Reply #15: it's a data problem, not a statistical problem [View All]

I imagine that the die was cast when the question wordings were set: "Did you vote for the Republican? The Democrat? Or someone else?"

I do consider SurveyUSA a legitimate firm, but in this case someone let you down. (Their methodology is controversial, but I have no reason to think that that is relevant here.)

But that part is speculative, although the inadequacy of generic questions is established. What I can say is that the small differences in the robopolls between the Senate and governor races in Hardee and Okeechobee are hard to reconcile with pre-election polls and historical precedent, as well as the official returns. The rate of ticket-splitting should be much higher. There isn't enough information to say exactly what respondents were doing, but I surmise that to some extent they were reporting that the incumbent(s) they liked were of the party they liked.

You're right to wonder, if this is a bad-data problem, how you ended up with statistically significant results. I think it was a fluke of the design: of your six competitive statewide races, five were paired with races that strongly favored the Democrat. That could be because of massive miscount, but looking at the races, I doubt it.