Measuring public opinion
Polls are the main tool for measuring public opinion. Good polls require careful sampling and wording, and even good polls carry uncertainty.
Early polling and its failures
In 1936 the Literary Digest magazine, using millions of mailed ballots to telephone and car owners, predicted Alf Landon would defeat Franklin Roosevelt; Roosevelt won in a landslide. The poll's sample over-represented wealthier Americans. George Gallup, using a much smaller but more representative sample, predicted the winner correctly, establishing scientific polling.
Random sampling
A scientific poll uses a sample in which every member of the population has a known chance of selection. With random sampling, a sample of about 1,000 people can estimate national opinion within a margin of error of about plus or minus 3 percentage points, 95% of the time. Larger samples reduce the margin of error, but with diminishing returns.
Margin of error
If a poll shows Candidate A at 48% and B at 45% with a margin of error of 3 points, the race is statistically close: A's true support could be 45% to 51%, B's 42% to 48%. The margin applies to each candidate's share, so the gap between them is even more uncertain. Subgroups, such as young voters, have larger margins because their samples are smaller.
Modern challenges
Response rates for telephone polls have fallen from over a third in the 1990s to the low single digits, and many people no longer have landlines. Pollsters now use mixed methods, including online panels and text messages, and weight results to match the population on characteristics such as age, education, and region. Weighting on education became more important after polls underestimated some candidates in 2016.
Question wording and order
How questions are worded strongly affects answers. Support for 'assistance to the poor' typically exceeds support for 'welfare', though they describe similar programmes. Question order can prime respondents, and offering 'don't know' options changes results. Push polls, disguised as surveys but designed to spread negative information, are a campaign tactic, not real research.
Types of polls
Tracking polls follow opinion over time; exit polls survey voters as they leave polling places on election day, helping explain results; benchmark polls measure a candidate's standing at the start of a campaign. Poll aggregators average many polls, reducing the error of any single poll, and election forecast models combine polls with other information.
Polls and democracy
Polls let leaders know what citizens think and give citizens a voice between elections, but critics worry that leaders follow polls rather than lead, and that coverage focuses on who is ahead (horse-race journalism) rather than issues. A bandwagon effect may encourage support for leading candidates, while an underdog effect may help those behind.
Worked example
A national poll of 1,000 likely voters finds 52% support for a policy, with a margin of error of plus or minus 3.1 points. True support is likely between about 49% and 55%, so a majority is probable but not certain. If the poll's sample has too few voters without university degrees, and such voters oppose the policy more, weighting to correct the education mix might lower the estimate to 50%, showing why methods matter.
Key takeaways
The 1936 Literary Digest failure showed that biased samples mislead; Gallup's representative sample succeeded. Random samples of about 1,000 give a margin of error near 3 points at 95% confidence. Margins apply to each share, and subgroup margins are larger. Falling response rates push pollsters to online and mixed methods with weighting, especially by education after 2016. Wording and order shape answers; push polls are campaign tactics. Tracking, exit, and benchmark polls serve different purposes; aggregators average polls. Polls inform democracy but risk horse-race coverage and bandwagon effects.
Flashcard
Failed 1936 prediction from a biased sample
Check your understanding
A poll of 1,000 people shows 46% for X and 44% for Y, with a 3-point margin of error. What is the best interpretation?
