The generic ballot poll for the House of Representatives, which was, until this week, the best evidence for the predicted "Blue Wave", asks a sample of adults, registered voters and/or likely voters which major party they favor for "Congress" (the word of both the House and the Senate has, in modern parlance, come to mean only the House).

The logical flaw in using such a poll is that no office in the Federal government is determined by a national vote–not the President, not the Senate, not the House. But, the poll does have the great advantage of being easy. The only alternative is to go by individual House district; I am not even 30% of the way through the "battleground" House districts, and I am - for the first time in my sedentary life–beginning to have a sense of what long-distance runners mean when they discuss "hitting the wall". I promise to finish; but getting the data for each House race is a long slog.

That is why I am following 538 and taking the easy road for an hour this morning.

But, I can't be a complete cheat. No political poll, of any kind, avoids GIGO if it does not disclose its sample's partisan allocation. It will not surprise any of you to learn that the New York Times and the other most frequently cited generic ballot polls are scrupulous about not providing any such useful data. I did find one polling organization that has gone against the GIGO trend. The YouGov people have been remarkably forthcoming about how they get their numbers. In their CBS/You Gov Tracker poll, their sample data for party ID is R-42, I/O-15, D-43; for political affiliation it is Conservative-38, Moderate-32, Liberal-30. What they do not disclose is their methodology for converting these sample numbers into their overall prediction: R-48 and D-52.

I can explain how they get there, but I was not able to find any explanation of the why. The recipe for How works by taking the middle third of voters - the Moderates - and asking the Lean question in such a way that more than 3 to 1 favor of the Democrats. The result of their party ID inquiries is to assign 13 points to the Democrats and 4 points to the Republicans. That reduces the "middle of the road" share of the electorate to 1/7th. To get to their final D +4 result, their algorithm assigns 60% of the Independent/Other vote to the Democrats and 40% to the Republicans.

When YouGov did a second generic ballot poll for the Economist, they queried 1500 adults. Their bottom line result was similar to the one from their CBS tracking poll: D +6. But, here, too the How and the Why seem very much at odds. Their political affiliation numbers show a slight leftward variation on the common 1/3rd, 1/3rd, 1/3rd distribution pattern. Their sample has Conservative - 31, Moderate - 34 and Liberal - 35. But, that somehow shifts to a Party ID landslide for the Democrats: Republican - 23, Independent - 42, Democrat - 35. Where the numbers get really strange is in the data for 2016 voting. The 931 respondents who voted in the last Presidential election produce a sample that is divided Republican - 45.8 and Democrat 54.2.

But we know that the actual vote count was Republican - 62.98M and Democrat - 65.84M. That was Republican - 48.9 and Democrat - 51.1. Somehow a sample that should, to be representative of known ballot numbers, have a D +2.2 margin ends up with a D +8.4 spread.

No wonder the statistical affiliates of the party of Native American heritage prefer national algorithms above all else. They are not only easier but they always give the right answer.





Speak your mind


Resources & Links