The wager
Large states have more potential movers. Nearby states impose less cost, risk, and friction. Therefore migration should rise with population and fall with distance.
The test
A useful model is not one that never fails. It is one whose failures teach us what the model left out.
01
Make a gravity-model prediction
The constant k is calibrated so total predicted migration equals the total of the published state flows. You decide how strongly distance should deter movement.
Who should send the most migrants?
The five leading predictions under your current distance assumption.
Pause before looking: Which states do you expect the model to badly underestimate? Write down one guess and a reason.
02
Test prediction against observation
Every dot is an origin state. Dots on the diagonal are predicted accurately; dots above it sent more migrants than expected, and dots below it sent fewer.
03
Map the model's failure
A scatterplot tells us how wrong the model is. A map asks whether the errors form a geographic pattern.
far belowbelowcloseabovefar abovenot numeric
What pattern do you see?
The strongest positive residuals cluster around Minnesota's neighbors, especially Wisconsin and the Dakotas. Distance matters, but a continuous regional system can matter more than distance alone.
Faraway California and Washington also outperform the simple model. This does not prove why. It gives us a geographical question worth investigating.
04
Repair the model—carefully
Add one crude “neighbor effect.” If the fit improves, you have evidence that a missing regional relationship matters. You have not yet explained that relationship.
The model improves. Have we explained migration?
Choose the strongest conclusion.
Exit prompt
Choose one state the model predicts poorly. Name one additional variable that might improve the prediction, explain how you would measure it, and state what evidence could prove your explanation wrong.
Data and model notes
Observed flows and margins of error are from the U.S. Census Bureau, 2024 American Community Survey 1-year State-to-State Migration Flows. The table measures current residence against residence one year earlier for people age 1 and over. Published estimates are sample-based. A few origin flows were not reported as numeric estimates and appear gray. State-center coordinates are approximate; the prediction model is a teaching instrument, not an official Census forecast.