Arkady
President
Throughout this election cycle, the LA Times/USC tracking poll has been an outlier in favor of Trump. For example, the RCP 10-poll average, including LA, currently shows a 5.8-point spread for Clinton, but if you exclude LA, it's 6.8 for Clinton. LA has Trump up by 2, and the best he's doing in the other nine is trailing by four (Fox News, which tends to lean a little right). If you take all the polls going back to mid-May, Clinton has averaged a five point lead, but in LA polling Trump has averaged a 1.9-point lead, so this isn't just a recent disparity between LA and all the rest of the polls. It's a pretty consistent gap of almost seven points.
So, I was curious about what as going on. It turns out that LA uses a method none of the others use, where people indicate voting preference with a probability -- for example, you self identify as being 80% likely to vote for a given candidate. This experimental method should be superior -- if people are honest, numerate, self-aware, and thoughtful. Even if they're not, it still might be superior, if people are equally likely to be those things on each side of the aisle. The problem comes if such symmetry doesn't exist.
Consider, for example, a simplified poll involving just my wife, a Clinton supporter, and my loony Trump-supporting cousin. My wife is a thoughtful and measured person, and a very honest one. If asked to identify her likelihood of voting for Clinton, she'd try to give her honest best estimate. Maybe my wife will die or become incapacitated or some emergency on election day will keep her from the polls. Maybe Clinton will do or say something so horrible that my wife feels compelled to vote third party. Maybe Trump will drop out and be replaced by someone decent enough that a third-party vote will look better to my wife than adding to Clinton's mandate. There's no way my wife will say 100%, when those are all possibilities. So, she might say something like 84%. As for my cousin, he's neither honest or thoughtful. He'll see it as his job to help create a winning image for Trump, so he'll just say 100%, even if he's actually more likely to flake on Election Day than my wife is.
If Clinton supporters tend to be more like my wife and Trump supporters more like my cousin, that can make a big difference. If there are 50 people leaning Clinton and 44 leaning Trump (the Quinnipac polling result), and those 50 Clinton supporters get weighted at 84% and the 44 Trump supporters get weighted at 100%, then it shows up as a 42-44 Trump lead, as with the latest LA poll.
I wouldn't be at all surprised if there were such tendencies in the broader voting populace. There seems to be a greater affinity for nuance with liberals, whereas conservatives tend to like black-and-white thinking. There are some other issues around party identification and how modeling is done in the poll, as well, but I'll leave those to the technical pollsters to hash out. Essentially, the LA poll weights its findings based on self-identified 2012 votes, rather than current party identification as is typical, which has some problems based on people mis-reporting past votes.
This election will test whether this particular experimental approach to polling is worthwhile. Since this time last month, there have been 35 national polls released and Trump won only five of them -- every single one being an LA Times/USC Tracking poll. If, on Election Day, polling looks like it currently does, and Trump pulls off the upset, LA's method will presumably be broadly adopted. If, on the other hand, Trump loses, and LA's the only poll that didn't see it coming, expect the experiment to be dismissed as a failure.
So, I was curious about what as going on. It turns out that LA uses a method none of the others use, where people indicate voting preference with a probability -- for example, you self identify as being 80% likely to vote for a given candidate. This experimental method should be superior -- if people are honest, numerate, self-aware, and thoughtful. Even if they're not, it still might be superior, if people are equally likely to be those things on each side of the aisle. The problem comes if such symmetry doesn't exist.
Consider, for example, a simplified poll involving just my wife, a Clinton supporter, and my loony Trump-supporting cousin. My wife is a thoughtful and measured person, and a very honest one. If asked to identify her likelihood of voting for Clinton, she'd try to give her honest best estimate. Maybe my wife will die or become incapacitated or some emergency on election day will keep her from the polls. Maybe Clinton will do or say something so horrible that my wife feels compelled to vote third party. Maybe Trump will drop out and be replaced by someone decent enough that a third-party vote will look better to my wife than adding to Clinton's mandate. There's no way my wife will say 100%, when those are all possibilities. So, she might say something like 84%. As for my cousin, he's neither honest or thoughtful. He'll see it as his job to help create a winning image for Trump, so he'll just say 100%, even if he's actually more likely to flake on Election Day than my wife is.
If Clinton supporters tend to be more like my wife and Trump supporters more like my cousin, that can make a big difference. If there are 50 people leaning Clinton and 44 leaning Trump (the Quinnipac polling result), and those 50 Clinton supporters get weighted at 84% and the 44 Trump supporters get weighted at 100%, then it shows up as a 42-44 Trump lead, as with the latest LA poll.
I wouldn't be at all surprised if there were such tendencies in the broader voting populace. There seems to be a greater affinity for nuance with liberals, whereas conservatives tend to like black-and-white thinking. There are some other issues around party identification and how modeling is done in the poll, as well, but I'll leave those to the technical pollsters to hash out. Essentially, the LA poll weights its findings based on self-identified 2012 votes, rather than current party identification as is typical, which has some problems based on people mis-reporting past votes.
This election will test whether this particular experimental approach to polling is worthwhile. Since this time last month, there have been 35 national polls released and Trump won only five of them -- every single one being an LA Times/USC Tracking poll. If, on Election Day, polling looks like it currently does, and Trump pulls off the upset, LA's method will presumably be broadly adopted. If, on the other hand, Trump loses, and LA's the only poll that didn't see it coming, expect the experiment to be dismissed as a failure.
Last edited:
