In the late summer and fall of 1936, clerks in the Literary Digest's New York offices addressed more than a quarter million envelopes a day. The magazine had polled presidential elections every four years since 1916, when its first poll covered five states; the four nationwide polls that followed had all called the winner correctly. Ten million ballots went out this time, the names pulled from telephone books, automobile-registration rolls, club rosters, and mail-order lists. Each ballot asked one more question, almost as an afterthought: who had the respondent voted for in 1932?
On October 31, the Digest published its forecast: Governor Alfred Landon of Kansas would beat President Franklin Roosevelt, carrying 370 of the nation's 531 electoral votes. Nearly 2.4 million of the ten million ballots had come back, a 24 percent response rate higher than most polls manage today. The editors said afterward that they had reported whatever the mail brought them, unweighted and unadjusted, trusting the sheer size of the count over any statistician's correction. On November 3, Roosevelt carried every state but Maine and Vermont and took more than 60 percent of the popular vote. It was the worst miss in the poll's twenty-year run, and the one that ended it.
The standard explanation blames the mailing list: telephone and automobile owners skewed wealthier and more Republican in the middle of the Depression, when either one was still a luxury. That bias was real, but a 2017 statistical reexamination by Sharon Lohr and Michael Brick found a more specific failure sitting inside the Digest's own returns. Because the ballot had asked how each respondent voted in 1932, the magazine already held a way to check its sample against a known result. Weighting the 1936 figures by that single answer, using methods already published and available to any statistician of the period, would have shifted ten states, including California and New York, and correctly projected Roosevelt as the winner of the electoral college. The Digest never ran that calculation. Its polls had leaned Republican in every election back to 1920 and still called the winner each time, so its editors trusted that pattern to hold rather than the number suggesting it might not.
George Gallup's new American Institute of Public Opinion took the opposite approach: a sample a small fraction of the Digest's size, built to approximate the electorate rather than whoever happened to mail back a ballot. It correctly predicted Roosevelt's win. The contrast made Gallup's reputation and ended the Digest's. The magazine sold off its remaining assets within two years, its run as the country's most trusted straw poll over.
A sample's honesty comes from what it represents, not from how large it is. Ten million ballots pulled from the same kind of household can be a worse witness than a far smaller sample built to mirror the whole electorate. The correction for that kind of blindness does not always require new data. Sometimes it is sitting in the file already, in a question nobody ran the numbers on.
The Literary Digest had the right question on its own ballot. It just never ran the numbers.