Total Recall Bias

1. Recall bias in polling has likely resulted in the industry overrepresenting Trump voters in surveys due to a disproportionate number no longer claiming they supported him in 2024. They’re largely concentrated in groups Trump made large gains with in 2024: lower propensity, younger, non-white, politically disaffected voters who are unhappy with his job performance.
2. Our estimate is that this probably results in a 2-point error in favor of Democrats in a standard generic ballot poll: a more conservative estimate based solely on data from Verasight points to a 1-point error in favor of Democrats.
3. Roughly half of polls in September and October have a viewable recall result, and on average the ones that do are likely shooting too conservatively for a midterm electorate.
4. A large non-response bias amongst Republicans is required to counteract weighting decisions made by pollsters and recall bias in order to cancel out the effect.
5. Our Autumnal polls have a recall result that is, on average, 10 points to the left of the actual 2024 presidential result. That’s mostly due to where we poll: state legislative districts with many non-college white voters that tend to drop off in a midterm election while Trump is in the White House, and most of which are concentrated in two states without a competitive statewide election to juice turnout.

Over half of public polls this year ask the same question we do: who did you vote for in the most recent presidential election? On average, the answer should closely match the actual result in a presidential year and be a couple of points to the left in a midterm year with an unpopular Republican President. Across polls released since September that publish the number, the average poll, excluding State Navigate's, shows a recalled presidential vote that is 2.2 points more left-leaning than its 2024 result. Ours is 10.3, an 8.1-point difference from the non-State Navigate mean in all polls since September.

Several factors explain this gap, but the biggest is that, in most of our polling, we don't weight on recalled presidential vote. The biggest weighting decision that affects a poll's toplines is how and where you weight your sample's partisanship. At State Navigate, we normally choose party registration in states that have it, and "leaned" party identification (which party respondents say they identify with, plus what party independents say they lean toward) for states without party registration. We take this approach because in a non-presidential election it usually relies on guesswork to decide where the electorate "should" be in its recall vote versus its actual presidential result. In other words, we have no prior state or district-level data to estimate what previous electorates look like based on their 2024 presidential vote, aside from the 2022 national electorate (which we'll get into later). We do, however, know what they look like by party registration and (roughly) party identification. We get party registration numbers from our voter file vendor, L2, and party identification from previous AP VoteCast data and Gallup quarterly party ID. All our team knows going into a poll is what our weights should roughly be: not what our recalled 2024 presidential vote "should be," nor what our topline "should be."

The most recent example of the perils of weighting to recalled presidential vote was last year in Virginia. In 2025, the Virginia gubernatorial election had an average recall vote (among pollsters that included one) in line with Virginia's 2024 presidential results; Noah Wyhof-Rudnick, who was collecting every pollster's partisan breakdown for that race, noticed it a month out. As students of Virginia history, we knew this was a doomed decision for any pollster, which led to State Navigate becoming the most accurate pollster in the 2025 election cycle. We chose the party identification weights for our Virginia gubernatorial polls not based on past vote, but instead on data from Catalist, Gallup, and AP VoteCast. As a result, our poll's presidential recall vote was 6 points to the left of 2024. Based on exit polling in the 2025 Virginia general election, the presidential recall margin was Harris +11.7, making our 2025 polls' recalled margin just 0.3 points off the exit poll. Our Virginia polling slightly underestimated Democrats by overshooting Winsome Earle-Sears’ performance among Trump voters in the electorate: our average result showed Sears earning 96.4% of the Trump vote when she actually earned closer to 93%. Remove the half-point of misreporting that the model detailed in this article expects for that electorate, and the Virginians who voted in 2025 actually voted about Harris +11.2 in 2024.

This year, our Virginia generic ballot poll has a presidential recall vote 8.2 points to the left of 2024 (and 2.8 points to the left of where the 2025 electorate had voted in 2024). That's slightly more friendly to Republicans than our average state legislative poll though, believe it or not. Part of the reason is where we poll: these districts have more low-propensity, non-college Republicans who typically drop off in a midterm election while Trump is in the White House, and they don't have a competitive statewide election.

The other part is what we call Total Recall Bias.

Recall bias usually refers to misremembering in the social sciences. While misremembering one's presidential vote may be a small part of what we call Total Recall Bias, we believe voters' attitudes and vote choice are the biggest drivers, and the academic literature supports that. The standard finding, from Dutch, Canadian, French and Norwegian panels that compared recalled votes with earlier reports, is that recall drifts toward the respondent's current preference (what Durand and her co-authors call reconciliation), that the drift is concentrated among the least engaged and least stable voters, and that social desirability plays a smaller part (van Elsas, Lubbe, van der Meer and van der Brug 2014; Durand, Deslauriers and Valois 2015). This pattern holds across every panel we examine in this article: voters that change their answer about their 2024 vote are loosely attached, their new recall choices are usually did-not-vote or a third party rather than the other candidate, and the rate barely moves with approval. Someone who is unhappy with Trump now says they never backed him, much like, in the Norwegian panels, voters who switched parties misremembered their earlier vote at several times the rate of consistent partisans, and mostly in favor of the party they moved to (Waldahl and Aardal 2000).

We believe voters increasingly use their answers to a presidential recall question to express opinions on the current presidential administration and reaffirm their upcoming vote choice, which we believe drove many of the 2024 polling misses and will partly contribute to a 2026 polling miss. In simple terms: 2024 polling overrepresented voters who voted for President Biden in 2020, and we believe this year's polling overrepresents voters who voted for President Trump in 2024. Our final estimate of how this will affect toplines in generic ballot polling depends on how the presidential recalled vote is weighted. It’s about 1.0 to 1.1 points when the targets include a did-not-vote share, and closer to 1.7 to 2.2 points when the poll weights only the Trump, Harris, and other/third-party shares among respondents who report a vote and leaves the did-not-vote answers alone. Most methodology statements do not say which. That said, not every survey will overrepresent Trump voters by that margin; the bias rate varies by a poll’s sampling, demographics, location, and, of course, the electorate's partisanship. We derive this estimate from our recontact surveys and from The Argument and Strength In Numbers (both partnered with Verasight, which uses an opt-in panel), which include data on who voters said they voted for immediately after the 2024 presidential election or on panel re-entry versus who they say they voted for now.

Since mid-September, State Navigate has fielded 28 polls, all text-to-web and sampled from a voter file from i360 or L2, with weights derived from our standard processes (except for a Vermont Governor poll and an alternative Iowa State Senate poll). Our field experiment includes 14 surveys with recontact data, with a total of 3,741 respondents who answered both surveys.

Here we have a group of people whose 2024 recall we recorded once, and for most of them Trump's approval has since fallen; the exceptions are Iowa and West Virginia, where Trump's approval had improved slightly since the first survey. We record their responses to see how their attitudes have changed and if they still give us the same answer in their 2024 presidential recall vote. Respondents were asked their voter likelihood, their approval of President Trump, their horserace(s), their party identification, and various demographic questions. These were all quick, easy polls to make it as painless as possible for someone to take time out of their day to respond, which is important for capturing attitudes among the least likely voters to fill out a survey.

Our hypotheses come from Lakshya Jain's work with Verasight, which recorded every panelist's 2024 vote the day after the election and asked again in February 2026. Trump voters who still approved of him recalled their vote correctly 98% of the time; Trump voters who had turned against him claimed a Harris vote 13% of the time and claimed not to have voted 12% of the time. Our surveys should then show three things. First, the people who now disapprove of Trump should be the ones whose recall moves, and the people who still approve should not move at all. Second, as the share of Trump voters who disapprove grows, the recalled vote should become "bluer." Third, Harris voters should barely move, unless, of course, they approved of Trump at wave one but now disapprove. The anchors we test against that pattern differ in two things: the level of misreporting, and how fast it moves as approval falls. The Jain model (named after The Argument/Verasight survey and Lakshya Jain) applies his February 2026 Verasight rates. The Morris model, named after G. Elliott Morris, uses the final Strength In Numbers/Verasight survey in October (fielded October 13 to 19). In that survey, 6.4% of recorded Trump voters gave a different answer the day after the election, compared with 6.0% in Jain's February wave.

The chart below is where the hypotheses start. Across 135 polls from nine pollsters that publish Trump approval by 2024 vote, the share of his own voters who disapprove of him rose from about 15% in February to about 21% in October: 0.63 points for every point of national net approval he lost this year. If misreporting tracked disapproval the way the February rates say it should, the misreporting rate would have climbed right along with it, and the recalled vote in our surveys should have drifted a few tenths of a point bluer between waves. That is the slope our model gets tested on. The second graph is the other side of the coin: Harris voters who approve of Trump, about 4% in February and 2% by the fall, which is why the Harris-side term in every model is small and getting smaller.

Nine pollsters, 135 polls from January 2025 to October 22, 2026. Pollsters sit at different levels, so the fit uses the change in Trump-voter disapproval as national approval falls, with a house effect for each pollster. Economist/YouGov sampled adult citizens through August 2026 and registered voters from the September 4 to 8 release, marked on the chart; one house effect covers both. The 2026 fit is the one the published model uses.

Our sample is not demographically representative of the nation, though, and it runs much “redder” as well. Trump's 2024 share runs from 46% in Virginia to 69% in WV SD-1, against 50% nationally, and our thirteen district samples are whiter, older, and more rural. With more Trump voters available, they should also have larger misreporting rates: you’d expect the effect to rise with Trump’s share. Unless, that is, the extra Trump voters are older and whiter and rarely change their presidential recall vote. In our surveys, they are. The voter file is our control. For every respondent, the state's own record tells us whether they cast a 2024 ballot, so we count people claiming a vote they never cast, not estimate them.

The first question behind our interest in this experiment is what drove our polls’ recall bluer than the actual 2024 result: what effect did respondents who changed their recalled vote in our Pennsylvania follow-up polls have? For each Pennsylvania district polled twice, the change in recalled margin between the first poll and its October follow-up reflects two effects: respondents who answered both times and gave a different 2024 answer the second time, and a different mix of people answering. Of the nine districts with an October follow-up poll, changes in recall vote answer accounted for, on average, 32% of the overall change in recall composition, ranging from -0.1 to 1.8. The remainder of the change in recall is who answered the second poll.

Of the change in a district's recall between its first poll and October, the orange share is respondents who changed their own answer; the grey is a different mix of people answering. Likely-voter weighted.

The second question is the slope. Between Jain's February wave and Morris's October wave, Trump's net approval fell about nine points, from -18 to -27, and the share of recorded Trump voters who no longer say Trump moved from 6.0% to 6.4%. Four tenths of a point for nine points of approval. Our own Pennsylvania surveys had suggested something steeper, 0.135 points of recalled margin for every point of national net approval lost, which, run from the inauguration, would put misreporting at 4 to 6 points by October. Neither the national panel nor our non-Pennsylvanian contacts bore it out: the recalled margin moved slightly redder between those waves, not bluer. In other words: misreporting is a level, not a slope.

The model we publish, which we call Nuttycombe III, starts from Jain’s February rates applied two-sided to each district’s actual 2024 Trump and Harris voters, so a Trump voter who claims Harris moves the margin two points and one who claims not to have voted moves it one, mirrored on the Harris side. It then moves those rates along the February-to-October Verasight slope, run on each district’s own modeled net approval rather than the national figure, because Trump’s approval hasn’t fallen evenly nationwide. On our district approval model it fell 6 points in PA SD-6 since February and 7 in SD-36, against 9 nationally, and it actually rose a point in both West Virginia districts between their first poll and the recontact. By October’s approval, the district term moves the answer by hundredths of a point; it matters as approval diverges by place. Nuttycombe III adds who the Trump and Harris voters in a given sample are. Jain’s pooled 2026 cells say the rate at which a disapproving Trump voter disowns his vote falls steeply with age, from about 30% under 30 to 6% at 65 and over. It is higher among voters without a four-year college degree, about 21% against 13%, and about four times higher among non-white Trump voters: across all recorded Trump voters, approving or not, 14.8% of non-white Trump voters gave a different answer against 3.7% of white ones, and 8.3% against 1.2% claimed Harris. Our model scales each approval group’s rate by age, degree and race, each factor set to 1 at Verasight’s own mix and averaged over the recallers in the poll’s own sample. A sample that looks like Verasight’s keeps the national rate. Ours, which are older and whiter, get a lower one. The full definition is in the methodology.

The result is 1.0 to 3.7 points of expected misreporting in the Pennsylvania districts, 2.0 on average, 2.1 in Iowa SD-35, 2.2 and 1.8 in the two West Virginia districts, and 1.5 in Virginia. The two highest are Pennsylvania HD-41 and HD-187, where many of the disapproving Trump recallers are non-white; the two lowest, Pennsylvania HD-131 and HD-18, are where they are oldest. That is about 16% on average of what makes our recalls more Harris-friendly than 2024 across the 14 polls, from 7% in PA HD-131 to 28% in HD-41 and HD-187. The rest is who answered and the electorate we weighted to, which we get to below.

The cross-state check says the same thing. Across the 14 polls, the total gap rises 0.19 points for every point of Trump's 2024 share, but expected misreporting rises only 0.02 per point, because it is the young and non-white Trump voters who disown him, and the districts with the most Trump voters are the ones whose Trump voters are oldest and whitest. What remains after misreporting rises 0.17. The rise with Trump's share sits almost entirely in who answered and the electorate we weighted to.

The surveys are the direct test of the level. Of the people who told us Trump the first time, the share who no longer say Trump runs from 1.5% in WV SD-1 to 6.1% in PA HD-187, with Virginia at 3.9% of 388 over a full year. That sits between Morris's August benchmark of 3% and Morris's October rate of 6.4%. On the recalled margin, every level model predicts no change between waves and the slope models predict a few tenths. What we observed, +0.0 to +3.8 in Pennsylvania between the two waves and -1.9 to -0.9 in the October recontacts, scatters around zero by several points in both directions. With 110 to 917 people per survey, moves of that size are within sampling error, and they point both ways.

Those rates are averages. The follow-up surveys let us ask who, specifically, gives us a different answer about 2024 the second time we ask. Pooling all five, 2,966 people answered the recall question at both waves and 129 changed it, 4.1% after weighting.

The changers are not a cross-section of the sample. Respondents without a college degree changed their answer at 5.0%, against 3.2% for degree holders. The under-30s changed at 5.7% and the over-65s at 3.6%. Men changed at 5.0% and women at 3.2%. Households under $50,000 changed at 6.3%, against 3.1% at $150,000 or more, and people whose household income fell a bracket between the waves changed at 8.0%.

But the biggest factors influencing who does and does not change their vote recall are political preferences and identifications. Independents at wave one changed at 16.8%; partisans of either side at 3 to 4%. People who switched their party identification between the waves changed at 16.5%, and people who moved to or from independent at 15.1%, against 3.2% for everyone whose party ID held. The same goes for the ballot: people who switched their vote between the parties in the race we were polling changed their recall at 13.0%, those who moved to or from undecided at 9.0%, and those with the same vote choice both waves at 3.3%. Soft approval behaves the same way. Somewhat-disapprovers of Trump at wave one changed at 10.9%, against 2.0% among his strong approvers and 4.2% among his strong disapprovers, and people whose approval moved to or from "don't know" between the waves changed at 28.8%. Simply put: the people who change their answer about 2024 are the people whose answers about everything are loosely held, which holds up against the panel studies cited above: van Elsas and her co-authors found the drift concentrated among the least engaged and least stable voters, and Waldahl and Aardal found party switchers misremembering their earlier vote at several times the rate of consistent partisans.

Share of re-interviewed respondents whose 2024 presidential vote differed from their first answer, in our five pooled surveys, the Strength In Numbers / Verasight October 2026 wave and The Argument's pooled Verasight file for January to October 2026. All three are on the same base, respondents whose first answer was Trump or Harris, so our pooled rate reads 2.7% here rather than the 4.1% of all re-interviewed respondents. A row carries a dot only for the surveys that have that cut; the Strength In Numbers between-waves rows are Morris's linked subset, unweighted and small.

So where do the recall changers go? Of the 82 people who left a major-party answer, 61 went to a third party or to did-not-vote and 21 crossed to the other candidate: 13 from Harris to Trump and 8 from Trump to Harris. People don't so much change sides as stop claiming a side. A third of the changers were independents at wave one, a quarter changed their party identification between the waves, a fifth changed the party of their vote choice, and a third changed their Trump approval, mostly negatively. On a recalled margin, an exit counts for about half of a switch, because it removes a Trump vote without adding a Harris one. It is part of why the misreporting piece lands at 1 to 4 points rather than higher, and why the soft disapprovers move the margin less than their change rate suggests.

The national Verasight panels that Jain and Morris were so gracious to provide us with measure the same thing a different way. Verasight recorded every panelist's 2024 vote when they joined the panel, so a change there is a self-report checked against a record, where ours is two self-reports weeks to a year apart. That is why their rates run higher, but the ordering is the same.

Lakshya Jain's pooled January-to-October file, 11,530 registered voters with each panelist counted once at their latest interview, is the base we lean on, and it adds the direction of the change. In it, 8.8% of registered voters give a different answer than their profile, against our 4.1%, which is no surprise since our surveys run far whiter in demographics than the country’s. Respondents without a degree change at 10.9% against 4.7% with one; the under-30s at 14.1% and the 30-to-44s at 12.6% against 3.2% for the over-65s; non-white respondents at 14.2% against 6.4% for white respondents, with Black respondents at 16.5%; households under $50,000 at 12.2% against 5.6% at $200,000 or more. Gender barely impacts the rate in Verasight, where it does in ours.

Collapsed approval has a slight difference, 9.2% among Trump disapprovers against 8.1% among approvers, but that’s not noise: the disapprovers' recalled margin moves 4.2 points toward Harris and the approvers' 2.0 toward Trump, a net 1.9 points toward Harris across the whole panel. The groups whose recall moves most toward Harris in the Verasight panels are the under-45s, non-white voters and those in the lowest income brackets; their opposites (above 45, white, higher income brackets) are unsurprisingly the least likely. This explains a large chunk of the gap between Verasight's 8.8% and our 4.1%: the geographies we polled are older and whiter than Verasight's registered voters.

The monthly series settles the slope question more firmly than our two waves ever could. The share of recorded Trump voters who no longer say Trump sat between 5.3% and 6.8% in every month from January to August, with no trend as approval fell; Harris voters ran between 2.1% and 4.2%. September looked like a potential outlier, 3.2% of Trump voters on a sample half the size of the other months, but it did not last. Morris's October wave put the rate back at 6.4%, and Lakshya's pooled January-to-October file, which counts each panelist once at their latest interview and so leans on the October interviews, sits at 6.1% (2.7% claiming Harris, 1.0% another candidate and 2.5% not having voted), with Harris voters at 3.2%, 1.2% of them claiming Trump, within a point of Morris on both sides; no surprise in these small differences since both use Verasight. The level holds month after month, which is why the published model takes its level from the Verasight panels and runs only a miniscule slope on top.

Inspired by a recent article of Noah Wyhof-Rudnick, which we strongly recommend reading ahead of Election Day, we decided to conduct our own check against recall in polling this year. We’ve analyzed 667 polls with a field date ending September 1st or later to check how many have a published recall vote. Fewer than half, 296, publish one: 43% of statewide polls, 43% of House district polls and 52% of generic ballot polls. The remaining 371 have no usable recall result, with 266 of them not mentioning a presidential recall result anywhere in the toplines, crosstabs, methodology, or release. Here’s where things get really frustrating, though. There’s about 105 “bad apples”: 74 say they “weight to the 2024 vote,” but they never show a number, and another 31 have the results either paywalled or unavailable.

667 polls from other pollsters ending September 1 or later, through October 30. Left: the share that publish a 2024 recall we can compare with the result, by type. Right: what the rest publish instead.

The bottom bar of the graphic shows the partisan weighting schemes by pollster. 34% say they weight their sample by recall, including the ones that say they “weight to the 2024 presidential result” but we couldn’t find actual numbers for. 36% weight on other partisan variables, and the remaining 30% we either couldn’t find the partisan weighting decisions in their methodology or they didn’t describe their weighting at all. Again, the biggest mover of a topline is the partisan weighting decision, not demographics.

Here’s the big question at hand: what to do with the bad apples that say they they weight to the 2024 presidential result but do not disclose the actual percentages? If we assume that they quite literally mean they are weighting to the 2024 presidential result, we are very likely looking at a 2025 repeat where polling averages drastically underestimate Democrats, and that’s not even including the potential effect of our measurements on recall bias this year.