I reanalyzed the raw data from Wisconsin’s primary polls. Here’s what actually went wrong.

Author: G. Elliott Morris | Published: 2026-08-13 | Source: article

Summary

The only source in the corpus that decomposes the Wisconsin polling miss with microdata rather than hypothesis. Morris re-weights State Navigate’s July and August polls (Hong +29, then +22 five days before Crowley won by 0.5) and attributes the ~22-point error to three separable causes, in descending order:

1. Wrong demographic targets (5–10 points of margin). The polls modeled an electorate that was far too young: July’s sample was 17% under-30 and 36% seniors; the L2 voter-file universes of actual Democratic primary participants are 2–3% under-30 and 52–53% seniors. Re-weighting to the voter file cuts the August error from 24 to about 14 points; adding ideology (pinned to 2016 exit-poll levels, 25% “very liberal” vs. 38% in the weighted sample) and income brings it to about 3.
2. Late movement (6–10 points). A 633-person recontact panel shows the Hong–Crowley margin fell 7 points in ten days before fieldwork ended; Barnes’s withdrawal and undecideds broke toward Crowley (40% Crowley, 20% Hong among late deciders); the moderate lane “finished [consolidating] on Election Day.”
3. Non-ignorable nonresponse (2–5 points, “0 to 5” depending on weights). Forcing the poll to match the result required upweighting Crowley voters within every demographic cell, so the Hong-friendly skew was not just too many young people. A recontact check found no differential response on the second wave (51.2% for both camps). Applying the Andridge/West/Little MUBP sensitivity test, Hong’s overstatement is explainable by ordinary observable selection under the best weights but requires a nonresponse parameter of about 0.7 (on a 0–1 scale where 0.5 is a typical guess) under demographic-only weights; Crowley’s curve never reaches his actual share under two of three electorate models.

The bottom line Morris draws is deliberately anti-alarmist: most of the miss was fixable weighting plus late movement, not “pervasive non-response by moderates,” and re-weighted the polls missed by about the primary average (~13 points). His recommendations: weight to voter-file primary participants, field into the final weekend, run ideology sensitivity analyses, and run a nonresponse diagnostic before publishing.

Relevance to the wiki’s open question. This source does not test issue polling directly, but its mechanism findings are the best evidence available: the largest component (age targets) is specific to low-turnout primary electorates and would not transfer to a general-population issue poll, which has census age benchmarks; the ideology component (a weighted sample 38% “very liberal” against no official benchmark) and the within-cell nonresponse residual (2–5 points) are the parts that plausibly do transfer, because issue polls also lack ideology benchmarks and also depend on who answers. That implies the progressive-overpolling discount on youth-socialism favorability numbers is real but small, on the order of a few points, not the 20-point primary miss.

Key Claims

• State Navigate: Hong +29 (July), +22 (Aug, ended Aug 6); result Crowley 39.8–39.3.
• Sample age: 17% under-30 / 36% seniors vs. voter-file primary electorate 2–3% / 52–53%.
• Weighting to voter file cuts error 24 → ~14; adding ideology and income → ~3.
• Recontact panel (n=633): margin fell 7 points in ten days; Crowley gained ~2/3 point per day.
• Undecided allocation using late-decider flows reproduces a 0.1-point race.
• Nonresponse (MUBP): 0–5 points of Hong margin; about a tenth of the total gap.
• Weighted sample 38% “very liberal”; 2016 exit benchmark 25%; 17% self-identified democratic socialist in State Navigate’s published toplines.
• Average Senate/governor primary poll error is about 13 points.
• Morris built the pollster’s weighting program (disclosed); analysis not previewed by the pollster.

Direct quotes worth preserving

“Most of the miss in polls in Wisconsin is attributable to faulty demographic targets (too many young people). This inflated Hong’s vote margin by somewhere between 5 and 10 points.” — G. Elliott Morris

Why citable: the quantified primary cause of the 2026 progressive polling miss, from the one analyst with the raw data.

“Non-ignorable non-response within demographic categories likely further inflated Hong’s vote margin by 2-5 points.” — G. Elliott Morris

Why citable: the size of the part of the miss that is actually nonresponse bias, as opposed to weighting error or late movement.

“The older and more moderate voters who did take the poll were still more Hong-friendly than the older and more moderate voters who actually showed up.” — G. Elliott Morris

Why citable: names the within-cell mechanism that demographic weighting cannot fix, which is the mechanism that would transfer to issue polling.

“It is reassuring that most of the 20-point miss by the polls came from two problems that are either (1) fixable (weighting model) or (2) unavoidable in primaries (late movement), rather than, e.g., pervasive non-response by moderates.” — G. Elliott Morris

Why citable: a direct, evidence-based narrowing of the WaPo nonresponse hypothesis.

Connections

Entities mentioned: Francesca Hong Events: 2026 Wisconsin Democratic Gubernatorial Primary, 2026 Michigan Democratic Senate Primary Debates: Does socialism win elections in America?, Why is socialism popular with young Americans?

Contradictions / Tensions

• Narrows Wisconsin surprise raises big questions about polls and Democratic voters: WaPo’s headline hypothesis (“one candidate’s voters wanted to answer polls”) accounts for ~2–5 points, not the bulk; weighting targets and late movement account for the rest. Nuttycombe (quoted in WaPo) is the source of State Navigate’s data here.
• Does not address the Median Strategies fabrication claim in Abdul el Sayed outperforms Mike Rogers in latest Michigan Senate poll (Susquehanna, Aug 11–17, 2026); State Navigate’s data were real and reproducible.
• Wisconsin margin: Morris uses 0.5 (39.8–39.3); the wiki’s Aug 13 reconciliation used 0.4 pending certification.

Notes

Lean is by author orientation (Morris is a former Economist/ABC data journalist generally read as left-of-center); the piece is methodological. Ingested for the overview’s open question on whether progressive nonresponse bias generalizes to issue polling; the answer it supports is “partially, and by a few points.”