PUTNAM COUNTY, INDIANA · 2026 GENERAL ELECTION
Checking a prediction against the result, and a look at the contact program behind it.
VAN assigns every registered voter a turnout score from 0 to 100, fixed early in the cycle, meant to read as their probability of voting. This checks it against what 23,570 people on the Putnam County file actually did on 3 November 2026, when 34.1% of them voted.
Those are two separate questions. Does a higher score mean a likelier voter? And when the score says 70, do 70% of them turn out? The answer here is yes to the first and no to the second.
Each bar is a 5-point band of the turnout score. Whiskers are 95% Wilson intervals. The dashed diagonal is where bars would sit if the score read directly as a turnout percentage. Every band falls below it.
Pick a voter and a non-voter at random and the voter holds the higher score 86.8% of the time, counting a tied pair as half — an AUC of 0.868 (95% CI 0.863–0.873). Turnout climbs almost perfectly in step with the score: a linear trend accounts for 96.9% of the association between the two (Cochran–Armitage χ² = 8,921 on 1 df). For sorting a universe into who is worth a knock, the score does its job.
Taken at face value the score expected 10,735 voters out of this universe. 8,037 turned out — it over-predicted by 34%. This is not noise in one or two bands: every band came in under its own score, and 18 of 20 do so beyond a 5% false-discovery rate. The gap is widest in the middle of the scale, where the 70–75 band scored 72 and turned out 51.6%.
Fitting the observed rates against the score on a log-odds scale gives a calibration slope of 0.83 (perfect is 1.00) and an intercept of -0.81 (perfect is 0.00) — a scale that is both shifted down and flattened. A two-parameter correction of that form tracks the real rates within a few points across most of the range, so the information is there; it is the labeling that is off.
| Score band | People | Voted | Score says | Actually voted | Gap (pts) | 95% CI |
|---|---|---|---|---|---|---|
| 0–5 | 3,728 | 63 | 2% | 1.7% | -0.3 | 1.3–2.2% |
| 5–10 | 2,188 | 147 | 7% | 6.7% | -0.3 | 5.7–7.8% |
| 10–15 | 698 | 31 | 12% | 4.4% | -7.6 | 3.1–6.2% |
| 15–20 | 576 | 56 | 17% | 9.7% | -7.3 | 7.6–12.4% |
| 20–25 | 984 | 107 | 22% | 10.9% | -11.1 | 9.1–13.0% |
| 25–30 | 1,340 | 166 | 27% | 12.4% | -14.6 | 10.7–14.3% |
| 30–35 | 1,183 | 206 | 32% | 17.4% | -14.6 | 15.4–19.7% |
| 35–40 | 870 | 175 | 37% | 20.1% | -16.9 | 17.6–22.9% |
| 40–45 | 683 | 149 | 42% | 21.8% | -20.2 | 18.9–25.1% |
| 45–50 | 821 | 253 | 47% | 30.8% | -16.2 | 27.8–34.1% |
| 50–55 | 759 | 267 | 52% | 35.2% | -16.8 | 31.9–38.6% |
| 55–60 | 656 | 254 | 57% | 38.7% | -18.3 | 35.1–42.5% |
| 60–65 | 616 | 268 | 62% | 43.5% | -18.5 | 39.6–47.4% |
| 65–70 | 683 | 320 | 67% | 46.9% | -20.1 | 43.1–50.6% |
| 70–75 | 1,085 | 560 | 72% | 51.6% | -20.4 | 48.6–54.6% |
| 75–80 | 1,124 | 660 | 77% | 58.7% | -18.3 | 55.8–61.6% |
| 80–85 | 1,099 | 714 | 82% | 65.0% | -17.0 | 62.1–67.7% |
| 85–90 | 792 | 560 | 87% | 70.7% | -16.3 | 67.4–73.8% |
| 90–95 | 910 | 698 | 92% | 76.7% | -15.3 | 73.8–79.3% |
| 95–100 | 2,775 | 2,383 | 97% | 85.9% | -11.1 | 84.5–87.1% |
23,556 contact attempts between 1 August and 3 November 2026, reaching 12,519 distinct people across 56 active days. Behind those attempts sit 19,884 individual phone dials and 7,851 text messages, logged by the dialer and texter themselves.
Disposition recorded for each attempt. “Left Message” is the most common at 26.8%. Results under 1% are grouped.
Door-knocking accounts for 2,313 attempts (9.8%); the rest is phone and text. Channels under 1% are grouped.
One bar per active day. The program runs in bursts, with the heaviest single day on 14 October (5,792 attempts).
The contact log records that an attempt was made. The dialer's own log records what came of it — and most of those dials never had a volunteer on them at all.
Two kinds of dial. Of 19,884 dials in the general, only 3,370 had a volunteer on the line. The other 16,514 were the dialer working on its own, and they are a different thing entirely: not one of them was ever logged as a conversation, a refusal or a wrong number, while “Dropped” and “Error” appear nowhere else. Mean duration is 47 seconds on a dial that reached someone against about two seconds on an automated one.
Mixing the two gives a pick-up rate of 6.2%. Counted over the dials a volunteer actually handled, it is 36.5%. The second number is the one that describes the work.
Each step as a share of the 19,884 dials at the top. A volunteer was on 16.9% of dials; of those, someone picked up on 36.5%, a conversation was logged on 16.4% and a survey answer was recorded on 13.5%.
Outcome logged by the caller, over the 3,370 dials a volunteer handled. 41.8% still reached an answering machine. Outcomes under 1% are grouped.
The 16,514 dials with no volunteer on them. 2,839 were dropped by the system after dialing — someone may well have answered a silent line — and 1,666 failed outright. None produced a logged conversation.
70.8% of messages were confirmed delivered by the carrier and 1,749 failed outright, most often because the number was not a mobile. 208 replies came back, 2.6% of messages sent.
30 people placed the 3,370 volunteer-handled dials, and the top five placed 63.7% of them. Hover for each caller's pick-up rate. Showing the busiest 12.
30 people knocked 2,313 doors, and the top five knocked 49.7% of them. Someone answered at 806 (34.8%), counting Canvassed, Refused and Moved as answered and Lit Dropped, Not Home and Inaccessible as not — the same rule the phone log uses, so the doorstep and the phone are measured alike. Showing the busiest 12.
Both charts label people by rank only; this edition carries no volunteer names. Note that rank order alone can identify someone on a team this small. Calls are attributed from the dialer's own log and doors from the contact history, so the two counts come from different systems: a dial the dialer placed on its own is credited to nobody, which is why 3,370 of 19,884 dials carry a name.
The same program ran through the May primary on a different universe. It placed 17,183 dials, 2,235 of them volunteer-handled, and picked up on 55.7% of those against 36.5% in the general. Texting went the other way: 408 replies to 4,702 messages, 8.7%, against 2.6% in the general.
| Primary | General | |
|---|---|---|
| Numbers dialed | 17,183 | 19,884 |
| Volunteer on the line | 2,235 | 3,370 |
| Someone picked up | 1,245 | 1,231 |
| Pick-up rate, handled dials | 55.7% | 36.5% |
| Conversations | 473 | 552 |
| Hours on the phone | 21 | 16 |
| Texts sent | 4,702 | 7,851 |
| Replies | 408 | 208 |
| People on the phones | 25 | 30 |
Every day from the first day of early voting to election day, a list of people who had already voted was loaded against the contact file. Set that against the people the program reached and you get a first read on how much of the outreach landed on people who turned out.
These counts are a floor, not a rate. The vote record on file is incomplete — 7,836 credits loaded between 6 October and 3 November, and more are still to be added. Final vote history is forthcoming. Read “no credit on file” as not yet matched to a vote, not as did not vote: some of those people did vote and are not yet recorded.
Of the 12,519 people reached, at least 4,901 voted. The remainder is an upper bound on non-voters, not a count of them.
People reached by each channel who have a vote credit. Channels reaching fewer than 100 people are left out.
This is not a measure of what contact achieved. The program chose whom to call, text and visit using the turnout score, so the people it reached were always going to vote more than the county as a whole, and the channels differ because the people assigned to them differ. Separating the effect of being contacted from the effect of being the sort of person who gets contacted needs vote history for the people who were not contacted too, matched on turnout score. That data is not in this report.
Turnout scores and vote history are a VAN export for the Putnam County universe, cross-tabulated at 5-point score thresholds. The export is cumulative — each row counts everyone at that score or higher — so the bands here are built by differencing adjacent rows. Contact records come from two places: the program's VAN contact history, which records that an attempt happened, and the dialer's and texter's own logs, which record what came of it.
7,836 records of type “Event” are vote credits rather than outreach: one per person, bulk-loaded daily from 6 October through 3 November, 3,965 of them on 3 November alone. They are excluded from every attempt and channel figure, because nobody was contacted by them, and used instead as the record of who voted. Excluding them is why the universe reached is 12,519 rather than 15,454; the 2,935 people who appear only as a credit were never contacted. The record is incomplete and final vote history is still to come.
A dial reached a person when the caller logged an outcome only a person can produce — a conversation, a refusal, a wrong number, not home, moved, deceased, or no language in common. The dialer also carries its own answering-machine detection, which flagged 1,299 human answers against the caller's 1,231. The two agree on 761 dials and disagree on 1,008, so the figures here are the campaign's own record rather than the machine's.
The contact figures on this page are counts only. Individual voter identifiers are not published here, nor carried in this file's data, and volunteers appear by rank rather than by name. There is nothing on this page that resolves to a person.
Placing each band at its midpoint assumes scores are spread evenly inside the band; pinning them to the floor or the ceiling instead moves the headline ratio only from 0.78 to 0.72, against 0.75 as reported, so that assumption is not driving the result. This is one election in one county: a score built to predict a presidential year will look miscalibrated against a midterm even when it is working as designed, and nothing here separates those two explanations. The contact data describes a universe that was itself being worked, so these are not an untouched control group.
The two contact sources do not reconcile exactly. 12% of dials in the general (2,370 of 19,884) carry a voter ID that appears nowhere in the VAN contact history, so the dial happened but no contact record was written for it. Those dials are counted here rather than discarded, which is why the dial total runs ahead of the phone attempts in the contact log.