You get the raw data, every X voter at a small set of precincts. You know the demographics for respondents, and you know the group that poll workers put all non-respondents into.
So you know in precinct 212 that for those who answered the poll 83% of (D) went for Barrett, 17% went for Walker; that 84% of (R) went for Walker and 16% went for Barrett. What does that get you, when 30% of those asked didn't respond at all?
Well, you know, for respondents, how many blacks and women and people in each age group went for Walker or Barrett. And you know how poll workers categorized the 30% of voters that didn't respond--black/white/Latino/Asian, male/female, age group. Even some idea of SES because of dress. You don't know if the blacks avoided the poll worker because they voted for Walker, or if the 30% were mostly (R) that refused to answer because the male poll worker was wearing an earring.
You made an assumption. Was it right? You don't know. And can you extrapolate that precinct's raw data?
Well, last 3 elections, on averge the blacks in Precinct 212 voted just like the blacks did statewide. So you know how blacks voted this year. Right? Well, last election was an off-year election, and were blacks as motivated to vote as this year? The previous one had Obama running, with very high black turnout. So that "on average" has a really large standard deviation. This year the numbers are fairly meaningless. You picked this district because the white voters were rock solid in their distribution the last three elections, not because of the black voter turnout or results. Yeah, there's data you can extrapolate--but it's not straightforward. You need to make an assumption. Was it right? You don't know.
The raw data gives you precious little. You need to have a model to interpret the data! So those early, first results are still based on a model. But it's a model that they adjust as time goes on.
The more they adjust their model, the more assumptions they make. And, at the end of the day, they have data from perhaps 5% of the precincts and make a prediction--using their model. If they're at odds with the election results, there are a lot of assumptions behind the pollster's prediction, from a valid computer program to poll workers that followed the protocol to no systematic bias and a correct model. The only assumption behind the election results is that nobody tampered with the results in a way that the final canvassing of the voting machines won't turn up. (And most 'election fraud' that I've seen here was turned up during the official count, not the rushed early count.)
In the WI recall election, more than one precinct had massive new voter registration. Models are based on past elections, so anything with a lot of new voters renders that assumption problematic. They were single-issue voters, whatever their issue was, and that makes predictions about voter demographics harder. Lots of assumptions--so little validation. That makes for nervous pollsters because the data are useless with a valid model.
In other words, you don't know if your prediction was right or wrong. There are external polls, but they have their own biases. (Note that with the WI elections, the external polls mostly side with the election results.) So it's a damned good thing that the prediction isn't even the main reason for the exit polls. I know, it's the reason we want them to have to running the polls. But it's not.
Instead you get a wealth of data on who votes for whom and why. Politicians, reporters, social scientists all want this data. Homeowners voted for Obama a lot less than apt. dwellers? There's a mortgage interest deduction tie-in. Those making over $200k/year voted 60% for Obama and self-identify as (D)? So much for the upper class being all stodgy (R). Those with some education or just a high school degree vote mostly (R), but high school drop outs and those with grad degrees vote (D)? Good luck classifying (D) as all educated, and lack of education equating to being (R).
Except that if you don't know if your prediction was right you don't know if any of this kind of data is right. However, given facts you can validate your model and adjust your model. This is what they do--not given election results for the state, but for each precinct that they polled. In the end, the adjusted numbers can still be off--yeah, you fit your precinct data to the final vote, but there's still error. It's in the big name result figures, it's in the details. Why?
Because a lot of voters don't answer. Paragraphs earlier I said you knew some information "for respondents." But there are always non-respondents. Often 10% or more are non-respondents. Sometimes 30% or more. Imagine: For 30% of your sample you have virtually no data. Cool!
Now, you need to randomly sample a precinct. If your sample's not random, you don't have anything close to a valid sample. You may pick every 10th voter; given a few hundred voters, it'll do. If one #10 doesn't answer, you don't ever just pick the 11th. That lets the respondent self-select and that's non-random. But you can't let 10, 20, 30% of your data walk away.
There's a work-around. Poll workers eye up the #10s that get away and class them. "Okay, that's one's black, late 20s, female, has a wedding ring and seems to have a small child with her." Plop goes that #10 into the right "box." No clue how she voted, but at the end of the day the non-respondents will be assigned Walker/Barrett votes based on how the known respondent pool voted. (Just think about the ways the assumptions for this could be wrong!)
Even knowing the final results per the election board you've just estimated from 10 to >30% of your voters. That means estimation and rounding for each little subgroup. When you're done and you add all the subgroups together for a precinct, and all the precincts together, you're no longer likely to match the final result #s. And they're the numbers you have the least error for.
Europe doesn't have this problem. They sample nearly every precinct. They have more data and less extrapolation, making for fewer assumptions. They have greater buy-in because their exit polls are quicker because they collect less data per person and their elections are simpler. And yet they still blow some really close elections. We just don't notice because it doesn't fit our preconceived ideas.
Similarly, American election polls produced some hilarious results. Decades ago. In one case, commentators were struggling to explain a predicted surprise win for a (D) candidate based upon a surprisingly large black and (D) vote and lowered white and (R) vote, only to have the (R) win in the end. The exit poll workers polled near the front of the polling station. The polling station was urban, in a depressed minority neighborhood, with a bus stop out front. The pollster identified the anomalous precinct among the few precincts polled and interviewed the poll workers. They followed the rules. Then the pollster went to the polling station itself. The front door was by the bus stop; pedestrians used it, so people living nearby used it. The back door led to the parking lot. If you came from the middle-class community more than walking distance away you were almost certainly white and almost certainly (R). And almost certainly drove. The poll workers didn't get a random sample: They basically only got the occasional white, (R) voter who parked on the street. (Now we'd scream "election fraud"--and the raw data, coupled with a really bad assumption, would certainly support it!)
= new reply since forum marked as read
