Friday, October 31, 2008

Betting on Bradley?

The sophisticated average of polls fivethirtyeight.com shows Obama ahead by almost 6 percentage points in the popular vote and on track to win about 350 electoral votes – almost twice McCain’s total (accessed October 31, 2008). There is sampling error in polls, so fivethirtyeight.com also simulates the sampling error 10,000 times and tells us in over 97% of those simulations Obama wins most of the electoral votes.

These predictions tells us the likely outcome if the main source of error in polls is sampling error and if votes do not change much in the next few days.

In contrast, the betting site intrade.com puts McCain’s winning chances at around 17%.

One interpretation of the dramatically higher betting odds on McCain than in the simulated voting is that the polling data is unbiased measures on the day taken, but bettors believe votes change a lot in the last few days. That interpretation seems dubious, as the polls’s estimated vote share tends to change by about 1 percentage point a week. It would take an extraordinarily large amount of news to move the polls 6 points in 4 days.

A second interpretation of these odds is that polls are biased in an unknown direction. For example, polls rely on models of who is likely to vote that are typically based on historical patterns. In addition, most polls miss voters who use only a cell phone. The systematic noise in polls, while surely there, is probably not a good explanation for why bettors favor McCain more than polls. For example, blacks are likely to vote more than historic norms, and cell phone users probably favor Obama as well. Thus, these factors – while surely present – would not explain a high betting line on McCain.

The third interpretation is that many bettors believe a “Bradley effect” exists; that is, that many voters will say they are voting Democratic or are undecided, but in the privacy of the polling place will not vote for an African-American to be president. The evidence for a Bradley effect in recent elections is weak, but it is extremely difficult to prove it does not exist. As best I read the evidence, it appears some people are betting this form of discrimination remains strong in America.

(updated by lerxst to correct misspelling of Bradley)

1 comment:

Anonymous said...

It's called the Bradley effect...
http://en.wikipedia.org/wiki/Bradley_effect

Also... while this been discussed, ad-nauseum, in the press lately... even the some of the pollsters and campaign workers who were involved in the original "Bradley" race refute the idea that the Bradley effect is more than a very minor influence.

In this election, some polls are likely to miss the target by much greater margins simply because of the difficulties in coming up with accurate likely voter models. These aren't really biases in the traditional sense of the term, rather, just different interpretations of the limited data we have right now about who will actually turn out to vote. Thus a pollster whose likely voter model predicts a lower turnout might show McCain as running very close... and may defend that assumption with reasonable facts (although the early voting statistics are starting looking fairly persuasive to the contrary).

Finally... I wanted to point out that the Intrade markets have already been shown to be easily manipulated due to a serious lack of liquidity. Earlier in the election cycle, a bidder was identified who was throwing money away in order to rig the Intrade numbers for McCain. In terms of the media war... this might be money well spent.