2004 Presidential Electoral College Predictions

Jump to: Intro | Simulations | Basic prediction | Raw data | Acknowledgements | Home

From Andrea Moro, University of Minnesota. Last auto-update: November 02nd, 12:26 AM Central

The new 2008 Presidential Election Forecast page is up !

Note to readers

Last auto-update of regular polls: November 02nd, 12:26 AM Central

I wish to thank everybody who sent comments and requests for improvements. At some point in the near future, and time permitting, I plan to post an analysis of the performance of various pollsters in predicting state level outcomes.

Summary

 * now renamed as www.nowchannel.com

 ** realclearpolitics.com using latest state polls, not their RCP average

*** tradesports.com prediction based on an idea by Chris from noteconomics.blogspot.com

Many questions are answered in the introduction below.

Scenarios

* current state poll in parenthesis

This figure displays the recent trends in Bush's winning probability (only one data point per day per source).

Introduction

This page presents some predictions of the 2004 presidential electoral college outcome (if elections were held today) based on state level polls collected by http://2.004k.com and realclearpolitics.com. At the first minute of every hour, a script fetches those site's polls table. If the script manages to download all pages (sometimes there may be a connection problem), it strips the data from the html and saves it, computes simulations of election outcomes (details below), generates this html page and sends it to the web server for your enjoyment.

Please use caution in using these results to predict who will win the election in November (see notes 2, 3, and 5 below). My computations assume that the only problem with these polls is sampling error. That is, every voter has an equal probability of being sampled. Obviously, pollsters may also make systematic mistakes (for example, they may not sample enough young voters, or they may have a biased method to pick likely voters ...). These mistakes are not corrected by my methodology. I prefer to view these results as a tool to interpret current state polls.

FAQ:
- why have you started tracking another source of polls? Are these sources biased? Probably not (see Note 9)
- why do the numbers change so frequently? (Note 6)
- what is the expected winning margin and why is one candidate ahead? (Note 7)

Simulation-based predictions

Polls report only imprecise information about the electorate's opinion. Using simple statistical techniques, however, it is possible to use the poll results together with the reported margin of error to compute the probability that the at least 50% of the population is in favor of one candidate. I compute this probability in each state, and use it to simulate several electoral college outcomes. Then, I compute the fraction of elections won by each candidate and other summary statistics. For a more technical explanation, see Note 1.

The following graphs show the probability distribution of the difference Bush-Kerry electoral votes. Each histogram represents how likely each electoral college outcome is. The probability that Bush wins is the fraction of the area to the right of zero (the dark line) out of the total area

www.realclearpolitics.com

2.004k.com

The trends are at the top of the page

Basic prediction from polls

The candidate with more preferences in the state poll gets all state electoral votes (see note 2)

Sometimes one candidate is ahead in this table, but his probability of winning computed from the simulation is not above 50%. Why? See Note 8

The figure displays the trends:

Raw data

Notes

1. I compute Bush's vote share out of the total of Bush+Kerry shares (i.e. I ignore Nader's share and undecided voters). I assume that, in each state, the sampled Bush's vote share follows a normal distribution, with mean equal to the share and standard deviation equal to the reported margin of error divided by 1.96 (see note 4). I then compute the probability that Bush's vote share in the population is greater than 50%. Finally, I use this probability to independently simulate 100000 election outcomes in each state. Using the results from the simulations, I count the electoral votes in each simulation and report summary statistics in the table and figure below.
2. I assume that the state winner gets all state's electoral votes. However, Maine (4 electoral votes) and Nebraska (5 votes) use a winner-take-all system at the congressional district level, with two electoral votes going to the statewide winner. This introduces a possible error in the prediction of vote difference of no more than 3 votes. Colorado is voting on an amendment that would split its 9 electoral votes proportionally according to statewide results (starting this year). In case the amendment passes and Colorado turns out to be decisive, constitutional challenges are likely.
3. It is not clear how to count undecided votes. This article argues that in an election with incumbents, the incumbent's share in the last poll is a good predictor of his share in the final election, meaning that all undecided eventually end up voting for the challenger. If this were the case, the predictions reported in this page would be biased in favor of Bush. Unfortunately, I cannot in most cases distinguish between the shares for Nader, others, or undecided, which means that I cannot run a simulation under this assumption.
4. Pollsters report the margin of error to provide a 95% confidence interval around each candidate's reported vote share. There is a big variance in prediction across pollsters and time, as documented by this table from electoral-vote.com. Somebody believes that the margins of error are understated by a factor of 1.4.
5. Please note that results are accurate as long as the polls are accurate, and as long as the page I use to fetch poll results is accurate and unbiased in the choice of polls to report. While I have no reasons to believe it is not, it is also true that a particular choice of pollsters over the 50 states can skew the outcome of the electoral college one way or the other. The simulation based prediction is to some extent robust to this kind of data-coaching, but is nevertheless affected by it.
6. On October 21, some readers noticed huge swings in Bush's winning probability. First, Florida went from a 1% Kerry lead to a 2% Bush lead with only 2.8% margin of error. This caused the probability that Bush wins to increase from about 40% to about 60%. In the evening, first Ohio went from Bush to Kerry, then Iowa from tied to Bush, which generated two more swings in the winning probability. This was not a mistake, these are pivotal states and the simulations are sensitive to even small changes in winning margin (the figure does not report these swings because it records only the probability at midnight of each day). If this causes you motion sickness, I suggest you rather look at the average winning margin and its confidence interval. Think about it this way, consider the case when Florida goes from Kerry to Bush as a result of a change in the poll's prediction. Then, simple counts of the electoral college vote imply that the whole electoral college goes from Kerry to Bush if Florida is pivotal. All my computations told you instead is that the probability of Bush winning went from 40 to 60%, only a 20% change. It is evident that a few percentage points change in Ohio and Florida can cause big changes in the probabilities of winning from the simulations. The reasons are obvious to those who know the math behind these computations. That is why I prefer to look at the confidence interval around the expected winning margin.
7. The expected winning margin gives an idea of what we should expect the winning margin to be. It is computed as the the average of the winning margins in all simulations of the electoral college. The 95% confidence interval reports the accuracy of the prediction of the winning margin.NOTE that the average winning margin may be in favor of Kerry even if the probability that Kerry wins is below 50%. This happens when Kerry wins less than half of the simulations, but the winning margins in those cases are larger than Bush's winning margin in the rest of the simulations. A possibility when the probabilities of Bush or Kerry winning are close to 50%.
8. In some circumstances, the simulation reports one candidate with probability of winning above 50%, but the same candidate is not ahead in the basic predictions table. This is one of the cases where my page provides the best contribution to understanding what is going on in the polls. This may happen when one candidate is winning by small margins in states that have a majority of electoral college votes. The other candidate is winning by wider margin in the other states. The basic prediction does not care about whether the lead of one candidate is 1% or 30%: it assigns all state votes to that candidate. However, intuitively, the probability that a candidate wins that state is different if the polls reports a 1% lead or a 30% lead. Therefore the probabilistic prediction may give a probability above 50% to the candidate that is actually losing a majority of electoral college votes.
9. I have not started tracking more polls because I thought either of the polls aggregator was biased. The truth is, there are plenty of polls to choose from, and both of these sites provides a great public service. By tracking two sources I thought I could provide a better picture of the significance of the simulations. In both of the simulations derived from the two sources, zero is within the confidence interval by a wide margin. I have been monitoring these sources for a few days and it is pretty obvious by now that none of them is particularly biased. There are several circumstances in which the allegedly Kerry-biased source has one state poll with bush winning, while the allegedly Bush-biased source has the same state going to Kerry.

Acknowledgements and links

I thank 2.004k.com and realclearpolitics.com for letting me use and download their data. I also thank colleagues Erzo Luttmer and Chris Phelan for help and discussions on how to set up the simulations.

Other polls sites

• http://www.electoral-vote.com/

Other probabilistic-based predictions (results may differ widely because of different polls used and different update frequency)

• Sam Wang
• Matthew Hubbard
• Larry Allen
• Ken Martinian
• Srikanta Mishra

email home research