Ricochet is the best place on the internet to discuss the issues of the day, either through commenting on posts or writing your own for our active and dynamic community in a fully moderated environment. In addition, the Ricochet Audio Network offers over 50 original podcasts with new episodes released every day.
The Poll Perplex Returns
If you turn to Real Clear Politics today and look at the latest polls focused on a head-to-head conflict between Trump and Hillary, you will once again be perplexed. IBD/TIPP, which was the most accurate of the polls in 2012, has it all tied up. ABC/Wapo has Hillary ahead by a point. The LA Times has Trump up by six, and NBC/SM has Hillary up by seven. That is a thirteen-point spread. One of these last two polls is way off. Both may be. It could be all tied up.
If you go to the same site and look at the latest polls focused on a four-way race, IBD/Tipp, ABC/Wapo, and Rasmussen have it all tied up, and NBC/SM has Hillary up by six. The LA Times does not do a poll covering the race in this fashion.
This, too, would suggest a very close race. In normal years, the odds would be good that on 8 November we would see results more or less consistent with this. But this is not a normal year, and polling is as much an art as a science. Pollsters adjust what they learn in an attempt to make their sample reflect the public at large. When the public is in motion, when the old rules to do not apply, when Democrats are to an unusual degree apt to vote Republican and vice-versa, when turnout is exceedingly hard to predict, they are apt to stumble, and surprises can very easily present themselves.
Update at 4:13 p.m. EST. Real Clear Politics asserts on the basis of the polls that, if the election were to be held today, Hillary would win 273 electoral college votes and Trump, 265.
There is motion in the numbers that the pollsters are producing. Except in the case of NBC/SM, the motion is in Trump’s direction. In 1980, there was a sudden avalanche in Reagan’s direction at the end as voters pondered just how bad Carter had been and puzzled over the question whether Reagan was really a madman in the manner in which the mainstream media strongly suggested. That could happen this year. The polls show that the vast majority of Americans are aware that the mainstream media is in the tank for Hillary. That suggests that they are apt to discount what they are told by Pravda-on-the-Airwaves. And people do get annoyed when they think that they are being manipulated. So, who knows?
If you are a prognosticator, one thing to look for is signs of panic on the part of one campaign or the other. They all do internal polling, and they spend larger sums on it than do the outfits mentioned above. Knowing what is going on in the electorate is a life-or-death matter for the campaigns. They also poll individual states — especially, the ones that really matter — much more intensively than anyone else.
One final observation. Sean Trende, who is an able prognosticator, has an article out on early voting. He thinks it a poor predictor of final results. So, go figure.
Published in General
That assumptions intended to produce a more-representative sample are a potential source of error, and might cause more error than they eliminate, is a possibility. But that has to be weighed against the possibility that they actually reduce more error than they create.
Experimental design is full of methods other than brute-force randomization for producing more-accurate samples. For example, psychologists often use twin studies. Why would they do that if random sampling were superior?
If I were interested in finding out what all students, no matter what age, think of a prospective event intended for all ages at a school, why risk the likelihood of the random sample giving me a disproportionate number of seniors? Why not draw random samples grade by grade then compare proportionally?
If I were interested in finding out what all residence of a city, no matter where they lived, thought of a city proposal, why wouldn’t I divide the city into sampling cells of equal population, then draw a sample from each? On the other hand, if I want an opinion about a local feature that some are likely to use more than others, is it really so unrealistic to build my model with factors weighting for who is most likely to use this feature? Yes, those factors will include assumptions. But twin studies, sampling by grade, or sampling by equipopulous regions also include assumptions – they’re just assumptions we’re less likely to worry about having gotten wrong.
“Black voter turnout is reverting back to pre-Obama levels, which is a terrible sign for team Hillary. According to Old North State Politics, early voting data out of North Carolina suggests that black voter turnout is down roughly 7 points, as a percent of the overall electorate versus 2012, implying that cumulative black votes are down around 16%. As a reminder, Mitt Romney won North Carolina in 2012 by slightly over 2 points. Given that black voters usually favor democrats by 80-90%, simple math implies that a 7-point reduction in blacks as a percentage of the overall electorate would hurt Hillary by roughly 6 points versus the 2012 results.”
“The outlook is even more troubling in Florida as black composition of early votes is down 9.5 points versus 2012. According to a recent article by Politico, in 2008 and 2012, Obama received 95% of the 1.7mm votes cast by black voters in Florida. Given that, simple math would imply that a 9.5-point reduction in blacks as a percentage of the overall electorate would hurt Hillary by roughly 9 points, versus the 2012 results, which is disastrous news for a state that Obama narrowly won by less than 1 point.”
http://www.zerohedge.com/news/2016-11-02/hillary-camp-panics-early-black-voter-turnout-plunges-key-swing-states
Isn’t that why you control for variables when you’re analyzing the results? It doesn’t matter if your sample has a disproportionately large number of seniors if your analysis establishes that grade level does not affect the results.
But yes, that is assuming that your sample size is large enough to allow you to slice and dice the data set like that. I take Bereket Kelile’s point, that with media polls we’re usually not dealing with sample sizes that are anywhere near large enough. And that some of the algorithms are very sophisticated and have been fine-tuned over many iterations.
My only point was that any weighting of sample data is another potential source of error because of the assumptions it has to make. It’s a way of trying to fix a flawed data set; in an ideal world we wouldn’t have to do it. This isn’t an ideal world, and I suspect that for the polling business this election is even less ideal than usual.
On Monday, a group of ministers and theologians affiliated with the Church of God in Christ (COGIC), an evangelical Christian denomination with more than 6 million members, marched to Clinton’s Brooklyn headquarters to drop off a sternly worded “Open Letter to Hillary Clinton Regarding Religious Freedom for Black America.”
What they want, according to the manifesto, is a conversation with Clinton — and, by extension, the Democratic Party — about the hypocrisy of rounding up black votes through community churches while undermining core values of the clergy and members of those same congregations.
Up to now, pundits and pollsters have assumed that numbers suggesting a falloff in black support for the Democratic ticket are the result of young black millennials who see Clinton as too conservative and too much a part of entrenched power.
That may be true, but it’s not the whole story.”
http://www.nydailynews.com/opinion/errol-louis-black-pushback-hillary-clinton-article-1.2852671
I agree that the weighting method used could fail to produce a representative sample. Researchers certainly aren’t infallible and weighting doesn’t automatically make improvement simply by nature.
Research objectives should drive the methodology, not the other way around. And there’s multiple methods used in survey research, as in other areas. My argument is less about the design and more about how you make the sample representative. In your examples you mention different populations which raise different challenges when it comes to sampling. That is going to determine
Again, speaking from experience, I can tell you that taking a sample as is from the field without any weights to make the sample representative is just a bad idea.
“Republican Donald Trump has a three-point lead in Rasmussen Reports’ White House Watch survey. Among voters who are certain how they will vote, Trump now has over 50% support”
http://www.rasmussenreports.com/public_content/politics/elections/election_2016/white_house_watch_nov3
Guys, please! Stop citing Rasmussen, they’re not credible.
Well, again, what if your analysis that grade level does not reflect results is somehow mistaken? Controlling for variables is one technique, good experimental design another… many options here.
Understood, and that point is well-taken.
Sure. There are all sorts of interesting trade-offs between low variance and bias, for example. For some statistical measurements (I’m not saying polls, the examples I’ve heard of are from elsewhere), permitting some bias in order to get lower-variance measurements is actually more useful than keeping the variance high in order to eliminate that bias. Actuaries are sexier than we know.