Des Moines Register / Mediacom Iowa Poll, final wave before the 2024 election. 808 likely voters, published 3 November 2024. Below are two versions of its method note. One is what was published. The other is what three political scientists computed afterwards, using the same interviews. Find the line that differs.
Version A
Version B
Same interviews, same week, same people. One line differs. What do you think it changes?
The same 808 people, interviewed in the same week, by the same firm. One line in the method note changed, and the winner flipped. The margin moved nine points.
Version A is what J. Ann Selzer published. She has weighted on demographics only for decades, for a stated reason: she "has not found voters' recollections of their past votes to be highly reliable" (DeRagon, Osborn and Lewis-Beck, 2024). Version B is what three University of Iowa political scientists computed from the same data afterwards.
Version A missed by 16 points on the margin. Version B missed by 7. One line in a method note got nine points closer to the answer and still left seven on the table.
That is the article in one case. Weighting on politics helps. It does not repair. And the choice of whether to do it, which target to use, and when the recall was collected is the most consequential invisible decision in every poll you read.
Two cautions. The reweighted figure is reported by DeRagon, Osborn and Lewis-Beck (2024) from a table in Selzer's own technical review, not read here first-hand: two searches failed to find that review online. And AAPOR (2025) lists Iowa among states where individual surveys erred by 10 points or more, so this case is vivid partly because it is extreme.
Sources. Des Moines Register / Mediacom Iowa Poll, final wave, 3 November 2024. DeRagon, Osborn and Lewis-Beck (2024), Sabato's Crystal Ball. AAPOR (2025), 2024 post-election evaluation. Full entries in the article's references.