CENSUS-26 · HOW IT WORKS
The model reads 22 numbers about who lives in each district, learns from past House elections how those numbers map to votes, and adds a national mood taken from history. Here is each piece, and a version you can play with.
01 What it sees
For every district the model reads the American Community Survey: education, race and origin, age, marriage, income, home values, how people get to work, how dense the place is. It learns from 2012–2024 House results how those numbers map to votes, and then predicts each district from its profile alone.
The single strongest signal is the share of white adults without a college degree: the more of them, the more Republican the district. Marriage comes next, then how many people ride public transit, a stand-in for how urban a place is. Income on its own barely matters once those are known.
Pick a variable: every dot is a district, placed by that census number and by how it voted for the House in 2024.
02 Twin districts
The model's bet is that places with the same census profile vote alike, wherever they are. These pairs of districts, in different states and regions, have almost identical profiles.
Hundreds of miles apart, different states, different candidates: the census profile alone tells you how they vote.
| White adults without a degree | 63.5% | 63.0% |
|---|---|---|
| Bachelor's degree or higher | 28.2% | 30.0% |
| Married adults | 49.6% | 50.7% |
| Hispanic | 3.5% | 2.9% |
| Black | 5.1% | 4.1% |
| Median age | 43.0 years | 42.9 years |
| People per square mile | 323 | 199 |
| White adults without a degree | 66.9% | 66.4% |
|---|---|---|
| Bachelor's degree or higher | 23.6% | 25.9% |
| Married adults | 51.4% | 51.5% |
| Hispanic | 5.5% | 4.3% |
| Black | 3.8% | 2.4% |
| Median age | 42.3 years | 42.4 years |
| People per square mile | 148 | 109 |
| White adults without a degree | 39.3% | 37.7% |
|---|---|---|
| Bachelor's degree or higher | 36.0% | 34.4% |
| Married adults | 43.8% | 43.6% |
| Hispanic | 8.8% | 11.8% |
| Black | 21.2% | 21.8% |
| Median age | 38.1 years | 36.8 years |
| People per square mile | 2,432 | 1,829 |
The exceptions: the profile matches but the vote does not. Local history, a candidate or a state party the census cannot see.
| White adults without a degree | 62.1% | 65.2% |
|---|---|---|
| Bachelor's degree or higher | 26.8% | 28.5% |
| Married adults | 51.5% | 49.9% |
| Hispanic | 10.4% | 3.6% |
| Black | 1.9% | 1.1% |
| Median age | 38.6 years | 39.5 years |
| People per square mile | 37 | 66 |
| White adults without a degree | 44.4% | 45.5% |
|---|---|---|
| Bachelor's degree or higher | 36.4% | 32.8% |
| Married adults | 51.2% | 52.1% |
| Hispanic | 11.6% | 11.8% |
| Black | 11.6% | 15.2% |
| Median age | 40.2 years | 40.4 years |
| People per square mile | 211 | 409 |
03 Play the model
Move the sliders and the census tells you how your imaginary district would vote. Start from a real one, then change it: more graduates, fewer married couples, a denser city, another region.
The census would say
R+40.2Safe RepublicanVA-09 actually voted R+45.3 in 2024, a year when the country voted R+2.9.
Starting point: national mood R+2.9, region R+9.0.
A simplified, linear version of the model with eight variables. It misses the 2024 districts by 14.3 points on average; the full model, by 9.6.
04 The national mood
Demographics say which districts lean which way, not how big the wave is. Without polls, the national mood comes from history alone: start from the 2024 House popular vote and apply the average midterm swing against the president's party. That swing has ranged from a small gain to a 17-point loss, so the uncertainty is wide on purpose.
05 The engine
Two learners, one formula, 20,000 simulated elections.
Each district's margin is the national House vote plus a structural term learned from its census profile and region. The structural term averages two learners: gradient-boosted trees, and a ridge regression whose leftovers are passed to a second set of trees. 20,000 simulations draw one national error shared by every seat and an independent error per district.
A state's margin is the average of its districts' census margins; single-district states are their district. Weighting districts by votes cast was tested and tied with equal weights. The state error is measured on every two-party Senate race since 2014 and the same national error is shared with the House.