for our model to better cope with the restricted data.

Our data was scaled using the logistic and standard scaling as there were outliers that we needed to account for before training, fitting, and testing our model.

from sklearn.model_selection import train_test_split
from sklearn.neural_network import MLPRegressor
from sklearn.preprocessing import StandardScaler

machine_learning_df = pd.merge(
    county_df[["County Name", "Median Income", "Median Age", "Population"]],
    contributions_df,
    on="County Name",
    how="inner",
)
machine_learning_df = machine_learning_df[0 <= machine_learning_df["Contributions"]]

# scaling using inter-quartile range
q1 = machine_learning_df["Contributions"].quantile(0.25)
q3 = machine_learning_df["Contributions"].quantile(0.75)
iqr = q3 - q1
machine_learning_df = machine_learning_df[
    q1 - 1.5 * iqr <= machine_learning_df["Contributions"]
]
machine_learning_df = machine_learning_df[
    machine_learning_df["Contributions"] <= q3 + 1.5 * iqr
]

# taking the log to further account for outliers
machine_learning_df = np.log1p(
    machine_learning_df[["Median Income", "Median Age", "Population", "Contributions"]]
)
machine_learning_df = machine_learning_df.dropna(how="any")
machine_learning_df.boxplot(column=["Contributions"])
plt.show()

/Users/davidli/cmsc320_project/env/lib/python3.9/site-packages/pandas/core/internals/blocks.py:395: RuntimeWarning: invalid value encountered in log1p
  result = func(self.values, **kwargs)

# train, test, split
X = machine_learning_df[["Median Income", "Median Age", "Population"]]
y = machine_learning_df["Contributions"]
X_train, X_test, y_train, y_test = train_test_split(
    X, y, random_state=42, test_size=0.2, stratify=pd.qcut(y, q=5)
)

# standard scaling
scaler_X = StandardScaler().fit(X_train)
X_train_scaled = scaler_X.transform(X_train)
X_test_scaled = scaler_X.transform(X_test)

# standard scaling
scaler_y = StandardScaler().fit(y_train.values.reshape(-1, 1))
y_train_scaled = scaler_y.transform(y_train.values.reshape(-1, 1)).ravel()
y_test_scaled = scaler_y.transform(y_test.values.reshape(-1, 1)).ravel()

With a set seed, our model has a $R^2$ coefficient of determination of $0.747$, meaning that $74.7\%$ of the variance in $Y$ can be explained by $X$. The alpha parameter controls the strength of the L2 regularization term, and this was increased from its base value of $0.0001$ to $0.1$ to decrease the amount of overfitting and thus generalize the model better.

np.seterr("ignore")  # https://github.com/numpy/numpy/issues/28687

# building and fitting model
regressor = MLPRegressor(
    hidden_layer_sizes=(128, 64, 32), alpha=0.1, max_iter=3000, random_state=42
)
regressor.fit(X_train_scaled, y_train_scaled)

MLPRegressor(alpha=0.1, hidden_layer_sizes=(128, 64, 32), max_iter=3000,
             random_state=42)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

from sklearn.metrics import mean_squared_error, r2_score, root_mean_squared_error

# predicting
y_pred_scaled = regressor.predict(X_test_scaled)
y_pred = scaler_y.inverse_transform(y_pred_scaled.reshape(-1, 1)).ravel()

# statistical tests
mse = mean_squared_error(y_test, y_pred)
r2 = r2_score(y_test, y_pred)
rmse = root_mean_squared_error(y_test, y_pred)

print(f"Mean Squared Error: {mse:.4f}")
print(f"R-squared: {r2:.4f}")
print(f"Root Mean Squared Error: {rmse:.4f}")

Mean Squared Error: 0.4861
R-squared: 0.7515
Root Mean Squared Error: 0.6972

plt.scatter(y_test, y_pred, alpha=0.5)
plt.plot([y_test.min(), y_test.max()], [y_test.min(), y_test.max()], "k--")
plt.xlabel("Actual Log(Contributions)")
plt.ylabel("Predicted Log(Contributions)")
plt.title("Actual vs Predicted Values")
plt.show()

print(f"Train score (R²): {regressor.score(X_train_scaled, y_train_scaled):.4f}")
print(f"Test score (R²): {regressor.score(X_test_scaled, y_test_scaled):.4f}")

Train score (R²): 0.7635
Test score (R²): 0.7515

print(f'Median Contribution: {machine_learning_df["Contributions"].median()}')

Median Contribution: 10.957511756214629

Model Analysis

To assess the effectiveness of our predictive model, we trained a Multi-Layer Perceptron (MLP) regressor to estimate total campaign contributions at the county level, using only three input features:

• Median income
• Median age
• Total population

The MLP Configuration

We used an 80/20 train-test split (train_test_split with random_state=42) to evaluate performance.

MLPRegressor( hidden_layer_sizes=(128, 64, 32), activation="relu", alpha=0.1, max_iter=3000, random_state=42 )

• The hidden layer sizes were gradually increased from a smaller initial configuration
• The alpha parameter (L2 regularization strength) was tuned from 0.001 to 0.1, resulting in improved R² values.
• max_iter was increased from the default 1,000 to 3,000 to ensure convergence.

Preprocessing Enhancements

To improve predictive performance we applied a combination of outlier handling, transformation, and scaling:

1. Outlier Handling using IQR:
   • We removed counties with total contributions outside the range of [Q1 - 1.5×IQR, Q3 + 1.5×IQR]
2. Log Transformation (np.log1p):
   • Applied log transformations to all of the features and the target to handle a possibly skewed distributiont
3. Feature Scaling:
   • Applied StandardScaler seperately to the inputs and the target

Performance Metrics

Metric	Value
Train $R^2$ Score	$0.7635$
Test $R^2$ Score	$0.7515$
Root MSE	0.692

• The Train $R^2$ of $0.7635$: the model explains over $76\%$ of the variance in the training set, indicating it effectively learned patterns from the transformed data.

• The Test $R^2$ of $0.7515$: the model explains about $75\%$ of the variance in county-level contributions in unseen data, which indicates that outlier handling and log transformation helped align the training and test distributions.

• The Train $R^2$ shows the model is learning (not underfitting) and the Test $R^2$ is reasonably close (not overfitting)

• RMSE = $0.6972$ (log scale):

  • Meaning the average prediction is off by a factor of about $e^{0.6972}$ $\approx$ $2.01$. This means the model's prediction is typically within a factor of 2 of the actual contribution amount.
  • This level of error is reasonable given the variability of financial contributions and the use of only three basic demographic predictors.

Model Analysis Conclusion

Even with only three input variables, out MLP model shows strong predictive power. Explaining 75% of the variance in political contributions using just income, age, and population is solid, especially given the complexity and noisiness of campaign finance data. This makes it a powerful starting point for campaign strategies like:

• Identifying counties with high fundraising potential
• Understading demographic drivers of contributions

With additional features such as education level, internet access, or political affiliation, this model could be further refined. Even in its current form, it serves as a valuable tool for providing insights for campaign strategists.

5. Insights and Conclusions

After reading through this project, an uninformed reader will gain some insight, primarily through the visualizations, of the relationships between the intersection of demographics like income, age, and population, and campaign contributions and election outcomes.

A reader who does already know about the topic will be able to see visualizations as well as the outcomes of statistical tests such as the $R^2$ or the Pearson, which will likely align with their background knowledge. They may also gain some insight into the Data Science side of political information, as our process of data collection and exploratory data analysis are displayed.