Introducing Matt Hubbard’s Weekly Electoral College Status. From Matt:
The idea for this website came from two sources. The first was Ryan Lizza’s webpage during the 2000 election where he analyzed state polling data and came up with the remarkable prediction that it all hinged on Florida.
The second idea came to me last year when I was teaching statistics for the first time. The thing that struck me is how few people – including myself, my students and several good mathematicians I know – knew about the margin of error being tied to a confidence level, and that the standard confidence level is 95%. This means that if a candidate has 47% in a poll with a +/-3% margin of error, then 95% of the time, the true polling numbers should be between 44% and 50%; the other 5% of the time, the polling numbers may go above 50% or below 44%, and it is assumed both the high and the low have the same probability, 2.5% each.
You can change the margin of error with poll data, but it will also change the confidence level. My idea is to make the margin of error half the lead; this way the candidate with the lead in a particular state should win if his total stays inside the new margin of error or goes higher than the new margin of error, and he will only lose that state if his totals go below the new margin of error. (This assumes that percentage gains for one candidate are reflected in percentage losses for the other; while there are more than two candidates, they are getting only marginal numbers; if we had an election like 1992 where Perot was getting double digit numbers in many states, I would have to change my assumptions considerably. The math wouldn’t be impossible, but it would be much harder.)
Here’s an example of the method: Let’s say Candidate A leads Candidate B 47%-45% in a poll with a 3% margin of error. The margin of error should always be the 95% confidence level, or +/-1.96 standard deviations (SD) around the statistic; I then use that to get the confidence level for a margin that is half the lead, which in this case is 1 point; 1/3 ? 1.96 = .65; taking the integral of the normal distribution from -/+.65 SD, we get that about 48.4% of the time Candidate A will stay inside the 48%-46% range and Candidate B will stay inside the 46%-44% range and Candidate A will win; it is also possible Candidate A will do better than expected about 25.8% of the time, and also do worse than expected 25.8% of the time. This adds up to a 74.2% chance for victory for Candidate A and a 25.8% chance of victory for Candidate B in this state.
I collect data from state polls and put the information into a computer program written in C; if the leader’s chance to win a particular state is better than 99.5%, which is to say if half the lead is greater than 2.6 SD, then I consider that state a lock and add the electoral votes to the party’s total. For all the n states that are not locks, I then put them into a calculation pool of 2n¬ possibilities and figure out the probability for each of the possibilities and add that probability into its proper category, either GOP victory, DEM victory or tie.
For this week’s stats, Matt writes,
For the first time since the Republican National Convention, my data shows Kerry in the lead. It’s still very close and hinges largely on the outcome of Ohio, but it looks like those undecided voters (who are these people?) are starting to see Our Resolute Leader as petulant and impatient instead of forceful and resolved.
Write to Matt at [email protected]