Public Opinion Measurement and Survey Methods

A poll is only as good as the people in it. For most of polling's history, the hard part was reaching a random slice of the public. Today the hard part is that most people you reach decline to take part, and the decision to answer is not random. People who are more engaged with politics are more likely to pick up, and in some elections the supporters of one side have been noticeably less willing to talk to pollsters than the supporters of the other. When the people who stay silent vote differently from the people who answer, no amount of sample size fixes the result.

Pollsters have not stood still. They now adjust their samples for how people say they voted last time or which party they identify with. They build models of who will actually turn out. They estimate opinion in hundreds of small areas from a single national sample. They field surveys by text message and online as well as by phone. And a new generation of tools promises to go further: AI interviewers that hold a conversation, language models that play the part of voters, and methods that read opinion from social media posts without asking anyone at all. Each of these is a response to the same problem, and each brings a new one.

Why this matters to you

If you read polls, the most useful thing you can learn is that the published number depends on choices made behind the scenes. Two pollsters can interview similar people and publish different leads because they weighted differently or counted likely voters differently. The same policy can look popular or unpopular depending on how the question was worded. If you commission polls, this chapter helps you ask the questions that reveal those choices before a headline does it for you. If you are tempted by cheaper, faster methods, from opt-in online panels to synthetic voters, it tells you what the evidence says they can and cannot replace.

What you'll find inside

The chapter starts with the problem itself: partisan nonresponse, the political pattern in who agrees to be polled. It then examines the main repairs. You will see what weighting on past vote or party identity did to the 2024 polls, why cheap opt-in panels can produce wildly different numbers for the same question, and how the mode of contact, whether phone, web, text or a mix, changes who is reached and what they say. You will watch the same set of respondents produce very different leads depending only on the rule for deciding who will vote, and see how small-area models can move seat projections by large amounts with no change in the electorate. The chapter covers why issue polling is so sensitive to wording, and what happens when you ask people who they think will win instead of how they will vote. It then turns to the new instruments: AI-led interviews, language models standing in for voters, opinion read from digital traces, and the surprisingly unsettled question of how to measure polarisation.

The honest note

The characteristic caveat here is timing. Many of the repairs pollsters made after 2020 were tuned on the electorate of that election, and electorates move. A fix that corrected one year's error can create the next one if the people who stopped answering change. The newest tools deserve the most caution: an AI interviewer or a synthetic voter may look impressive in a demonstration and still be untested on the question that matters, which is whether it gets real opinion right in a real election. This chapter reports those tests where they exist, and says so plainly where they do not.

The twelve pieces in this chapter

• P1-01 · Partisan Nonresponse: Who picks up the phone is political
• P1-02 · Weighting on Politics: Recalled vote, party ID and the price of stability
• P1-03 · Opt-In Panels in Politics: Cheap, fast and vulnerable
• P1-04 · Modes and the Missing Voter: Text, web, phone, mixed
• P1-05 · Modelling Who Will Vote: The electorate that didn't show up as planned
• P1-06 · MRP Comes of Age: Six hundred constituencies from one sample
• P1-07 · Issue Polling's Fragility: Support depends on the sentence
• P1-08 · Asking Who Will Win: Citizen forecasting
• P1-09 · Talking Politics to a Machine: AI-led interviews in public opinion
• P1-10 · Synthetic Voters: Can a model stand in for the electorate?
• P1-11 · Polling Without Asking: Opinion from digital traces, structured by a model
• P1-12 · Measuring Polarisation: The most-quoted construct with the least settled ruler