Take The NBC/Marist poll from late August. It said that the voters in Texas were nearly split, 49-45 (advantage Cruz).
Read the fine print and that's including those who are "leaning" to one or the other, however weakly. 6% undecided when asked if they had any leaning.
The Emerson poll from about a week later had Cruz up by 1%, the MOE being over 4%. So "neck-and-neck" (when we need to include the MOE). But if you actually pay attention, it was 39 Cruz, 38 O'Rourke. 33% of the electorate wasn't in either camp, but the Emerson poll didn't follow up with the equivalent of "so, in a pinch, who'd you think you'd vote for if you had to vote, and the election was today?"
And that was in August, 2 months ago, and people weren't really paying attention too much since then. Hence the "33% are undecided" from the Emerson poll. Since then there's been Kavanaugh; O'Rourke's background's a bit better know, his "I crashed while drunk, and only really admitted to driving while a bit inebriated" news story broke, and it looks like affluenza to those not disposed.
Mostly the undecideds are saying who they've decided for. Now, whether the models that the pollsters are using are anything close to reality, that's another question. Remember, raw poll numbers are pretty much trash unless the sample is very, very large and demonstrably random. When you poll 1000 people, you then have to put their demographics into the sausage-machine called a "model" and see how the various weightings--all of them based on assumptions about the future, and mostly based on the past--change things. A poll of a 1000 people is only okay unweighted if the total population is maybe 1050-1100, and even then you have to make sure that nobody's left out in a systematic way; or, if the people are pre-selected based on the model, so you do your weighting before you ask the questions.
= new reply since forum marked as read
