Actions Republicans Should Take
Precinct 523 · Princeton, Maine
These numbers come from a computer, not from asking people. In June 2026 we asked AI stand-ins, not real voters, how they would vote. Each stand-in is built from public census and voter records. No real votes were counted, nobody was asked, and this does not predict the election.
The recommendation
| Higher-leverage move | Persuade |
|---|---|
| Priority vs all polled precincts | 88 / 100 |
Many undecideds — name the choice and own the framing first. Introduce the candidate around a single clear promise (affordability, safety) and make the contrast concrete.
This is an example, not advice. No campaign has tested it or approved it. We work it out by adding up the four things shown below: how close the race is, how many people have not made up their mind, how many voters live here, and how many of them usually stay home. Nothing is hidden. You can follow the same steps yourself.
Why — the four factors
Each is this precinct's percentile against every polled precinct in the country. Priority rises with a bigger, more competitive electorate that has a large undecided pool (persuade) or a favourable but low-turnout base (mobilize).
What the simulated poll says
| Democratic share (two-party) | 42.3% |
|---|---|
| Republican share (two-party) | 57.7% |
| 95% confidence interval (Dem two-party) | 29–55% |
| Undecided | 4.5% |
| Simulated personas sampled | 62 |
See this precinct on the live map →
Why the model expects this result
A typical precinct in this model sits at 65.3% Democratic on the two-party vote. This precinct's neighbourhood traits move that to 45.7%.
What pushes it toward the Republicans: % non-Hispanic White, % limited English and % mobile homes.
It projected came out at 42.3%, close to what the traits suggest.
See these numbers as a table
| Fact | Running guess | Change (points) |
|---|---|---|
| Starting point: a typical area | 65.3% | — |
| % non-Hispanic White | 61.4% | -4.0 |
| % limited English | 58.0% | -3.4 |
| % mobile homes | 54.9% | -3.1 |
| % without internet | 52.4% | -2.5 |
| % uninsured | 49.9% | -2.5 |
| Average annual pay | 47.5% | -2.3 |
| All other facts together | 45.7% | -1.8 |
| What the model expects here | 45.7% | — |
- % non-Hispanic White — Racial makeup — white, non-Hispanic residents.
- % limited English — Language — households with limited English.
- % mobile homes — Housing type — mobile/manufactured homes (rural marker).
- % without internet — Digital access — households without internet.
- % uninsured — Health-insurance coverage gap.
- Average annual pay — Local wages — average annual pay.
Across all precincts this model explains about 40% of the variation in results, and is typically off by 10 points for any one precinct. The model only knows facts about the area, such as age, pay and housing. It knows nothing about the people standing for election, what they said, what mattered locally, or who turned out to vote. That is why the real result can be a long way from its guess. These are patterns across whole areas. They do not tell you how any one person voted, and they do not show that one thing causes another. When two things go together, such as low pay and low rents, the model cannot tell them apart, so it shares the credit between them.
How does this area compare with the rest of the state?
One row for each fact. Every dot is an area in this state, placed by how far that fact moved the guess there. This area is the ringed dot.
See this area's numbers as a table
| Fact | Value here | Moves the guess by | Towards |
|---|---|---|---|
| % non-Hispanic White | — | -4.0 points | Republicans |
| % limited English | 0.4% | -3.4 points | Republicans |
| % mobile homes | 11.4% | -3.1 points | Republicans |
| % without internet | 18.3% | -2.5 points | Republicans |
| % uninsured | 10.5% | -2.5 points | Republicans |
| Average annual pay | $50,219 | -2.3 points | Republicans |
Every other dot is one of the 564 areas in Maine that the same computer looked at. For each fact we only show the areas where it was one of the six that mattered most, so you are seeing that fact where it counted, not everywhere.
How the model was built
We give a computer about 20 facts about each area and ask it to guess how the area voted. It is never shown any votes. The facts are things like age, race, pay, jobs, health cover and the kind of homes people live in. They come from the 2020 Census and from other government figures.
To check the guesses are fair, we train the computer on some states and then test it on states it has never seen. That way it cannot simply remember the answer.
Each bar shows how much one fact moved the guess up or down for this area. Some facts are measured for the whole county, so they are the same for every area in that county. Where two neighbouring areas differ, it is their own 2020 Census figures that pull them apart.
For the technically minded: the model is gradient-boosted trees (XGBoost), the target is the two-party Democratic share, the bars are SHAP values, and it is validated by holding out whole states. County-level inputs include the Census/CDC Social Vulnerability Index 2018–2022, County Business Patterns and wage data.
Who lives here — 2020 Census
| Population | 846 |
|---|---|
| Voting-age population (18+) | 677 |
Counted exactly for this precinct: every 2020 census block inside the precinct boundary, summed. Not an estimate.