In a linear fit (if I’m remembering my math) adding in a point on the fit won’t change the fit at all (although it would, erroneously, increase the R-squared). So adding in the “projected” 2008 won’t change the 2012 expectation.
Yes, of course that’s right. I assume the second 2012 in the fourth paragraph is a typo?
ETA: And I guess it goes to show that my intuitive extrapolation was not sufficiently accounting for 1992.
Yes, that should read: “As an aside, the “excluding 2008” model would have given a non-white turnout of 78.5% in 2008 (rather than the 75.7% actually found).”
I don’t think any of this matters as the only swing state with a sizeable AA population is Virginia at 19%. Ohio has 12% but it is moving in Obama’s direction. Florida I consider safe for Romney.
The real tossup states like Iowa, New Hampshire, and Colorado have almost no AA’s at all.
Let’s remember the context of this discussion.
Someone - I think you - cited some Obama people as alleging that the Romney people were projecting lower minority turnout in 2012 than in 2008. We don’t know the basis for the Romney people’s alleged projection. But the question was whether this assumption can be refuted based on observable facts.
You claimed that the historical pattern proved that the minority turnout would have to be higher in 2012 than in 2008. I rejoinded that this pattern might be thrown off by unique circumstances in 2008.
I’m not claiming to know how likely it is that 2012 will be lower than 2008. Only that your claim that it was clearly unlikely based on the pattern since 1992 doesn’t hold.
I don’t know what you’d extrapolate it to be. As previous, I don’t think there’s a compelling reason to assume the pattern is linear, and in addition, I think you need to model African Americans separately from Hispanics, and you also have to look at the growth of the underlying populations over that time versus just extrapolating from historical patterns.
Again, I have no idea what the minority turnout will be in 2012 and whether it will or will not be higher than 2008. All I’ve said in this discussion is that the assumption that it will be lower than 2008 - FWIW - cannot be refuted simply by pointing to the fact that it has consistently increased since 1992.
Interesting thing is that the polls are starting to converge somewhat, at least the ones on RCP.
Romney’s lead on Gallup is down from a high to 7 to only 3. Conversely, Obama’s lead on IBD is down from 4 to 2. These were the two outliers in each direction. And some of the other polls have also moved, in the general direction of a 1 point Romney lead.
Expecting it to regress to some mean appears to be at odds with the whole notion of “likely voters”. The percentage of voters who were nonwhite is a “variable,” yes, and we assume variables will settle in to their mean after some perturbation. But the percentage of non-white voters is also an estimate of a group of people, behaving. The best predictor of future behavior is past behavior. And likely voter models assume that having voted is a good predictor of likelihood of voting. In that sense, there is no really strong reason to assume that a substantial number of non-white voters who voted in 2008 will not vote in 2012. Unless white turn out is changing drastically, I don’t see why the percentage of the nonWhite vote would change so very much from that in 2008. It’ll be an interesting empirical question.
WADR ISTM that you’re just ignoring the reason and then declaring that there’s no good reason. All this stuff about past behaviour being the best predictor etc. etc. is not under dispute in this discussion.
But - as a hypothetical thought experiment - suppose you knew that in the previous election all voters in some category were paid $1,000 to vote and that in this election they were not going to be paid. Would you still trot out the stuff about past behaviour being a guide etc. and likely voter models? I assume you wouldn’t, and that you understand the concept of behavior varying under varying circumstances. In that sense, the question is whether there was anything exciting about Barack Obama being possibly the first ever African American president, such that it might constitute varying circumstances. Maybe yes, maybe not. But it’s a reasonable possibility.
Another day, another round of national trackers showing Romney up by ~1% on average and a whole slew of swing state polls (NH, CO, IA, OH) showing Obama ahead by ~2% or more.
The only “good” state poll for Romney today is a FL R+5 poll from some company I’ve never heard of. It also includes this tidbit: “The poll also suggests that Obama may be losing some support in the black community, as 21 percent of the respondents identifying themselves as African-American plan to vote for Romney.” If Romney gets 21% of the black vote in Florida I’ll eat my hat.
That’s obviously a ridiculous number.
What I’m wondering is that that poll seems to have been one of these automated phone polls, and I wonder if a stat of this sort can be skewed by people accidently hitting the wrong button.
IOW if 20% of whites mistakenly identify themselves as black and 20% of blacks mistakenly identify as white, it would seriously skew the racial stats. But if 20% of Romney supporters mistakenly say they’re voting for Obama and vice versa, the impact would be small.
I think it’s more likely a bad sample. That is, they just didn’t get enough blacks to answer the questions (or, perhaps, didn’t get a very representative group of blacks). So then they weight the sub-sample up and get a silly number for that cohort.
It probably doesn’t effect the top-line very much, but when you add that bit to the fact that this doesn’t seem to be a very prominent pollster and you can see why weighting polls based on some measure of reliability can be important to a meta-analysis.
At this point I think I’m most interested in seeing how closely the RAND survey matches the final results. It’s a quite different methodology, and one that seems to have some obvious advantages and disadvantages. It’s also showing easily the strongest result for Obama right now.
It would have to be a very very bad sample to be off by a factor of 10.
I have a hard time taking seriously a polling outfit which accepted a result like that at face value.
Interesting if you’re interested in poll methodology. At this point I would think new methodology plus outlier result equals low predictive value.
I think this explanation from Dean Chamber (the Unskewed Polls guy) should make clear why Nate Silver is not to be trusted:
Clearly, Dean Chambers has the sort of physique which inspires confidence in one’s integrity and intellect; why, he looks just like Rush Limbaugh in his prime!
If truth comes by the pound, Mr. Chambers looks like our guy!
Nate’s reaction on Twitter (he seems kind of amused):
Appropos of a line of discussion from earlier in this thread, from Politico:
[OTOH, there is also an eschange with Joe Scarborough quoted in subsequent paragraphs - I agree with Nate Silver on that one.]
Calling rough predictions wizardry is just as silly as treating them like gospel truth.
Of course you can predict the election based on polls. You can estimate the likelihood that a curveball will move the polls by 5 points before election day by looking at how often that happened in past elections. As it turns out, it happens pretty rarely, and you can quantify how often it happens and make a prediction. Even a snapshot poll is a prediction based on history, specifically, historical turnout that helps you weight the sample.
The problem is if you try to hold election predictions to a higher standard than we hold any other part of social science, or even natural science.
We cannot perfectly predict when it will rain. That doesn’t make Al Roker a wizard.
That uncertainty is priced into the forecast, is essentially the entire point of having a range of outcomes, and is why 538 has two different projections: Nov. 6 and the ‘Now-Cast’. The ‘Now-Cast’ is his projection for what would happen if the election were held today. The Nov. 6 forecast accounts for the possibility of October Surpisish events by looking at past history. The two forecasts get closer together as we get closer to the election, until they are identical on election day.
You can’t predict everything, which Nate Silver has been very clear about, and his detractors consistently ignore his admission of. In this instance, Brooks is setting up a straw-man argument saying that Silver is ‘calling’ the election, which is 100% not the case. He gives a probabilistic estimate that Obama will win the election. Today’s 75% chance of such includes the possibility that an unknown unknown will dramatically alter the election between now and then. To imply otherwise is to either be willfully obtuse or simply ignorant of how forecasting works, election or otherwise.
I love how Silver is at fault if he hedges anything AND he’s at fault if he claims to be certain.