Like many Americans these days, I spend lots of time checking pollster.com's polls of polls and their Electoral College map. The poll of polls, like the analogous product over at realclearpolitics.com and others, offers an aggregate of the most recent public polls.
The site's editors explained that the polling aggregate they offer is not merely an average of the most recent polls. Rather, they use a more complicated statistics analysis to determine the trend.
The pollster.com aggregate is of no use if isn't accurate, so I unscientifically analysed the site's performance in 2006. They were spot on, correctly calling 50 out of the 51 races I examined. While it is impossible to gauge how accurate today's polling is, two weeks out from election day, the results give weight to pollster.com's read of the trend.
I looked 29 gubernatorial races and 22 senate races with at least five polls for the site's algorithm to work with. Of those, Pollster.com's final aggregate was correct 50 times. They were only wrong in Minnesota's gubernatorial race, where Republican Tim Pawlenty won by one percentage point after Pollster.com had Democrat Mike Hatch winning by 2.6 percentage points. The 2006 results are available on CNN.com.
Also noteworthy, the pollster.com aggregates tended to under poll for the Democratic candidates. For example, just before Maryland's 2006 senatorial race, pollster.com gave Democrat Ben Cardin 48.4% and Republican Michael Steele 45.2%. The actual result was 54%-44% for Cardin.
My unscientific analysis demands several caveats. For one, it is impossible to tell how accurate the polling was two weeks before the election, where we are now. We can only compare the aggregates as the candidates crossed the finish line with the actual result. Also, pollster.com may have tinkered with their algorithm since 2006.
The unprecedented enthusiasm for Obama's candidacy among young and African-American voters, compounded with the expected surge in turnout, complicate sampling for the companies conducting the polls. The pollsters say they account for the changing electoral.
Lastly, it remains to be seen whether Obama will suffer from the "Bradley effect", under which African-American candidates under-perform late polling. It is a matter of debate whether the Bradley effect, named for a 1982 candidate for governor of California who lost the election after leading polls by more than 10 points, ever existed, or if it did once, whether it would affect the race this year. It was not universally apparent during the Democrats' long primary season.
The most important figures on pollster.com's page are those tracking the polls in the battleground states, not the national number, because the winner of the electoral college takes the White House, not the winner of the national popular vote. The national polls, which show McCain tightening Obama's lead in recent days, may no longer be a useful compass tracking the nation's leanings, since McCain may be pulling head in bright red states he is already certain to win. For example, McCain's leads in Nebraska, Texas and Arizona appear to be growing. Increased support there for McCain would show up on a national poll but would have no bearing on his ability to win in Ohio and Virginia.
