Why Polls Miss: Likely Voters, Weighting and Small Samples Explained
Dated editorial analysis; figures in this article are not recalculated automatically. Visit the linked race pages for current data.
After every election, some polls look close to the final result and others do not. In 2026, the races we tracked had an average gap of about 4.6 points between the final polling average and the unofficial result, with a few races off by much more. This explainer describes the main reasons polls and results can differ. It is general methodology, not a criticism of any particular pollster.
1. Sampling error is larger than the headline suggests
A poll interviews a sample, not every voter. The reported margin of error, often around plus or minus three or four points for a sample of about 600 to 1,000 people, applies to each candidate's share. The margin between two candidates can move by roughly twice as much. So a poll showing a four-point lead is consistent, within its stated error, with a tie or with a lead of eight points or more. Margins of error also assume random sampling, which no real poll achieves perfectly, so actual error is usually larger than the published figure.
2. Deciding who is a likely voter
Many adults do not vote in a midterm election, so pollsters try to estimate which respondents will. Common approaches include asking people how certain they are to vote, whether they voted in past elections, and how closely they follow the campaign; some pollsters instead match respondents to official voter files that show their voting history. Results among all registered voters and among likely voters can differ by several points. If turnout among some groups turns out higher or lower than a pollster assumed, the poll can miss even if every interview was accurate.
3. Weighting the sample
People who answer polls are not a perfect cross-section of the public. Pollsters adjust, or weight, their samples so that groups such as age, sex, race, education and region match known population figures. Many also weight by past vote or party registration. Weighting corrects some biases but depends on the targets chosen. If the people who respond within a group differ in their views from those who do not, weighting cannot fully fix that. This is called nonresponse bias, and it is one of the hardest problems in modern polling.
4. Undecided voters and late shifts
A poll measures opinion during its field dates, which may end a week or more before Election Day. Voters who are undecided at that point, or who change their minds, are not captured. Pollsters handle undecided respondents differently: some report them separately, and some ask which way they lean. Those choices can change the margin a poll reports.
5. One poll is a weak guide
Each poll carries its own sampling error and its own method choices. An average of several polls from different firms tends to cancel out some of those differences. In our 2026 comparison, races with three or four verified polls usually finished within about two points of the average, while the largest misses came in races with just one or two polls, such as South Dakota Senate. See How accurate were the 2026 polls? for the full comparison.
6. When the polls disagree with each other
A wide spread between polls of the same race is itself useful information. In Texas Senate, the four final polls ranged from the Republican ahead by about one point to the Democrat ahead by seven. When polls disagree that much, the average is less certain than a single number suggests, and the result could fall outside the range of all of them, as it did there.
How to read polls next time
- Look at the average and the number of polls behind it, not a single release.
- Check the field dates, the sample size and whether the poll measured likely or registered voters.
- Treat a lead smaller than about twice the margin of error as uncertain.
- Prefer pollsters that publish their full methodology and question wording.
Our own approach, including which polls count toward an average and why, is described in the methodology and how we verify polls.
2026 figures are based on unofficial results as of November 2026 and the verified polls logged on this site.
Relevant race pages: South Dakota Senate · Texas Senate · South Dakota Governor · Texas Governor