Precinct-level targeting for the proposed half-cent regional transit sales tax across San Francisco, Alameda, Contra Costa, San Mateo and Santa Clara. The support model is not a guess — it is fitted on an actual 2026 Bay Area countywide sales tax that appeared on the June ballot in one of these five counties, then applied to all 2,214 precincts.
Sources: UC Berkeley Statewide Database precinct file (Nov 2024 boundaries, registration, 2024 results); Contra Costa County Registrar of Voters Statement of Vote, June 2 2026; American Community Survey 2019–2023. Analysis by VoterForce.
A majority-threshold measure is decided by net votes, not ballots. Every 100 additional ballots cast in a precinct that supports the measure 65–35 is worth +30 net votes. The same 100 ballots in a 40–60 precinct is worth −20. So turnout is not a good in itself — it is only good where support already is. The Mobilisation value overlay sizes each precinct by exactly that quantity.
| Tier | Precincts | Registered | Projected ballots | Support | Net votes / 100 ballots |
|---|
In June 2026 a Bay Area county inside this five-county footprint put a countywide sales tax on the ballot. It lost, 43.1% to 56.9%, across 656 precincts. That result is the training set: it is a real, recent, precinct-level Bay Area verdict on a sales tax, in the same electorate and the same economy.
Of every variable tested, the strongest predictor by a distance was each precinct's 2024 vote on Proposition 5 — the statewide measure that would have lowered the approval threshold for local infrastructure bonds to 55%. It correlates with the actual sales-tax vote at r = +0.96. Adding Prop 4 (the climate bond) and Prop 36 (the crime measure, which enters negatively) produces a model with R² = 0.928 and a cross-validated error of 3.1 points per precinct.
Partisanship, by contrast, is a materially worse predictor: presidential vote share alone yields R² = 0.63 and roughly double the error. Precincts do not vote on a sales tax the way they vote for president. They vote on it the way they voted on Prop 5.
| Predictor of the real 2026 sales-tax vote | Correlation | Model R² | CV error |
|---|---|---|---|
| Prop 5 2024 — local infrastructure bond threshold | +0.96 | 0.914 | 3.33 pts |
| + Prop 4 (climate bond) + Prop 36 (crime) | — | 0.928 | 3.09 pts |
| Prop 2 2024 — school bond | +0.92 | — | — |
| Presidential vote share 2024 | +0.79 | 0.629 | 6.96 pts |
| Median household income | −0.57 | — | — |
| Renter share of households | +0.53 | — | — |
| Share commuting by transit | +0.50 | — | — |
| Share of registrants 65+ | +0.36 | — | — |
Raising turnout by a full 10 points across every base precinct in all five counties — roughly 99,000 additional ballots, a program far beyond what any regional campaign has executed — moves the result +0.69 points. Two points of persuasion moves it +2.00. Persuasion is close to three times the lever that mobilisation is, at every scenario level tested. A plan that spends its budget on base GOTV is optimising the smaller number.
On the real June 2026 sales tax, the campaign's strongest precincts did turn out lowest and the opposition's highest, exactly as the conventional read has it. But re-weighting that same electorate to presidential-turnout shape changes the result by 0.06 points, and to perfectly uniform turnout by 0.74. The differential-turnout effect is real and directionally adverse, and it is close to numerically irrelevant. The measure failed on support, not on who showed up.
At the current 56% poll the model puts San Francisco at 66% and Alameda at 60%, but Santa Clara at 51.0%, Contra Costa at 51.4% and San Mateo at 53.0%. Those three carry 60% of the projected ballots. The regional total is comfortable; the three counties that decide whether it is comfortable are not. A uniform 3-point slip — less than one standard poll-to-ballot decay — puts all three under water and the region at 53%.
The model transfers relative support with high confidence: it ranks precincts, and on the one election we can check it ranks them almost perfectly. It cannot set the level. The level here comes from published polling, and the training measure also polled above 50% before losing at 43%. Every tier of that measure fell 11–18 points from a near-identical measure six years earlier. If the 2026 sales-tax environment repeats, no targeting plan recovers it.
Transit commute share correlates +0.50 on its own, but once Prop 5 is in the model it adds nothing (t = 1.2). Riders are not a distinct constituency in this data; they are inside the general public-goods coalition. The caveat cuts both ways — the training county has low transit use, so its precincts cannot test the high-ridership range that San Francisco and inner Alameda occupy. Treat "target the riders" as unproven, not disproven, and worth a survey before it becomes a strategy.
This is the single largest methodological risk. It is mitigated but not eliminated: only 3 of 1,717 precincts outside the training county fall outside its Prop 5 range, so the model interpolates rather than extrapolates. But San Francisco's density, tenure mix and transit dependence have no real analogue in the training county, and a county-level fixed effect — a systematic San Francisco or Santa Clara offset — is exactly the kind of error a single-county fit cannot see. Treat the San Francisco and Santa Clara levels as the least certain numbers on this page.
A 5-point opposition turnout surge concentrated in the hardest No precincts costs the measure 0.11 to 0.25 points depending on scenario. That is reassuring, and it is the assumption most likely to be wrong: it prices organised opposition as a turnout effect only. A funded No campaign works through persuasion in the middle tier, where 1.32M registrants sit at 51% support — and the model has no way to price that.
Precincts that voted for tougher criminal sentencing vote against the sales tax, holding the bond axis constant. For a measure funding BART and Muni this is the one substantive interaction the data hints at, and it points at the safety-and-disorder critique rather than the price. It is a small term (t = −3.2) and should be read as a hypothesis to poll, not a finding.
Universe. 2,214 precincts with 200 or more registered voters across the five counties — 3.66M registrants, 99.2% of the regional roll. Precincts under 200 registrants are excluded: the county files are full of small mail-ballot pseudo-precincts that produce meaningless rates and dominate any ranking.
Support. Weighted least squares on the June 2026 countywide sales-tax result, 496 precincts with 30+ votes, weights normalised to mean 1 so the standard errors are honest. Predictors: Prop 5, Prop 4 and Prop 36 yes-shares from November 2024. Fitted values are shifted in log-odds so that the ballot-weighted regional projection equals the scenario level you select — the model supplies the shape, the scenario supplies the level.
Turnout. Log-odds regression of June 2026 precinct turnout on November 2024 turnout (R² = 0.85), re-anchored to a 62%-of-registered November 2026 regional turnout. Because the composition effect measured above is near zero, the projection is insensitive to this anchor; it changes the ballot counts, not the percentages.
What is not in here. No prior transit-measure results (none of the recent Bay Area transit measures share these precinct boundaries), no survey data, no modelled individual-level scores. This is precinct geography and public election returns.