Midterm Results Live

How to Read Polls: Margin of Error and Likely Voters

Dated editorial analysis; figures in this article are not recalculated automatically. Visit the linked race pages for current data.

Every week brings dozens of new polls, and headlines often treat each one as news. A few basic ideas make it much easier to tell a real shift from normal noise.

1. Check what the uncertainty measure describes

A reported margin of sampling error applies to a specified estimate and survey design. It is not automatically the uncertainty of the difference between two candidates, and it does not describe every possible source of survey error. Nonprobability surveys may use model-based measures with different assumptions. This site does not calculate a combined margin of error for its averages or use a fixed four-point threshold to declare a statistical tie. Check the original poll's methods and uncertainty statement. AAPOR's guide to polls and surveys explains these distinctions.

2. Sampling error is not the only error

Sampling error is one part of survey uncertainty. Nonresponse, question wording, coverage and turnout assumptions can also affect a result. A small sampling error alone does not establish that a poll is accurate. AAPOR's survey research guidance describes the information needed to assess survey quality.

3. Who was asked matters: adults, registered or likely voters

Polls of all adults include people who cannot or will not vote. Registered-voter (RV) polls are narrower, and likely-voter (LV) polls try to estimate who will turn out. The same poll can show different results by screen. The current dataset does not consistently store population types or paired RV/LV releases. We display the population when it is recorded; an automatic preference for LV results has not been implemented. A rule for choosing between multiple population releases requires an editorial decision.

4. Check who paid for it

Campaigns and party committees release "internal" polls, usually when the results help them. These are not necessarily wrong, but they are released selectively. Some independent pollsters also lean consistently toward one party compared with the average, an effect called "house effect." We label each poll with the pollster name so you can judge for yourself.

5. Read averages with their limitations

Individual polls bounce around. Averaging several recent polls can reduce random variation, but it does not reliably remove systematic bias or pollster effects. Our current race averages use valid logged polls ending in the 14 days through the dataset date, with equal weight for each poll. Older polls are shown as historical data and are not carried forward. See the methodology page for coverage and source limitations. When a race has only one or two recent polls, the average is fragile, and we show the number of polls next to it.

6. Polls are snapshots, not forecasts

A poll tells you about opinion during the days it was in the field. Late-deciding voters, turnout and news events can still move a race. Treat a 2-point lead in October as "close," not as a prediction.

Quick checklist

You can apply this checklist to any race on our latest polls page.

More analysis

All analysis · Latest polls · Senate races · Governor races