I’ve resisted saying much about the presidential race this year for many reasons, primarily because this is a topical issue that will soon be decided and may have a short shelf-life. However, the daily discussions about polls provide a great opportunity to think about some statistical and methodological issues.
Let's consider one poll result. Gallup’s daily tracking poll of “likely voters” conducted October 24-26 2008 found 50% supporting Obama and 45% in support of McCain. The poll had a sample size of 2,446, which might seem like a small number to estimate how the American public will vote. But if done well, polls can be somewhat accurate, and always include something called the margin of error to account for the possible error of not including the entire population. This poll result had a sampling error of +/- 3 %, thus the expected range of voting could be between 47%- 53% for Obama and 42%-48% for McCain. As practice for statistics students, the margin of error reported is for a 95% confidence level, which means that we can be 95% confident of the ranges listed above being a true representation of American likely voters. When we subtract the differences between Obama’s lowest range and McCain’s highest, we find a difference of just 1%.
McCain responds with optimism when asked about winning this election yet, with those poll results, there is a slight chance of his success, albeit very slight. Could the polls be wrong?
Commentators have been talking about various factors that the polls might not account for. The so-called "Bradley Effect" may introduce a potential source of error in these results. Named for Tom Bradley, the late mayor of Los Angeles who ran for governor of California in 1982. He was ahead in the polls but lost the race. Observers wondered whether people may say they are voting for an African American candidate, but have a last minute change of heart once they actually fill out their ballot and end up voting for the white candidate. Others countered that the polls did not adequately account for absentee voters, who tend to be more conservative, and that polls have been wrong with white candidates as well. It is possible, however, that racism might influence some to vote for McCain.
While some are also mentioning a reverse-Bradley effect, where people don’t want to admit that they are voting for Obama to their friends and family because of his race, I would advance a few other sources of potential error: The “McCain effect” and the “landline effect”.
The McCain effect is when people say they are voting for McCain yet actually vote for Obama because they are paying more attention to their economic health yet won’t alienate their conservative friends and family. Thus perceived economic self-interest will prevail over social congruity favoring Obama. They might have even voted for McCain if not for the current financial crisis, which polls indicate are most American’s number one issue of concern.
There’s one more potential polling error. The “Landline effect” holds that polls undercount those with only cell phones and over-count those with landline phones, thus undercounting people under 35 and over-counting those over 35. Normally this might not be a major issue, since young voters historically have only voted in low numbers. But this year young people are very energized and expected to vote at a much higher rate than in past elections. According to a Reuters article, more than fifty percent of eligible voters eighteen to 29 are expected to vote, and young voters prefer Obama to McCain by a two to one margin.
Thus the effects of aging, generations, technology, and lifestyle will skew the poll results and underplay the support for Obama while exaggerating the support for McCain.
I would imagine the Bradley effect and McCain effects will cancel each other out. Yet the Landline effect might ultimately give us an inaccurate picture of what how people may vote. Time will tell if the final count ends up reflecting these dynamics or others we have not yet discovered. In any case, pollsters will need to come up with ways for addressing these and other issues in predicting future elections.