Why the sample frame decides the result

Two and a half million responses and a confident forecast of the wrong winner. The 1936 poll was not too small. It drew from lists of car owners, telephone subscribers and magazine readers, in a year when owning those things predicted the vote.

What everyone sees

Accuracy is treated as a function of sample size, which is the number reported and the number readers use to judge credibility. Size governs only the random component. The frame — who could have been selected at all — sets a ceiling on accuracy that no additional respondents can lift, because the missing group is missing from every draw.

What is actually happening

Squire re-examined the 1936 Literary Digest failure and found the problem was not simply the list but the combination of a frame skewed towards higher-income households and a low response rate that skewed it further, with respondents differing systematically from non-respondents in the same direction. Kish’s treatment of survey sampling formalises the distinction that keeps being lost: sampling error is reduced by size, while coverage error is a property of the frame and is unaffected by it, so a huge sample from a defective frame is precise about the wrong population. The 2026 version is a frame made of app users, panel members and people who answer unknown numbers — and its exclusions are less visible than a list of car owners.

Why it stays hidden

The defect hides because the reported statistics are honest. Margin of error is computed correctly, weighting is applied, and none of those operations reference the frame. Anyone auditing the arithmetic finds it sound, and the question of who could never have been sampled has no field in the methodology table.