Poll estimates compared: results, margins, averages, and forecasts

What to take away

• A point estimate is not a guarantee or exact population value.
• A sampling margin covers one defined source of uncertainty.
• A model-based interval depends on the model's assumptions.
• A polling average combines inputs under its own rules.
• A forecast estimates a future outcome, not current opinion alone.

Poll coverage often places unlike numbers in one sentence. A survey result measures responses during fieldwork. An interval expresses uncertainty under a method. An average combines polls. A forecast adds assumptions about what will happen later. The reading order for a single poll is covered in how to read a public opinion poll.

Labeling each number by function prevents a forecast probability from being mistaken for vote share or a poll lead from being presented as a certain outcome.

Comparison table

Output	Main question	Depends on	Safe wording
Point estimate	What did this poll estimate?	Sample, response, weighting, question	The poll estimated 47 percent support
Sampling margin	How much sampling variability is reported?	Probability design and confidence level	The full-sample margin was plus or minus 3 points
Credibility interval	What range follows from a model?	Data plus model and prior assumptions	The model reported a 95 percent credibility interval
Polling average	What do selected polls show together?	Inclusion, adjustment, timing, weights	The aggregator's average was 46 percent
Forecast probability	How likely is a future event under the model?	Polls, fundamentals, turnout, uncertainty	The model assigned a 60 percent chance of winning

Point estimate

A point estimate is the single reported value, such as 47 percent. Its decimal precision should match the measurement. A result of 47.3 percent does not mean support is known to one-tenth of a point. The statistical claim checklist makes the margin or interval part of the claim before any decimal is repeated.

Report the population and field dates. "Voters support" is too broad if the poll sampled adults or used a defined likely-voter screen.

Margin of sampling error

For a probability sample, the reported margin describes sampling variability at a confidence level under the survey design. It does not include every error from nonresponse, question wording, coverage, turnout assumptions, or processing.

AAPOR's polling accuracy page explains this limit and notes that polling averages can reduce attention to the noise of individual results while showing a broader trend. The AAPOR poll accuracy explanation supports distinguishing sampling error from other errors and an average from a single poll. Its practical rule of thumb is not a substitute for studying a poll's own methods.

Credibility interval

A credibility interval is produced through a model and its probability assumptions. It may look like a conventional plus-or-minus range, but the interpretation and basis differ.

Before reporting one, identify the data source, model, prior assumptions, stated probability level, weighting, and population. Ask what would happen if the assumptions changed. The interval can be calculated correctly under the model while failing to represent an uncovered group or an incorrect response mechanism. Keep its technical label in the story so readers do not mistake one uncertainty framework for another.

AAPOR has warned against treating a credibility interval from nonprobability polling as if it were a conventional margin of sampling error, because the interval depends on model assumptions that can fail. The AAPOR statement on credibility intervals supports naming the interval correctly and describing its model basis. It does not mean model-based surveys are unusable.

Polling average

An average needs editorial treatment as a separate product. Review:

• which polls qualify;
• whether results are adjusted;
• how old polls lose weight;
• how sample size and pollster effects are handled;
• whether multiple releases from one sample are counted;
• how undecided responses are treated.

Two averages can differ because their rules differ, even when they draw from many of the same polls. When an average anchors a story, the broader guide to statistics and polls reporting supplies the method questions to ask.

Forecast probability

A forecast asks about a future outcome. It may use polls, economic indicators, historical patterns, incumbency, geography, turnout scenarios, and simulations. A 60 percent win probability does not predict 60 percent of the vote. It means the modeled event occurred in 60 percent of the model's probability distribution or simulations under its stated setup.

Probabilities near 50 percent describe substantial uncertainty. "Favored" does not mean "certain." Wording that survives events is a habit borrowed from developing news stories.

Do not cross the labels

Avoid these substitutions:

• poll support for forecast probability;
• candidate lead for probability of victory;
• credibility interval for sampling margin;
• full-sample uncertainty for subgroup uncertainty;
• average value for agreement among all polls.

Use a small method note whenever the main text cannot hold every distinction.

Common questions

Can a poll be accurate if its result falls outside the stated margin?

The margin is probabilistic and limited to sampling error. Other errors and ordinary tail outcomes remain possible.

Is a polling average always more accurate?

No. It can reduce random noise, but shared coverage or response errors can remain. The inclusion and weighting method matters.

What does a 70 percent forecast mean?

Under the model's data and assumptions, the event receives a 70 percent probability. A 30 percent outcome remains possible.

Should a headline use the point estimate or interval?

The point estimate can appear, but the wording should not hide uncertainty or announce a difference the design cannot distinguish.