Campaign Strategy
The project replaces blanket battleground spending with a county-level targeting logic. Instead of asking which state matters, it asks where the next marginal dollar is most likely to move votes.
250
Counties analyzed
3
Swing states studied
0.718
Best binary F1 (RF)
0.195
LOSO F1 across states
0.552
Diversity threshold
Plain-English Conclusion
The main takeaway is simple: campaigns should not treat battleground counties as interchangeable. A county profile that signals volatility in Pennsylvania does not necessarily behave the same way in Michigan or North Carolina, which is why pooled cross-state prediction breaks down so quickly.
Inside each state, though, a clear pattern emerges. Counties that are racially mixed, changing quickly, and often under economic stress are much more likely to move, which makes a state-specific persuasion ranking more useful than a single national template.
That is the role of the final PersuasionScore: it combines who looks persuadable, how much vote volume is at stake, and how likely the county is to behave like a swing county again.
Why It Matters
Civic & Media Value
The results also help explain why some counties deserve far more attention than others. They give a more nuanced map of competitive geography than statewide polling or winner-take-all narratives alone.
Why Further Funding Helps
The framework is already useful at county level, but it gets stronger with more states, better turnout data, voter-file enrichment, and repeated recalibration for future election cycles.
Research Question Explorer
Click any item below to see how the different model families answered the project’s original research questions and hypotheses.
-
Our final answer is the PersuasionScore, not the earlier ranking variants. In the final scoring update, predicted volatility carries half the total weight, while the other half is split across within-state electoral weight and a transparent persuadability blend built from diversity, poverty, inverse income, and population density. The storytelling spotlight starts with Philadelphia County, Oakland County, and Halifax County, but the official rankings remain state-specific because the score is normalized within state.
Classification Clustering Final composite scorePennsylvania leader Philadelphia County
PS = 0.838Large urban persuasion/turnout hybrid target with meaningful predicted volatility.
Michigan leader Oakland County
PS = 0.797High-volume suburban transition county with one of the strongest volatility signals in the state.
North Carolina leader Halifax County
PS = 0.697Smaller county, but the strongest volatility signal in the state pushes it to the top of the updated score.
PersuasionScore spotlight: the most strategically attractive counties are different in each state, even when they share similar battleground labels. -
Raw volatility highlights where margins moved the most, not necessarily where the next campaign dollar is most efficient. North Carolina dominates the extreme top of the volatility scale, while Pennsylvania and Michigan contribute fewer but still important high-swing counties. Once vote volume enters the picture, high-impact counties like Wayne, Philadelphia, Oakland, Chester, Montgomery, and Bucks matter much more than a raw-volatility list alone would suggest.
Temporal analysis Volatility engineering Election returnsPennsylvania
County Vol. Monroe 3.41 Pike 2.94 Lackawanna 2.32 Michigan
County Vol. Wayne 2.39 Leelanau 2.26 Grand Traverse 2.10 North Carolina
County Vol. Robeson 8.03 Anson 4.46 Scotland 4.00 High-impact reminder: the counties with the largest raw swings are not always the same counties that matter most strategically. Wayne County and Philadelphia County are the clearest examples of volatility becoming much more consequential once vote volume is considered. -
The cluster story is not random noise. The clearest unstable archetype is a poor, racially diverse rural profile concentrated in North Carolina, while affluent exurban and aging rural counties look substantially more stable. SHAP-based archetypes reinforce the same conclusion from a different angle: volatility comes from a small number of recognizable county types, not one universal statewide recipe.
K-means clustering SHAP archetypes ClassificationHigh-Poverty Rural
NC-only concentration, about 21% poverty, small rural population, and the highest average volatility.
High-Diversity Urban
The largest and most educated cluster; volatility is not always extreme, but it stays meaningfully elevated.
Affluent Exurban
High income, high homeownership, lower volatility, and fewer signs of rapid political movement.
Aging Rural
Older, more homogeneous counties with the lowest average volatility and the most stable voting patterns.
Model-explanation clustering reveals five functional county archetypes and shows why the North Carolina poor-and-diverse pattern does not transfer neatly to other states. -
This hypothesis is supported overall, but with clear state nuance. Lower-income and housing-pressure features repeatedly appear among top predictors, while poverty is especially decisive in North Carolina. The broader lesson is that economic stress matters, but the pathway differs by state: poverty dominates in North Carolina, rent pressure matters heavily in Pennsylvania, and Michigan is more education-driven.
Regression Classification Cross-method convergenceState Best Regression R2 What the model says Pennsylvania 0.674 Rent, unemployment, and diversity matter most in the strongest state model. Michigan 0.331 Education and transition features matter more than pure income measures. North Carolina 0.435 Poverty is the clearest statewide volatility signal. -
Yes, and this is the strongest single result in the project. The normalized racial diversity measure
race_entropy_normrepeatedly outruns individual race-share variables in permutation importance, SHAP analysis, threshold analysis, and interaction analysis. Crossing the 0.552 threshold makes a county about 7.15 times more likely to be highly volatile, especially when diversity and poverty appear together.SHAP Threshold analysis Strongest finding
The diversity switch: above race_entropy_norm = 0.552, counties become far more likely to land in the high-volatility bucket. -
The temporal evidence points in that direction. Volatile counties usually move with broader statewide or national direction rather than inventing a completely separate path: Blue Bounceback is common, Steady Red Drift is the most volatile, and true red-to-blue bouncebacks are extremely rare. That makes these places look more like amplifiers of larger trends than isolated anomalies, although this remains an interpretation built from trajectory evidence rather than a standalone formal test.
Temporal analysis Trajectory typingTrajectory type Counties % High volatility Blue Bounceback 117 34% Steady Red Drift 94 50% Steady Blue Drift 35 37% Red Bounceback 4 0% -
No. Leave-One-State-Out validation collapses to near-random performance, which means a model trained on two states struggles to predict the third. Pennsylvania behaves like a diversity-and-rent story, Michigan looks more like an education-and-transition story, and North Carolina is much more poverty-driven.
LOSO validation Regression State signatures
Cross-state failure is substantive, not cosmetic: pooled models do not carry cleanly from one swing state to another. Pennsylvania
Diversity and housing pressure matter most.
Michigan
Education and transition features dominate.
North Carolina
Poverty-driven volatility is the clearest pattern.
-
The swing map has some continuity, but it is not just the same counties repeating the same move. The most common pattern is Blue Bounceback, where counties swung toward Democrats in 2020 and snapped back in 2024. The most volatile pattern is Steady Red Drift, and North Carolina stands out because nearly half of its counties follow that path.
Temporal analysis Election trajectories
County trajectories show that 2020 often behaved like a temporary Democratic surge, especially outside the most stable urban strongholds. -
This question was not fully completed in the official milestone pipeline. We tracked vote volume and margin movement, but we did not finish a turnout-variance analysis that was strong enough to publish as a core finding. It remains an important next step, especially if the project expands to precinct-level data or voter files.
Future work Data expansionWhat would strengthen this answer? Better turnout-specific data, finer-grained geography, and a dedicated turnout model separate from the volatility pipeline used in Milestone 4.
PersuasionScore Atlas
This is the final milestone-4 deliverable. The updated PersuasionScore uses three ingredients chosen for clarity and interpretability, with state-specific predicted volatility now carrying the most weight.
Persuadability
Linear blend of racial diversity, poverty, inverse income, and population density. This is the transparent answer to who looks movable.
Electoral Weight
A softened version of within-state vote share. This keeps large counties important without letting them completely dominate the score.
Predicted Volatility
State-level out-of-fold Random Forest probability of high volatility. This is now the largest component, so repeatable swing behavior matters most in the final ranking.
Philadelphia County
PS = 0.838
Large urban persuasion/turnout hybrid target.
Oakland County
PS = 0.797
High-volume suburban transition county.
Halifax County
PS = 0.697
High-volatility rural persuasion test case.
Technical note: PersuasionScore is normalized within state, so the tables below are the
official rankings. The combined three-state view is a spotlight, not a strict national leaderboard.
| Rank | County | PersuasionScore | P(high volatility) | 2024 D margin | Why it matters |
|---|---|---|---|---|---|
| 1 | Philadelphia County | 0.838 | 0.744 | +58.8 pp | Large urban persuasion/turnout hybrid target |
| 2 | Allegheny County | 0.800 | 0.849 | +20.3 pp | High-volume defensive hold with swing risk |
| 3 | Delaware County | 0.736 | 0.867 | +23.7 pp | High-volume suburban transition county |
| 4 | Montgomery County | 0.690 | 0.750 | +22.8 pp | Affluent suburban transition county |
| 5 | Lehigh County | 0.633 | 0.748 | +2.7 pp | Bilingual outreach bellwether county |
| Rank | County | PersuasionScore | P(high volatility) | 2024 D margin | Why it matters |
|---|---|---|---|---|---|
| 1 | Oakland County | 0.797 | 0.880 | +10.6 pp | High-volume suburban transition county |
| 2 | Wayne County | 0.793 | 0.700 | +29.0 pp | Large urban persuasion/turnout hybrid target |
| 3 | Washtenaw County | 0.720 | 0.919 | +44.4 pp | Education-driven turnout anchor |
| 4 | Kent County | 0.712 | 0.857 | +5.4 pp | Suburban transition county |
| 5 | Ingham County | 0.654 | 0.795 | +29.7 pp | College-centered persuasion and mobilization target |
| Rank | County | PersuasionScore | P(high volatility) | 2024 D margin | Why it matters |
|---|---|---|---|---|---|
| 1 | Halifax County | 0.697 | 0.954 | +17.7 pp | High-volatility rural persuasion test case |
| 2 | Wake County | 0.689 | 0.610 | +25.4 pp | High-volume metro persuasion/turnout hybrid target |
| 3 | Lenoir County | 0.688 | 0.918 | -6.8 pp | Diverse rural warning signal |
| 4 | Washington County | 0.672 | 0.961 | +6.2 pp | Small-county volatility hotspot |
| 5 | Richmond County | 0.669 | 0.887 | -20.9 pp | Diverse rural warning signal |
Conclusion, Limitations, and Future Work
Non-Technical Summary
The project shows that county-level persuasion opportunity is real, but only when we respect local context. The strongest counties for future Democratic attention are not the same in every swing state, and the best signals often come from diversity, economic stress, housing pressure, and state-specific political realignment rather than from a single national battleground formula.
Key Discoveries and Real-World Impact
- Binary framing worked far better than five-class framing, which tells us that identifying high vs. low volatility is more actionable than over-classifying the middle.
- The diversity threshold at
race_entropy_norm = 0.552is the clearest empirical finding in the project. - Cross-state transfer fails, which means campaign strategy should be tailored to each state rather than copied across the map.
- The final PersuasionScore translates the research into a county-level targeting product that can support social sharing, strategic planning, and future funding conversations.
Limitations
What this page does not claim
- The sample is still small at 250 counties across three states.
- Cross-state transfer failure means the model is not a universal battleground engine.
- PersuasionScore is a county-level targeting aid, not a causal model of voter behavior.
- The turnout-variance question was not fully completed in the official pipeline.
- We do not have voter-file or precinct-level personalization.
- Demographic data can lag fast real-world population change.
Future Work
What stronger funding would unlock
- Expand the framework to more swing states.
- Add precinct-level or voter-file enrichment.
- Build a dedicated turnout-targeting extension instead of stopping at persuasion.
- Recalibrate the score for each new election cycle.
- Test whether the county signals remain stable as the 2028 environment takes shape.
Bottom line: the project moved from descriptive analysis to a practical targeting framework.
That is why the final deliverable is a persuasion map, not just a model scorecard.