What decided the El-Sayed–Stevens primary

Section one

What predicted the margin, precinct by precinct

Each dot is one precinct; dot area is votes cast. The vertical axis is El-Sayed's margin over Stevens throughout.

Axes are trimmed to the 1st–99th percentile of each predictor so a few extreme precincts don't compress every panel; all statistics use the full data. Colour doubles the vertical position and carries no information on its own.

Section two

A reversal hiding inside the 2020–2024 swing

Pooled across every precinct, El-Sayed ran better where Democrats lost ground (r = −0.25). Among the 96% of precincts at or under 10% Arab ancestry the sign flips to +0.18 — he ran better where Democrats gained. The pooled figure is a Simpson's paradox produced entirely by 115 precincts holding 4.5% of the vote.

Democratic two-party swing, 2020 → 2024  (left = moved toward Republicans)

x-scales differ by panel. Each subset spans a very different swing range and a common scale would collapse the middle panel into a sliver. Compare slopes and bin lines, not panel widths.

Section three

Where a precinct's presidential winner changed, so did the primary

The two sequences that broke a precinct's pattern point to two different coalitions moving in opposite directions. Dem → Dem → Rep is El-Sayed's strongest ground at 67.8%, and is overwhelmingly Dearborn and Dearborn Heights (mean Arab ancestry 30%, against 2% statewide). Rep → Rep → Dem, the reverse move, is his second-strongest at 53.0% — West Michigan and Traverse City, 49% with a bachelor's degree. Stevens led only in the two most Republican sequences.

No precinct produced Dem → Rep → Dem, so seven of eight sequences appear. Shares do not sum to 100 — Mallory McMorrow took the remainder. Dem → Rep → Rep rests on 9 precincts; read its interval with care. Note also that Rep → Rep → Dem and Rep → Dem → Rep are largely threshold crossings: 95% of each sat within three points of 50% in the pivotal year. Dem → Dem → Rep is not — it moved 24 points.

Section four

Education and income pulled in opposite directions

Read down a column — income held roughly fixed, education rising — and El-Sayed's margin climbs. Read across a row — education held fixed, income rising — and it falls. Among the most educated precincts he wins the poorest band by 50 points and the richest by 3. These are raw cell means with no model imposed.

Education and income correlate at +0.77, so the off-diagonal corners are thin: cells under 15 precincts are greyed and the low-education / high-income corner is empty altogether. That sparsity is why the linear income coefficient is fragile, and it is worth showing rather than hiding. The bottom-left cell is mostly heavily-Black precincts, where race rather than income drives the result — matching on race and age as well shrinks the income gap from about 21 points to about 12, but does not remove it.

Section five

Who showed up: young voters closed a quarter of the turnout gap

The age skew of August-primary turnout barely moved across four Trump-era elections, then broke in 2026. But it broke from a very low base: the youngest fifth of precincts still turned out 11 points below the oldest. Everyone voted more in 2026 — the young simply rose faster, 1.41× against 1.15×.

Age here is a precinct characteristic — the share of its registrants aged 18–34 — not a voter-level measure. Voter-level rates by age bracket exist for 2024 only, and they are stark: registered 65–74s voted at 47.5% against 8.0% for 20–24s, a six-fold gap among people already registered.

Turnout is each year's voters over one fixed denominator (registered voters in the 2024 L2 file), so the series is internally comparable; the levels are not the rates reported at the time. L2 credits vote history to a voter's address at the file snapshot rather than at the election, so attribution loosens for the earliest years — 2018 is the earliest year worth reading. The all-precinct ratio runs higher than the Democratic-leaning one only because 2026 counts Democratic ballots and young precincts lean left.