Slate/Votecastr real time election projections

[Angrie]

This is the model that big campaigns have used for years, and one of the consultants on this project is Obama's former director of microtargeting. That was a very successful electoral forecast model. It's also the model used by networks to make state calls. Exit polls are less accurate in general since their sample size is smaller than this method.

Yeah, this is the most accurate available indicator of what is the actual current situation. But it is really just *current* situation based on the modeled votes of the people who have already voted. The people who have already voted may not be perfectly representative of those whose ballots will be turned in later in the day (and probably/predictably are not).

FWIW, they showed their current county map on the Vice News stream, and it looked exactly the same as the 2012 CO map, except with Conjeos country flipped R and Alamosa county a tie.

[Angrie]

Is CLinton +2 in Colorado now, good news or bad news for her?

About in line with projections from e.g. 538. A bit better for Trump than them, but that is countered by the fact that the additional votes that come in are more likely than not to be a bit more Dem leaning.

[Angrie]

Nobody knows how reliable these slate projections are.

Wrong. They are pretty damn reliable. And certainly much more so than other things we will see later like exit polls.

[Angrie]

Haha, Dems were OK with CO projection, where Hillary did much better than Obama, but now it a trash because of Nevada

CO is not a projection, it is the modeled vote of individual people matched against the voter file who have already actually voted.

This is apparently different from other states where they do not have voter file individual level data on who has voted, but only know the total # of people who have voted in particular precincts (but not necessarily which individuals have voted in those precincts).

And including Stein in Nevada is of course obviously a dumb screw up.

And there is a difference in what early vote means in different states.

So basically there are large differences in what they are reporting for different states.

[Angrie]

Haha, Dems were OK with CO projection, where Hillary did much better than Obama, but now it a trash because of Nevada

CO is not a projection, it is the modeled vote of individual people matched against the voter file who have already actually voted.

This is apparently different from other states where they do not have voter file individual level data on who has voted, but only know the total # of people who have voted in particular precincts (but not necessarily which individuals have voted in those precincts).

And including Stein in Nevada is of course obviously a dumb screw up.

And there is a difference in what early vote means in different states.

So basically there are large differences in what they are reporting for different states.

But those with large EV, as Nevada? Does they not use voter file for EV there?

As I undrstand they use voter file for all EV?

You must be logged in to read this quote.

So it is projection. By using voter file. Other only by their model of turnout and polls

Well, it is not a projection of who is going to win the state. It is a projection of how the particular individuals that they know already voted have voted. That projection is extrapolated from a large phone poll (much larger sample size than normal polls) they did before the election, calling people in the voter file. Unlike in a regular poll, they record who each individual says they are going to support. They can then cross reference this with all the demographic information they have about each voter who answered the poll in the voter file - things they have hard data for like age, gender, vote history, and other things that are not filled out by the voter on their voter registration form in most states and so are usually modeled (like race and ethnicity), and possibly other commercial data on each person. Then basically they run more sophisticated versions of a regression of which candidate people support against their demographic characteristics. This then gives them estimated probabilities that each individual in the voter file supports each candidate.

But they can only do this when they have individual level data on which particular people have voted (like they apparently do for CO, but not other states).

So for other states where they only have a raw total count of how many people voted in each precinct, they basically just take the average support score of registered voters in that precinct multiplied by their turnout score, and then that gives them an estimate of what the vote is in that precinct. Since this is not individual level data of who has actually voted, this can be biased, since it relies on people accurately self-reporting their likelihood to vote (and self-reported voting intention is not accurate).

What I am describing is basically what Obama's campaign did. I am presuming that they are in fact doing the same thing, since that is what they say they are doing, and since Sasha Issenberg is apparently a founder of Votecastr.

So the bottom line is it is (presuming it is done properly, like Obama's campaign did) very accurate for the voters/states you have individual level turnout data for, but less so for the voters/states that you only have aggregated turnout data for.

[Angrie]

Local officials collect and report information about who voted early in each state, and VoteCastr then compares that public info with its own private early voter files.

The key point that is somewhat glossed over is that the "information" that the "local officials" report can be different in different states. In some states, that information may be a list of the voter ID numbers of which specific individuals have already voted. For other states, that "information" may just be a report that 578 voters have voted early in precinct 42, without specifying which particular voters it is who have voted.

[Angrie]

But they can only do this when they have individual level data on which particular people have voted (like they apparently do for CO, but not other states).

Are you sure, as I understand, they project early voting in other stated as well.

Projection of early voters. I didn't say anything about projection of who win the state. It is pretty meaningless.

But we can compare their projection of early voting with polls. Some polls ask voters who already voted. For instance if their projections is much Trump-friendly = polls probably underestimate him. And vice versa.

Sorry, what I mean (but forgot to explicitly say) is that once they have the modeled support score for each registered voter in the voter file, they will have data such as Amanda X has a .7 Clinton support score. That means they estimate that there is a 70% probability that Amanda X supports Trump. And they have data that Joe Y has a .4 Clinton support score.

So then if they have individual data for who has voted, they can then just add up each individual's support score and average it by the total number of people who have voted. So for example, if the only 2 people who have voted in a precinct are Amanda X and Joe Y, then the estimated vote is (.7 + .4) / 2 = .55. So they estimate Clinton has 55% support there (out of 2 votes cast, 1.1 votes for Clinton, .9 votes for Trump).

But you can only do this if you know specifically which individuals have voted - if you know that Amanda X and Joe Y are the particular people who have voted.

If you don't know that, but instead just know that 5 people have voted in Precinct Z and 3 people have voted in Precinct Q, then what you do for both of those precincts is to take the support scores of all the registered voters in each of those precincts weighted by their turnout scores, and that gives you your estimated vote percentage in that precinct. Then you can get your statewide or countywide vote estimate by averaging those weighted by the known aggregate turnout in each precinct. For example, if the average turnout-score weighted support-score in precinct Z is .7 Trump and for precinct Q is .2 Trump, then the estimated vote would be .7 * 5 + .2 * 3 = 4.1 votes estimated for Trump. Since there are a total of 8 votes, that then means 3.9 estimated votes for Clinton (not counting 3rd parties for this example).

So that is what they do for states that have early vote, but in which the state does not report to them which particular individuals have early voted.

[Angrie]

Eh. They clearly stated that they would use voter file. If it was only "report that 578 voters have voted early in precinct 42", they'd likely tell us about it.

At least historically it has varied by state/locality what they report, and what format they report it in. I don't know which particularly states report what at this point on the early vote - which are reporting exactly which individuals have voted, and which are reporting only an aggregated count. Since they said specifically this was the case for CO, we know that, but we should not assume that it is the case for states that they have not specifically said this.

But for the election day vote, all of it should be just this aggregated count using the method I described, coming from observers they have at selected precincts in each state.

[Angrie]

Uh-oh, they have Trump ahead 46-45 in Iowa now that some of the election day vote has started...

OH has also narrowed to just 46-45 Hillary with some election day vote now feeding into their model. I wonder if it will flip when they get more vote in...

But CO has expanded to 47-42 Hillary, and PA looks fine (48-44 Hillary) now that they have some election day vote.

[Angrie]

They don't even really have EV. They "estimate" it

SAD!

You must be logged in to read this quote.

So basically, the lists of voter ID #s of people who have already voted is reported in a decentralized manner by country clerks/election administrators and the like, and some of them are slow in reporting this. So they may have data that is up to date for some counties, but a day or two behind for other counties. And perhaps entirely absent in other countries (more likely to be small rural counties with an election administration consisting of one person who also does 5 other things).

And there will be other such variation from state to state as well.

[Angrie]

FL is shot. It's over if they are right.

Los Hispanos estan Votando, y no estan Votando por el pequeno Donald...

[Angrie]

What is up with Menominee County WI

Modeling...

Yes. When you have a small subgroup that is way overconcentrated in some particular way, you will get weird results. It works by averaging things out, so it will be most accurate on larger scales and for larger groups.