The polling folks are always terrified that they will repeat the mistake of 1948 – when the Chicago Tribune gloriously announced that Dewey had won the election. Then again CBS declared Gore the winner in 2000.
The reason for the 1948 error was bad sampling – back then, many people did not have a telephone, and the polling was done by phone. The result was that richer people who were more likely to be Republican – were more likely to own a phone and were oversampled compared to the people who actually voted.
In the first day of statistics class you are given a premise that the entire science of statistics is based on. When you have a “normal distribution”, you can… the problem is there are very few things that have a normal distribution.
Let’s start with a simple case – if you know that 52% of likely voters are women, but your sample winds up with 56% women, you have a sampling error. If women and men vote similarly, no big deal – but if not, you need to figure out why women are more likely to answer the poll. One might thank women when they reach 52% and hang up or ask if there is a male in the household to improve the sample. If you are lazy, you just “weight” each vote higher or lower to fix up the faulty sample.
Rush framed the issue today much better than the foaming at the mouth Alex Jones. The problem is all of the polls are weighting based on what people state as their party affiliation. In North Carolina, Barack Obama got 48% of the vote and Romney got 50%. If you reach 1,000 people, some portion will say they are Republicans and some will say Democrats, and some will claim to be Independents – but they are lying. They are telling you which party they intend to vote for, not which party they belong to. They all will say they are going to vote, but 1/3 won’t actually vote.
So weighting the sample based on what people as a group did 4 years ago is a really bad idea. Get the age, sex, rural vs urban, race, income to a reasonable representative sample and don’t weight by party. The data coming in suggests that many blacks are not going to vote. You can’t assume if a certain portion of the NC population is black and voted for Barack Obama that they’ll turn out in the same numbers to vote for Hillary.
I hear that the Black vote is way down in early voting. That’s bad for Hillary as they now have to manufacture more pre-filled ballot boxes and tweak a bunch more voting machines. I think they have 3 million votes from dead people lined up already.
Another problem is as soon as you start adjusting your sample using more than one variable, that creates more distortion. In this specific case, there is a huge difference between women with college degrees and women without. Women without college degrees support Trump. So if you split women into two groups, you have to estimate the actual population, and see if the sample has the right number of each. Add in more variables like age and income level and you wind up with 2 or 3 people representing the entire group.
The big reputable polling groups understand most of that – most of the college and TV station polls do not. While Quinnipiac University does a pretty good job in New England, it is because they understand the people who live there. Their national polls I view more skeptically. With polls done by humans, you always have the problem of the person’s voice influencing whether the target agrees to be polled and how they answer. I suspect that most live polling is done using women. I wonder if a woman being polled by a man would get a different result, particularly on “women” issues. Undoubtedly accents play a role as well.
Absolutely. If the pollster sounds interesting or brings a good looking vision to mind, that will influence the outcome.
I say force all of that data into a bell curve and then layout the two sigma lines. Let’s see who says what in the polls 🙂
Yeah I heard that too Briand. It was like 12 or 15% down in NC and the ‘white devils’ were up 15% in NC. That is a big swing.
parrott
If I was brave and/or stupid I would go to the polling location on Wilkinson Boulevard and play Pokémon in the parking lot and watch the comings and goings.
Yeah , and see if there was any Pirate radio stations around ?
I havn’t found any here in SWVA (yet)
parrott