The polls look close. The early vote looks busy. What should we believe?

A two-point lead can dominate the news for a day. In Texas, Kansas, Michigan and Maine, it can also sit well inside the uncertainty of the poll that produced it. These Senate races deserve close attention. They deserve a more careful reading than either “the polls are broken” or “our candidate is ahead, so the poll must be right.” Sources: Texas October 8–11 survey, released October 14, Kansas September 15–16 survey, Michigan October 15–18 survey, released October 21, Maine October 22–25 survey, released October 27

Our assessment as of October 29 is that the evidence for competitive races is stronger than the evidence for any precise finishing order. Pollsters are measuring preferences through difficult recruitment conditions and different turnout models. Early ballots give us another view of participation, but they do not reveal candidate choices. The sensible approach is to test whether those two kinds of evidence tell a consistent story, without forcing one to validate the other. Sources: What low response rates mean for telephone surveys, Official one-week-out voting totals and mail comparison

Polling has had bad years. It has also recovered.

AAPOR’s retrospective separates the direction of error from its size. Across offices, late polls had almost no net partisan tilt in 2018 and a small Republican tilt in 2022. They leaned much too Democratic in 2020. The Democratic tilt was smaller in 2024, but it persisted. A near-zero average bias can conceal individual misses that cancel, so it is not a claim that every 2018 or 2022 race was measured correctly. Sources: 2024 pre-election polling evaluation, executive summary and Table 2.2.1; released October 2025

The historical lesson is limited but useful: persistent problems can coexist with improvement. Before 2026 votes are counted, no honest analysis can certify this year as more accurate than 2024, 2022, 2020 or 2018. Nor can an election’s partisan error simply be carried forward as a guaranteed adjustment. Sources: Key things to know about U.S. election polling

A low response rate is a warning, not a verdict

One response per 1,000 eligible people would be 0.1%. A 1.5% rate is about one per 67. Calls placed, unique people sampled, completed interviews and participation among already-recruited panel members are different denominators. A frightening statistic needs its definition before it can tell us much about the final sample. Sources: American Trends Panel Wave 192 methodology

Pew reported typical telephone response rates of 6% in 2018, a year in which election polls performed well historically. In a spring 2026 American Trends Panel survey, 87% of invited panel members participated, but the cumulative rate was 3% after recruitment and attrition. Both figures describe the same survey. Neither is an industry-wide 2026 election-poll response rate. Sources: Telephone survey response rates, including 2018, American Trends Panel Wave 192 methodology

The harder problem is political differences that remain after weighting. A sample can match age, education and geography while still missing less trusting or less engaged voters within those groups. More interviews from the same recruitment process may reduce random noise without repairing that gap. The familiar sampling margin also leaves room for errors from nonresponse, question wording and an electorate that turns out differently from the model. Sources: What low response rates mean for telephone surveys, Key things to know about U.S. election polling

Four Senate races, four reasons to read the details

Texas: competition is credible; a one-point order is fragile

Marist and Emerson both found a close Talarico–Paxton contest in their October surveys. That agreement gives the competition more substance than a single surprising release would. It does not establish who will finish ahead. Marist’s 47–45 result and Emerson’s initial 48–46 result also come from different dates and methods; they are not a clean movement series. Sources: Texas October 8–11 survey, released October 14, Texas October 17–19 survey, initial and leaned ballots

Emerson supplies a particularly useful caution. Its initial ballot put Talarico ahead by two points, with 5% undecided. After asking undecided respondents for a preference, the headline became 49–48, with 3% for the Libertarian. That narrower lead is a change in question treatment within the same survey, not a late swing in the electorate. We would be especially wary of a chart that mixed those versions. Sources: Texas October 17–19 survey, initial and leaned ballots

The early-vote signal is real but less partisan than some commentary suggests. Texas’s official October 26 release reported more than 4.2 million ballots through October 24, about 22% of 19 million registered voters. Counties submit prior-day figures and can appear in the statewide system with delays. This is a dated participation benchmark, not a current complete count, a party vote total or a victory indicator. Watch whether county reporting is complete before comparing metropolitan and rural participation. Sources: More than 4.2 million Texans cast ballots in first six days

Kansas: cross-check the ballot, and respect the age of the evidence

Emerson and YouGov both found Hamilton competitive with Marshall. The underlying ballots differ: Emerson named Libertarian David Graham and retained 8% undecided; YouGov asked undecided voters for a preference and reported a generic someone-else category. Their figures support a competitive reading, but subtracting one from the other would not establish momentum. Both surveys also predate the final weeks of the campaign. Sources: Kansas September 15–16 survey, Kansas September 24–October 4 likely-voter survey; native question 6

The state’s general-election dashboard displayed 328,522 advance ballots voted when checked on October 29. Ballots sent to voters are a separate administrative measure; they are not additional votes. The row carries an October 29 date without a precise completed-day cutoff, so it cannot support an exact same-stage growth percentage against 2018, 2020, 2022 or 2024 from the evidence reviewed here. Sources: 2026 General Advance Voting Counts; observed October 29 dashboard

Kansas is where an unexpected close contest most needs corroboration from fresh, independent surveys using the complete ballot. A large crowd, a suburban county’s busy polling sites or an advantage among one early-voter group cannot establish the statewide result. We would look for repeated agreement across methods and for turnout comparisons that include the same counties, modes and election-relative dates. Sources: Kansas September 15–16 survey, Kansas September 24–October 4 likely-voter survey; native question 6, 2026 General Advance Voting Counts; observed October 29 dashboard

Michigan: the likely-voter model is part of the story

Marist found El-Sayed ahead 50–48 among likely voters but ahead by seven points among registered voters. That five-point gap makes turnout assumptions consequential. It does not prove the likely-voter estimate is wrong; registered voters who do not vote cannot determine the result. It does mean that enthusiasm claims should be checked against who is actually returning ballots and against other surveys’ turnout methods. Sources: Michigan October 15–18 survey, released October 21

Michigan offers a proper comparison rather than just an impressive total. At seven days before Election Day, mail-ballot requests were 13.5% above the comparable 2022 point, while returns were 4.5% higher. The state also reported 215,795 early in-person ballots and 1,392,170 ballots overall as of 7 a.m. October 27. This is the first midterm statewide general election under universal early in-person voting, so adding the new channel and comparing it directly with 2022 mail returns would exaggerate growth. Sources: Official one-week-out voting totals and mail comparison

Maine: a large sample can still leave a wide interval

Marist’s latest reviewed Maine survey put Jackson ahead 49–47 among likely voters, compared with a nine-point lead among registered voters. Its published likely-voter margin was ±6.6 points, despite 1,449 interviews in the broader registered-voter survey. The larger headline sample is not a license to borrow the registered-voter precision for the likely-voter result. A two-point lead is plainly fragile here. Sources: Maine October 22–25 survey, released October 27

For Maine we would scrutinize the turnout model, compare other independent polls and keep the Senate and governor ballots separate. The same Marist survey found the governor’s race tied 44–44, with 6% for independent Rick Bennett and 6% undecided. Voters can split their tickets. An absentee registration advantage would not establish either race’s candidate margin. Sources: Maine October 22–25 survey, released October 27

The official Maine source publishes absentee requests, issuance and returns. We could not verify a comparable four-cycle series of returns at the same number of days before Election Day. That comparison stays unavailable rather than being filled with a percentage from mismatched snapshots. Requests must be separated from returned ballots, and voting an absentee ballot at a clerk’s office must not be mistaken for a newly independent voting channel. Sources: Official absentee voter data and latest file links

Early voting changed dramatically between the last four elections

The matched midterms show why more early activity need not mean a proportional increase in final turnout. From 2018 to 2022, early in-person voting rose 1.4 percentage points as a share of reported voters, while mail voting rose 8.7 points. Some growth in an early counter can reflect people moving their vote to an earlier date. That is participation timing, not automatically additional voters. Sources: CPS Table 14: Method of Voting, November 2018, CPS Table 14: Method of Voting, November 2022

The presidential comparison is different again. Mail voting accounted for 43.0% of reported voters in pandemic 2020 and 29.0% in 2024. Early in-person voting moved the other way, from 26.0% to 30.7%. A 2026 mail-return total below 2020 could therefore reflect a different choice of voting method rather than poor enthusiasm. Presidential totals also describe larger and differently composed electorates than a midterm does. Sources: CPS Table 14: Method of Voting, November 2020, CPS Table 14: Method of Voting, November 2024

For a fair early-vote test we need the same state, the same ballot status, comparable coverage and the same distance from Election Day. Even that is not enough if a new law changes when or how people can vote. County population growth, permanent-mail enrollment, newly available early sites and reporting lags also matter. The Michigan paired counts are useful precisely because they keep requests, returns and the new in-person channel apart. Sources: Official one-week-out voting totals and mail comparison, More than 4.2 million Texans cast ballots in first six days, Secretary Schwab announces advance voting

Where our confidence is stronger—and where it stops

The best-supported conclusion in these examples is competitive politics, not precise probabilities. Texas has agreement across two different October survey organizations. Kansas has corroboration across two organizations, with older and differently worded evidence. Michigan and Maine provide transparent registered-versus-likely-voter contrasts that show how much assumptions matter. None has yet passed the final test against 2026 election results. Sources: Texas October 8–11 survey, released October 14, Texas October 17–19 survey, initial and leaned ballots, Kansas September 15–16 survey, Kansas September 24–October 4 likely-voter survey; native question 6, Michigan October 15–18 survey, released October 21, Maine October 22–25 survey, released October 27

We would be leery of a claimed surge built from unlike ballot questions, a state projection resting on one stale survey, a demographic cross-tab treated as a full-size sample, or a candidate lead inferred from party registration. We would put more weight on repeated independent agreement, visible questionnaires and field dates, disclosed turnout rules, and administrative comparisons that survive a definition check. Sources: Key things to know about U.S. election polling, What low response rates mean for telephone surveys

For 21CEE, that also means owning the distinction between a computed estimate and measured precision. A displayed margin with two decimals does not make a close race knowable to two decimals. Polls can tell us that an apparently safe seat deserves attention. Early voting can tell us that people are participating. We should let the ballots determine who won. Sources: Key things to know about U.S. election polling