#+TITLE: Virginia 2026 amendment results: who crossed party lines
#+CREATED: 2026-11-04

* What this is
Scripts that pull Virginia's November 2026 general election results and
estimate how Democratic (Warner) and Republican (Mizusawa) voters split on
the three constitutional amendments: abortion, same-sex marriage and
restoring voting rights to people with felony convictions. Backs the post
in [[file:posts/spot-on-backwards/outline.org][posts/spot-on-backwards]].

* Running it
#+begin_src sh
bin/fetch      # results, turnout, absentee counts; Census race + income (-E / -C for one half)
bin/analyze    # writes out/analysis.json and prints the summary
bin/mkchart    # out/crossovers.png, the dot-grid chart used in the post
#+end_src
=data/= and =out/= are regenerated and not committed. =bin/fetch= needs
=jq=, =opassist= (Census key in 1Password: "api.census.gov API key") and
Chrome for the chart. Everything else is Python's standard library.

* Sources
- Virginia Department of Elections, Election Night Reporting, 2026 November
  General: [[https://enr.elections.virginia.gov/results/public/virginia/elections/2026-November-General/reports][reports page]]. Three downloads: "Election Results" (one row
  per candidate or Yes/No per precinct), "Election Turnout" (registered
  voters per precinct) and "EnrAbsenteeRawCSV" (early and mail ballots
  returned, by locality, age and gender; no votes).
- US Census Bureau, 2020 Census Redistricting Data (PL 94-171), table P4,
  adults by race, by voting district, via api.census.gov.
- US Census Bureau, American Community Survey 2019-2023 5-year, table
  B19013, median household income by county or city.

* Data notes
- The site's page is a JavaScript app. Its JSON record is at
  =/results/public/api/elections/virginia/2026-November-General=, and
  downloads live at =/cdn/results/<jurisdictionId>/<blobName>=. Blob names
  change on every publish, and a report sometimes drops out of a publish
  for a few minutes; =bin/fetch= keeps the previous copy when that happens.
- The results CSV is Latin-1, not UTF-8.
- *Early votes are missing for many localities.* As of the Nov 4 morning
  publishes, about 48 localities' totals look complete (half or more of
  registered voters), while about 72 look Election-Day-only (Arlington,
  Alexandria, Loudoun, Prince William, Chesterfield, Virginia Beach and
  others), with returned absentee ballots roughly equal to everything
  reported. =bin/analyze= reports both the all-precinct and the
  complete-localities-only numbers. The crossover rates barely differ;
  the statewide percentages do.
- Precincts are matched to 2020 Census voting districts by locality and
  simplified name, not by number: numbers shifted in several counties
  (all of Henrico is off by one). About 75% of precincts match.

* Method
Two-group (Goodman) ecological regression across precincts with at least
100 Senate votes, weighted by Senate votes: a precinct's yes share is
modeled as Republican-yes-rate x (1 - Warner share) + Democratic-yes-rate x
Warner share. The 95% ranges come from resampling whole localities. These
are estimates from precinct totals, not individual ballots, and assume each
party's voters behave about the same everywhere. That assumption holds
reasonably for abortion and voting rights and breaks for marriage, which
varies sharply with a precinct's Black share.

The analysis was built with Claude (Claude Code) and checked by an
independent Claude agent, ChatGPT (gpt-5.5) and Gemini; see the Nov 4
session.