Can Fast Food Predict How a State Votes?

Fast food is deeply embedded in American culture, but its significance extends beyond convenience. Campaign finance records consistently show Republicans and Democrats spend money at very different types of restaurants — GOP candidates favor McDonald's and Chick-fil-A while Democratic candidates tend toward Panera and Chipotle. These aren't coincidences. They reflect deeper cultural and economic identities that show up in voting behavior.

Fast food consumption isn't strictly a low-income phenomenon either. Research shows it actually peaks at middle income levels, then falls off among the highest earners — people in the top income group are 54.5% less likely to eat fast food than the lowest group. This complexity is part of why the data seemed interesting: franchise density encodes regional culture, economic patterns, and demographic identity all at once.

We focused on eleven chains that reflect a range of political, economic, and regional identities: Starbucks, Subway, McDonald's, Taco Bell, Domino's, Chick-fil-A, Arby's, Wendy's, KFC, Dunkin' Donuts, and Chipotle. Some of these (Chick-fil-A, Chipotle) have well-documented partisan associations in campaign spending data. Others are national staples with broad reach. Together they gave us a feature space rich enough to train a classifier on 50 data points — one per state.

Getting the data required two approaches. Some chains had clean Kaggle datasets we could load directly — Wendy's, McDonald's (via a CSUN research table), Starbucks, and Subway. The rest required web scraping the chains' official location pages using Selenium and BeautifulSoup, since the data either didn't exist in structured form or was paywalled.

• Dataset #1: 2024 County-Level Presidential Results (Kaggle, 3,088 rows) — aggregated to state level for modeling
• CSV chains: Wendy's, McDonald's, Starbucks, Subway — cleaned and grouped by 2-letter state abbreviation
• Scraped chains: Taco Bell, Domino's, KFC, Dunkin' Donuts, Chipotle, Arby's — parsed from each chain's official location pages
• Population data: pulled from the TotalPop field in the election dataset for per-capita normalization

My contribution was building and running the web scraping pipeline for the six chains without accessible datasets. Each chain's site had a different structure — dropdowns, paginated tables, flat lists — which meant writing custom parsing logic for each one rather than a single reusable scraper.

Most states are not close calls. The vote share histogram shows a mostly bimodal distribution — states tend to cluster toward one candidate rather than landing near 50/50. The geographic map makes regional patterns visible: the South, Midwest, and Mountain West went Republican; the West Coast and Northeast went Democratic.

Before building a model, we tested whether overall fast food density differed significantly between red and blue states. It didn't. The independent samples t-test returned a t-statistic of -0.103 and a p-value of 0.919. The Pearson correlation between density and vote margin was r = 0.139 with p = 0.337. No signal. The KDE plot below confirms it — the distributions for red and blue states almost completely overlap.

red = merged_df[merged_df['vote_diff'] > 0]['restaurants_per_100k']
blue = merged_df[merged_df['vote_diff'] < 0]['restaurants_per_100k']

t_stat, p_val = ttest_ind(red, blue)
r, r_p = pearsonr(merged_df['restaurants_per_100k'], merged_df['vote_diff'])

# Results:
# t = -0.103, p = 0.9187  → no significant difference between red and blue states
# r =  0.139, p = 0.3366  → no significant linear correlation with vote margin

Total density doesn't predict voting. But when we looked at individual chains, the picture changed entirely.

We built a logistic regression classifier where each state is one data point, the features are per-capita restaurant counts for all 11 chains, and the label is the 2024 winner. With only 50 states, cross-validation was essential. We ran 50 randomized 60/40 train-test splits to get a reliable accuracy estimate.

state_votepct['winner'] = state_votepct.apply(
    lambda row: 'red' if row['pct_trump'] > row['pct_harris'] else 'blue', axis=1
)

# Normalize each chain per 100k residents
for col in food_df.columns:
    model_df[col] = (model_df[col] / model_df['TotalPop']) * 100_000

X, y = model_df[food_df.columns], model_df['winner']

accuracies = []
for seed in range(50):
    X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.4, random_state=seed)
    clf = LogisticRegression(max_iter=1000).fit(X_tr, y_tr)
    accuracies.append(clf.score(X_te, y_te))

# mean(accuracies) ≈ 0.81
# one-sample t-test vs. 0.5 → t = 28.48, p < 0.000001

Averaging the logistic regression coefficients across all 50 runs reveals which chains are actually doing the predictive work. Positive coefficients indicate association with Republican voting; negative with Democratic.

• Republican-associated: Arby's, Wendy's, Chick-fil-A, KFC, Domino's — all significantly higher per-capita density in red states (p < 0.05 on t-tests)
• Democratic-associated: Dunkin' Donuts, Starbucks — significantly higher per-capita density in blue states
• No significant association: McDonald's and Chipotle — broadly distributed across both partisan leanings

State-level aggregation obscures a lot. Urban and rural areas within the same state vote very differently, and they have very different fast food environments too. Texas as a single data point papers over the gap between Houston and the Panhandle. A county-level analysis would likely surface stronger and more nuanced patterns — but matching county-level restaurant counts with county-level votes was outside the scope of this project.

The Modifiable Areal Unit Problem (MAUP) is real here: the patterns we see at the state level might look completely different at the county or ZIP code level. We're also working with 50 data points, which is enough to identify patterns but not enough to be confident about generalizability. Fast food density is shaped by income, urbanization, and regional infrastructure in ways that may fully explain the voting correlation without any direct causal relationship.

We're identifying correlations, not causes. The finding that Arby's per-capita count predicts Republican voting probably says more about the rural South and Midwest — where Arby's happens to cluster — than it does about Arby's specifically. That said, the signal is statistically significant and consistent across methodologies. Whether it's about the chains themselves or the places those chains call home is a question worth asking.