Again, the mistake is that you think those folks were “all Obama” a month ago. Most probably they were all “Obama? meh, I guess…”.
Only in some cases. Some pollsters ask a series of questions, and screen you based on your answers to those questions. A lot of those questions have to do with past actions. Those obviously wouldn’t show any change with new rounds of polling.
I don’t think any pollster’s likely voter model ignore or override the person’s own subjective likelihood to vote. I think instead they tend to seek corroboration of that expression in past voting. Do you know of any that would find someone to be a likely voter even if they don’t say they’re extremely likely to vote?
I dunno. I thought the purpose of the questions was to get a more “objective” answer than someone just saying whether they were likely to vote or not.
It is, but it only functions in one direction (for the ones I’ve seen).
So if you say you’re super likely, but you’re 30 and you’ve never voted in a general election, you’re not gonna get counted. But if you have voted in every election since 1960, and yet you’re not sure if you’ll vote this time, you also will not be counted.
I think there’s two things that a lot of people miss about models like Nate Silver’s.
- The model is not intended as a prediction of the winner or the likelihood of a given candidate winning. It’s an assessment of the likelihood of a given candidate winning if the election were held today. This distinction is highly significant. Because the probability that the models put on a given scenario is based on the probability that the current (weighted average of) the polls are wrong. That’s very different than the likelihood that the public sentiment might change.
If you have enough reliable data, you can have a situation where the likelihood of the actual results being different by even a small amount is very low. That doesn’t mean that the likelihood of the public sentiment changing by that small amount is equally low, or anything close to it.
- I’ve not looked too deeply at the intricacies of Nate Silver’s model, but it’s my understanding that he is not modeling popular vote but rather the likely winner of each state, and then putting together a composite likelihood of any given electoral vote scenario happening. This impacts the results, when looking at the likelihood of the polls being wrong. Because you’re looking at the likelihood of a whole bunch of state polls being wrong, versus the likelihood of much fewer national polls being wrong. This too has the impact of magnifying the likelihood that the front runner will actually win, which makes it that much more sensitive to small changes in public sentiment.
Another way of looking at it is to consider that when Obama was the 89% favorite to win (or thereabouts) it’s not like he was leading in the polls by 15 points. He was only a few points ahead, but the plethora of polls made it possible to declare an 89% likelihood that he was the true front-runner (at that time). But since this was based on a very small lead, that likelihood could shrink a lot with only a small movement in the underlying poll numbers.
Both descriptions of Nate Silver’s model are incorrect.
See herefor complete methodology.
I don’t see anything in your link which contradicts anything I’ve written. Perhaps you can be more specific.
Sure.
You’re wrong that he models “the likelihood of a given candidate winning if the election were held today.” That’s what’s projected under the “now-cast” tab. But the main model is projecting the result on election day, and often varies from that shown in the now-cast.
And you’re wrong about the methodology of the model. It incorporates both national-level popular vote polling and state-level polling.
AFAICT - based on your link - he’s basically allocating undecided voters (which is why it tends to vary by very small amounts). He’s not predicting the likelihood that voters might change their minds based on future events, which is what’s relevant here.
No, I think you’re wrong. He incorporates national-level popular vote polling, but - to the extent that it impacts his “predicted winner” numbers it’s only used as a way of adjusting or validating the state numbers. (He’s obviously be a fool if he did otherwise, because the winner is determined by the electoral college.)
No. He has examined how predictive polls were at a given distance from the election date, and weighted them according to their predictive value at that date and added other factors like economic statistics. By weighting the polls that way based on past predictive power, he is inherently accounting for the likelihood of shift before the election.
You’d need to read his detailed methodology posts for further explanation, which I mistakenly thought were included in the link I posted, but which are not.
That’s correct, and contradicts what you earlier wrote. Obviously, as you now acknowledge, he’s using the popular vote as a way to control for error in the electoral college prediciton.
Well if you can find where he writes what you say by all means please post it. (I’ve used up all my NYT free articles for the month so I can’t search for it, but I can access it if you link it here.)
That does not contradict what I wrote.
Despite the use of the popular vote, the odds are still longer on two states being off than on one. This is basic mathematics.
To use a very simple example, imagine you take polls in two states which each show Candidate X leading 51-49, and then you take an aggregate two-state poll which also shows a 51-49 margin, validating the individual polls. Now consider the likelihood of the following happening in an election held that day: A) Candidate Y wins one of the states. B) Candidate Y wins the combined two-state vote. Which possibility faces longer odds?
Can someone who isn’t as inept as I am with math help me with this?
Okay, if you use the Census Regions someone else posted, the NE has about 55 million people, the MW about 66 million, the South about 114 million, and the West 71 million. If you polled proportionately so that the polls reflect actual numbers of people, you would poll twice as many times in the South as you would in the NE or MW, and half again as many as you would in the West, right?
But if you look at electoral vote distribution, that would mean that you would be doing a ton of polling in very very red states. Almost 40 percent of it.
But the swing states are scattered across the country, with three in the MW, two in the West, three in the South, and one in the NE.
Can we extrapolate anything from a national poll like this to see state polling results? Because my very flawed eyeballing of it would be to say, “Romney running up the score in states where he was already expected to win big is immaterial.” But then I think, “I’m sure Gallup wouldn’t ignore swing states.” But Gallup doesn’t poll states, right?
So I’m just confused in general about what a national poll represents.
I’ll pass. You were the one making the claim which you’re free to substantiate, I just wanted people to know that it wasn’t correct. Anyone who doubts me can Google for the posts (which are not charged to the paywall limit). You can use the title “Election Forecast: Obama Begins With Tenuous Advantage”
It might not contradict what you meant to say, but it does contradict what you did say. Do I really need to quote the post back to you? I’d rather not. I’m just frustrated with people misrepresenting the model. There’s plenty to criticize without guessing at what it does or making things up.
I actually did google it a bit and couldn’t find anything that substantiates what you said. Whatever.
It doesn’t contradict what I said. And there’s no need for you to be so sensitive about Nate Silver’s model. I didn’t criticize the model, I was merely pointing out a couple of reasons why his probabilities are very sensitive to small movements in the poll numbers. Which is the reason they moved so much after the first debate.
I’m sensitive to not misrepresenting what we’re discussing. It could be 538 or ElectionProjection.com (it’s GOP equivalent) for all I care. I’ve got no dog in the model wars.
I’m just suggesting that before discussing the model’s methodology, you might want to actually find out what it’s methodology is. That’s all.
I grant you that the blog doesn’t make that very easy at the moment since the methodology posts are not aggregated anywhere. But I don’t think that gives you license to just guess (incorrectly, as it turns out) that the model is forecasting the result if the election were held today, or that it is “not modeling popular vote but rather the likely winner of each state, and then putting together a composite likelihood of any given electoral vote scenario happening.”
Thanks Rich (and others!) for the break down. Makes more sense to see the polling as “enthused” vs “not so much”. And it would explain the large drop.
I wonder tho, do the pollsters call back previously called folks? If prior to the first debate the Obama support was enthused and the explains the high number (to an extent) then after the debate they aren’t enthused (hence not a likely voter, and not counted) would they ever call those folks back to get a sense of the gap, or how their perceptions have changed?
There is at least one poll this cycle that is keeping tabs on the same set of voters and repeatedly calling them back, the YouGov polls. Discussed here.
Latest data released today:
Rasmussen: Romney +2
Survey USA: Romney +3
Gallup: Romney +6
If you’re going to cherry pick polls, you might want to at least state their results accurately. Gallup has it +5 today.