November 02, 2004
Until next time...
The last days have shown a clear and unexpected trend favoring Kerry in the presidential race. Polls in key states, particularly Florida, have led us to update our probabilities of who wins these states, and as a result the "roll-up" of all these probabilities at the Electoral College level, as reported on our main page, led to final odds favoring Kerry.
Surely, this "final answer" is but a probabilistic one, and should not be interpreted as us calling the winner.
First, we remind the reader that the final number reported is a probability of winning the White House -- not a forecast of the popular vote. Subtle leads in key battleground states can quickly snowball into noticeably biased odds for one candidate or another nationwide. This mechanism is not quantified by traditional media reports. Second, we must bear in mind that this final prediction is based on opinion polls conducted through Tuesday, 4 PM PST, and are not based on actual exit polls. Most analysts agree that two factors are being left out of any analysis based on such opinion polls: (1) turn-out effects (and preliminary indications this Tuesday seem to suggest a historically high turn-out for this election), (2) biased sample effects: it has been argued at length that “unpollable” cell phone users not owning or answering a landline are mostly comprised of younger voters, and several polls have shown a distinctively Democratic-leaning tendency in this age group.
In summary, this project has now reached the limit of what it can offer, and while we attempted to assess the odds of this race, the dice are now cast -- and out of our hands.
We do hope, however, that this website and this methodology stimulated our readers' curiosity, and offered interesting insights into the mechanics of the Electoral College, and particularly the discrepancy between classical statistical reports of state-by-state polls and a more rigorous analysis of the Electoral College outcome. We will be busy performing postmortem analysis of this and past elections to refine the reliability of our methodology, particularly the aggregation of polls over time. The fluctuations of polls over time have a large impact on the smoothness of our reported probabilities since we “roll up” all the state polls into a single probability of who will win the majority of the electoral votes. Our off-season research hopefully will shed insight to improve upon such concerns. We are most grateful to all our readers for their support, comments, error corrections, constructive criticism, and thoughtful commentaries.
Until 2008, then...
How do we come up with these numbers?
If voter sentiment is static, we would treat all polls (old and current) as equal – without discounting earlier polls. However, we know that it is far from the truth. Events influence voter preferences and undecided voters make up their minds. Thus, it becomes an “art” to find the most appropriate “discount” factor so that we can aggregate polls of different vintage in each state to compute the probability that each candidate will carry a state.
Heavy discounting gives more weight to the most recent poll, while lighter discounting distributes some of the weights to older polls. One can get a sense of voter trend by viewing our computed probabilities as a function of the discount factor used.
The following table shows the sensitivity of the computed probabilities as the discount factor goes from light to heavy. Consumers of our analysis can draw their own conclusion and make their inferences based on the numbers presented herein: who has a better chance to win the election?
What we choose to present
What we choose to present
We have chosen a “medium” discount factor to balance the “inertia” of earlier polls and trends suggested in recent ones. Had we treated all polls taken after October 1 equally, President Bush would have been the run-away winner of this election (with probability of winning over 90%). This choice of discount factor is our best educated guess about the result of this election. When all the election excitement fades, we will be performing postmortem analysis of this and past election, hoping to shed light on voter trends to better treat polls of various vintages. The following chart also provides some insight as to the tightness of this race, with the following rules to “categorize” the electoral votes:
- If a candidate has a 95% chance or high to win a state, we call it solid.
- Leaning is in the (85%, 95%) interval.
- Soft is in the (65%, 85%) interval.
- If neither candidate has a 65% or higher probability of capturing that state, we put it in the “toss-up” category.
We should point out that our analysis is based on the inferred probability (inferred from polling data) that a candidate will win a state’s popular votes (thus, its electoral votes), and NOT on the exact polling statistics.
The percentages (on the map below for each state) represent the probabilities that either candidate wins the state (thus its electoral votes) cummulatively based on the state-by-state poll data.
If you are unable to view the interactive java map above, please click on the electoral college map for the static image