Is it a blue tidal wave,
or are Republicans not
answering polls… again?
Wick ran a new recontact study to explain what that phrase means and how much we might be seeing it again.
David Burrell · Wick · October 2026
Illustration, not data. For decades, polling error stayed in a fairly narrow range. Around 2014 it widened, and it has stayed wide. What changed? Not the math. The incentive to answer a poll.
Our recontact study shows one part of the problem: an incentive to answer polls that didn't exist until recently. It adds a hidden bias of 3.3 points toward Democrats on the margin and understates Trump's net favorability by 3.2 points. Even weighting to results we already had couldn't eliminate either one.
The study
We sent a new survey to 16,358 people
from our accurate late-October polls.
16,358 invited
from accurate late-October
polls in 2022 and 2024
Recontacted
October 2026
1,874 came back
from states with competitive races in 2026
510 came back
from Pennsylvania, with no competitive race in 2026
This is the study. We used our 2022 and 2024 text-to-web polls in Arizona, Florida, Georgia, Michigan, North Carolina, Pennsylvania, South Carolina and Tennessee: 16,358 respondents, mostly from mid-to-late October, from polls that came close to the final result.
In October 2026 we invited all of them to take a new survey. 2,384 came back, or 14.6%. That is well above the 1% or 2% typical for a new poll, because they had taken our surveys before.
2,384 came back in all: 1,874 from the states with competitive races in 2026, and 510 from Pennsylvania, our comparison state with no competitive statewide race.
Because we have everyone's original answers, including the people who didn't come back, we can compare the two groups directly. Normally a pollster knows nothing about the people who don't respond.
The study
We gave weighting
the best possible test.
A normal poll
Has to guess who will vote. Even after the election, there's no true count of ideology.
Our test
We had everyone's original answers, so we weighted the people who came back to match them exactly.
This is the best case for weighting. A normal pre-election poll is predicting a result that hasn't happened yet. It has to estimate turnout and the party mix, and there's no true benchmark for ideology, even after an election.
We didn't have to guess. We had the original answers of everyone we invited, so we weighted the people who came back to match them exactly on age, sex, race, education, urbanicity, party, ideology and 2020 vote recall.
Then we checked their old ballot answers, which weren't part of the weighting. If weighting could fix the problem, those should have matched too. They didn't.
The study
What the study tells us
01Real shifts
Changes in the same people, beyond shifts in party shares.
4–5 point shifts in House and Senate
02Who’s missing?
Evidence on whether we’re undercounting Republicans again.
3–4 point nonresponse bias
03Why people answer
Changing incentives in an always-on world of smartphones and social media.
Coincides with the mass adoption of smartphones and social media
This is the guide to the three sections that follow. First, compare the same respondents’ earlier and current answers to observe individual-level movement, rather than inferring movement from changes in category shares. The reported shifts remain findings among recontact respondents. Second, use the original answers of everyone invited to examine whether the people returning in 2026 leave a different mix within parties, including patterns consistent with undercounting Republicans. Third, discuss changing incentives to participate in polls as an always-on political narrative coincides with the mass adoption of smartphones and social media. This is a temporal association, not an attribution of blame to the media or proof of causation. The incentive explanation is an interpretation of the evidence and the conceptual illustrations, not a randomized causal finding. The self-reported survey-frequency comparison is an additional association in the 2026 recontact. Illustrations are conceptual, not measured transitions or quantities.
Real shifts
Recontact showed a major swing left.
The same voters, compared with their 2022 and 2024 answers.
Senate & House
4–5 points
more Democratic
A 2028 template
Ossoff showing
how to do it in GA.
Republicans
No positive shifts
to report
Governor’s races
More neutral
than federal races
How much of this is real, and how much are we just undercounting Republicans again?
The recontact shows a major leftward swing among people who took the accurate 2022 and 2024 surveys. The same 2024 respondents are 4 to 5 points more Democratic in 2026 Senate and House races than in their 2024 presidential vote. The 2022 cohort also moved toward Democrats; do not imply that the same 4-to-5-point figure applies identically to every comparison. Ossoff is holding nearly all of Warnock’s voters and adding former undecided and third-party voters, and some former Republican voters: a coalition-building template for Democrats. There are no positive Republican shifts to highlight in this recontact readout. Governor’s races show more neutral movement than federal races. These are findings among recontact respondents; the next section asks how much reflects real change and how much reflects nonresponse bias. The survey-frequency and hidden-bias discussion is intentionally separate from these four observations. Illustrations carry the concepts; all findings and numbers are native slide text. Source: research-hidden-nonresponse-2026-polls.html, the recontact’s other findings; four-point editorial framing requested by David.
The wrong mix within Democrats
| Democrats2022 / 2024 answer | Recontact response2026 response rate | 2026 recontact answered… Would vote R for HouseApprove of Donald Trump | |
|---|---|---|---|
 democracy |
Too high |
0.3% | 0.0% |
 |
Too low |
8.1%+7.8 pp | 9.4%+9.4 pp |
Democrats whose top issue was threats to democracy were 47.9% of everyone invited and 56.6% of the weighted recontact responses. Source: existing deck charts s2 and s4a, also reproduced in research-hidden-nonresponse-2026-polls.html. Groups are defined by their original answers. House ballot and Trump job approval are measured in the 2026 recontact. The characters illustrate subgroup averages.
The wrong mix within Republican men
| Republican men2022 / 2024 answer | Recontact response2026 response rate | 2026 recontact answered… Would vote R for HouseApprove of Donald Trump | |
|---|---|---|---|
 |
Too high |
74.8% | 69.9% |
 |
Too low |
90.5%+15.7 pp | 82.9%+13.0 pp |
Vaccinated men were 49.7% of everyone invited and 63.4% of the weighted recontact responses. Source: existing deck charts s2 and s4c, also reproduced in research-hidden-nonresponse-2026-polls.html. Groups are defined by their original answers. House ballot and Trump job approval are measured in the 2026 recontact. The characters illustrate subgroup averages.
The wrong mix within Independents
| Independents2022 / 2024 answer | Recontact response2026 response rate | 2026 recontact answered… Would vote R for HouseApprove of Donald Trump | |
|---|---|---|---|
 |
Too high |
32.4% | 28.1% |
 |
Too low |
54.2%+21.8 pp | 41.5%+13.4 pp |
Moderate Independents were 49.5% of everyone invited and 57.4% of the weighted recontact responses. Source: existing deck charts s2 and s4b, also reproduced in research-hidden-nonresponse-2026-polls.html. Groups are defined by their original answers. House ballot and Trump job approval are measured in the 2026 recontact. The characters illustrate subgroup averages.
Why polls get party shares wrong
1Not adjusting quotas
for party composition
+3.1points
What we do know
Registration points to Republican gains on the margin
in swing states since 2022.
The first two scenes are illustrative: four of five people cast ballots while the original Republican stays seated. These character counts are not turnout estimates. The Independent-to-Republican move and red cap are narrative symbols, not an estimate of individual switching.
Two separate moving inputs affect the party shares a poll needs to represent: changes in party composition and differences in turnout. The drawings are conceptual, not quantities or measured transitions. Registration is an indicator, not an interchangeable measure of party identification or vote choice.
The swing-state average of a 3.1 percentage-point Republican registration-margin increase versus 2022 was supplied by David. It has not been independently reproduced for this draft; the state set, observation dates, denominator, and averaging method remain to be documented before publication.
The turnout statement is a directional expectation supplied by David, not a known final 2026 electorate. Current reporting on North Carolina describes strong Democratic early participation, but early-vote shares do not establish final turnout composition nationally. Source: https://www.axios.com/local/charlotte/2026/10/19/north-carolina-early-voting-2026-democrats-gop . Additional context: https://apnews.com/article/44df2c18b387d28022dbb480e99feaba .
The two indicators can move in different directions: registration can become more Republican while differential turnout increases Democrats’ share of those who vote.
Why polls get party shares wrong
2Getting turnout
predictions wrong
What we do know
- Multiple signals point to a larger Democratic share of voters than in 2022.
The first two scenes are illustrative: four of five people cast ballots while the original Republican stays seated. These character counts are not turnout estimates. The Independent-to-Republican move and red cap are narrative symbols, not an estimate of individual switching.
Two separate moving inputs affect the party shares a poll needs to represent: changes in party composition and differences in turnout. The drawings are conceptual, not quantities or measured transitions. Registration is an indicator, not an interchangeable measure of party identification or vote choice.
The swing-state average of a 3.1 percentage-point Republican registration-margin increase versus 2022 was supplied by David. It has not been independently reproduced for this draft; the state set, observation dates, denominator, and averaging method remain to be documented before publication.
The turnout statement is a directional expectation supplied by David, not a known final 2026 electorate. Current reporting on North Carolina describes strong Democratic early participation, but early-vote shares do not establish final turnout composition nationally. Source: https://www.axios.com/local/charlotte/2026/10/19/north-carolina-early-voting-2026-democrats-gop . Additional context: https://apnews.com/article/44df2c18b387d28022dbb480e99feaba .
The two indicators can move in different directions: registration can become more Republican while differential turnout increases Democrats’ share of those who vote.
Why polls get party shares wrong
3Double counting
Unexpected
finding
finding
Label recategorization, not voter realignment
Some of the most reliable Democratic voters
now identify as Independents.
Editorial interpretation: some of the most reliable Democratic voters now identify as Independents. This is a change in party label rather than evidence of a change in vote alignment. The duplicated character is a conceptual illustration of recategorization, not proof of duplicate survey records. The duplicate represents the same respondent counted in two party categories. The original remains in the Democratic category while a second entry appears in the Independent category. Both entries remain in the sample. This illustrates double counting, not a change in the respondent’s political alignment. The scene does not quantify the frequency of the finding.
Why people answer
Answering a poll used to be
civic expression.
Then
Few polls. Editors decided what ran. Civic habit was enough to get people to pick up.
Then: answering a poll was civic expression. There were few polls, editors decided what ran, and civic habit was enough to get people to pick up.
The main drivers of who answered, like age, gender and education, were visible and benchmarkable. The census and voter files show the right mix, so weighting fixed it.
Why people answer
Now there’s a new incentive:
feeding the narrative.
Then
The incentive was civic expression. Few polls, and editors decided what ran.
Now
The new incentive is feeding the narrative. The feed never turns off, and every answer shapes the story.
Now: answering feeds the narrative. The feed never turns off, and every answer shapes the story.
Polls are now part of an always-on political story that matters a lot to some people. For some, that creates an incentive to answer and support their side. Others stop answering, especially when their side is losing.
Why people answer
Same party, same story,
different willingness to answer.
Answers
The poll story gives her
a reason to pick up.
Tunes out
The same story gives her
a reason to tune out.
Same party ID. Same narrative. Different willingness to answer. Both groups still vote.
You can't benchmark which type of person reacts which way, because these differences run inside each party and each demographic group.
We call this hidden nonresponse.
Why people answer
Other dynamics move
response rates, too.
The turnoff
When the polls look like a blowout, people on the losing side stop answering.
The tune-in
As Election Day gets close, people of every type start answering again.
Hidden nonresponse isn't the only thing moving response rates. Two other dynamics matter, and in late October they pull against each other.
We think there's a turnoff point. When the polls start to look like a blowout, people on the losing side are more likely to ignore the next text or call, like fans who turn off a football game when their team is down by 21 in the third quarter. Most of those fans still watch next week, and most people who ignore a poll still vote.
Then, as Election Day gets close, people tune back in. The election feels real, more people of every type have a reason to answer, and the gaps in who responds shrink.
Right now, in late October, both forces are at work at the same time. That's why pollsters can point to their final polls and say they weren't that bad.
But campaigns and outside groups make their biggest decisions, and spend hundreds of millions of dollars, in the summer and September. Judged by the polls from those months, the record looks worse.
Why it matters
Bad polls misrepresent voters. And no one
likes paying $1.2 billion for landslides.
Spending: candidates and outside groups, from OpenSecrets and news reports. A + means at least; Florida and Texas count candidate spending only.
Polls have a lot of practical uses, so if they're off for 23 of every 24 months in a cycle, that's a real problem. Elected officials need to know what people actually think to do their jobs well.
In the final stretch before the election, we're running large polls in many states built for nonresponse checks, and we'll check what AI finds against the results.
It doesn't mean every state has a 3.3-point error. Some could be close to that, others a little less and others a lot more.
The fix
How do we fix it? Find the
symptoms we can benchmark.
Too many
Polls reach too many people
from places with this profile.
Too few
And too few from places like this.
That mix we can measure and correct.
So how do we fix it? We have to find symptoms of hidden nonresponse that we can benchmark.
These two women are both Democrats, both moderate, both suburban. A standard poll treats them the same. But one answers polls and the other doesn't, and they live in different kinds of places.
Some symptoms can't be benchmarked. No census or voter file says how many voters rank threats to democracy as their top issue, so there's nothing to weight it to.
Where people live is different. Polls get too many people from places like the first one, with higher education, health care and economic measures, and too few from places like the second. There are hard statistics on every community, so that mix is something we can measure and correct, even inside the same suburbs.
That's the work ahead: find which community measures track who answers, then check whether matching on them brings the ballot error down. Illustrative profiles, not data.