Tired of constantly refreshing CNN election returns on election night? Frustrated by uninformative pundit blather? Crave numbers instead of spin?
Well, Stochastic Democracy plans to satiate your irrational impatience with a live election projection system on election day.
Instead of "calling" races when the probability of victory crosses some arbitrary threshold, we plan to post actual probabilities, as well as demographic trends, in near real time.
Details below the fold. or at
http://stochasticdemocracy.blogspot.com/
Come back to the site later tonight to see today's poll update.
*********Cross Posted at StochasticDemocracy**********
After nearly a year of poll watching, election night can be particularly stressful for those who follow politics. Many a political junkie have refreshed CNN's election page every 30 seconds, desperate to know how the races turned out, and are forced to rely on John King and his touch screen for wisdom. No more!
On election day, Stochastic Democracy plans take periodically post "probability of victory" for each state(Presidential, Senate, and Governor), each house district, and a couple of ballot measures.
This is essentially what CNN does, except that CNN only tells you whether or not the probability of victory is greater or less than 99%(That is the threshold for when they "call" a race).
Due to the odd federalized nature of it's elections, the United States is really one of the only countries that turns poll counting into such an emotionally charged spectacle. This, and the fact that the networks have outsourced most of this work to two companies since the dawn of time, makes it incredibly difficult to find information on the methodology of calling elections(Everything beyond "We use advanced statistics" seems to be a trade secret).
So consequently, I've had to create my own methodology(So I might be completely off base). I'll skip the heavy duty math for now(Basically, a lot of Bayesian inference), and just give a skim of procedure.
We are considering 4 sources of information(In order from highest weight to lowest weight):
- Actual live vote returns(When possible, this is done on a county level)
- Exit Polls (Exit poll numbers tend to shift through the night, but we'll try and use the best value possible)
3)Pre-Election day polls
- Regression(When possible, this is done on a county level)
The idea of using regression here is to take counties where there are returns, and extrapolate to other counties using demographic trends. For example, if the race in Indiana is very close, and the only county that hasn't submitted it's votes yet is a university town with a history of voting democratic, the regression will take that into account.
To a lesser extent, regression will be used across races as well. If vote returns in New York pick up a massive Bradley Effect, the model will downgrade probabilities of victory in the rest of the states accordingly. (But obviously, regression is weighed very poorly against actual vote returns).
Pre-Election polls and exit polls with be used as a prior, though we won't call races based solely on polls alone.
That's all the methodology I'll go over now, but I'll discuss it more over the coming day, and am always open to questions and ideas.
*********Cross Posted at StochasticDemocracy**********