Precinct Precinct

Three weeks of political research, done in an afternoon.

Precinct is an AI political scientist. Ask it a question about elections or public opinion. It reads the published research, builds the dataset, and hands back an estimate with the error bars attached.

Real result

Fig. 1

Did same-day registration raise turnout among 18-to-24-year-olds?

Yes, by 6.5 percentage points (90% interval 3.9 to 9.0). A published study of the same data reported 3.1. Its standard two-way fixed effects method is known to misstate effects when states adopt a law in different years. A method built for that case gives twice the estimate.

Two estimates of the same effect, with 90% intervals Published method 3.1 Same data, staggered-adoption method 6.5 0 5 10 points

Change in turnout among 18-to-24-year-olds where same-day registration took effect, with 90% intervals. Current Population Survey, 1980 to 2018.

Also caught. The county election file most analysts use puts Jackson County, Missouri's 2020 turnout at 37% of citizens of voting age. Kansas City's votes are filed outside the county. With them, turnout is 63%.

Track record. Run five weeks before each election with only what was public then, Precinct's forecasts picked the winner in 99 of 104 Senate races from 2020 to 2024. The actual seat count landed inside the forecast range in 8 of 8 tests.

Estimate6.5 percentage points
90% interval3.9 to 9.0
Published estimate3.10, reproduced to the printed digit
SourcesGrumbach and Hill (2022) replication data; Current Population Survey voting supplement

How it works

1

Reads the research

Finds the published studies, polls, and past estimates on your question, and shows where they disagree and why.

2

Builds the dataset

Combines election returns, census data, and survey microdata into one table. Every merge and every dropped row is logged, so a colleague can check the work.

3

Reports the range

Every estimate comes with an interval and its sources. If the data can't answer the question, Precinct says so rather than guess.

Who it's for

Research organizations

Check a claim against the full literature before it goes into a report.

Newsrooms

Get a defensible number on election night, with the caveats your editor will ask about.

Campaigns

Test which explanation of the last cycle holds up before you spend on the next one.

3,240federal campaigns spent money in the 2023-24 cycle.

$60,883spent by the median one, for everything.

$262 millionpaid for polling, surveys, and research, at the least.

Precinct's count of Federal Election Commission filings. Federal races only; research billed under other labels is not included.