We're now getting polls left, right (a lot of those) and center. In my last post I showed that this is The Year of the Crappy Pollster. But in order to document their crappiness, we need to know a little more about how polling error behaves. Today we look at the question, Why is there so much scatter among the polls, even from good pollsters?
Sorting through old polling data, we can find two answers, at least, that are not directly related to the idiosyncrasies of individual pollsters. First, scatter is related to how far away an election is, and second, it is related to what type of race is being polled.
The lines on the graph below represent how much scatter there is among the polls, on average, for four different Florida races, after taking into account changing voter preferences.
For all four races, the scatter is generally worse for polls taken about 100 days before the election than those taken 10 days before the election. Additionally, the Senate races clearly have more error—more scatter about the polling average—than the presidential race. The primary race appears to have more scatter than either Senate race as well.
In other words, different types of elections have different amounts of scatter, but polls converge as the election approaches, even after you account for changing voter preferences.
For those who want it, there's lots more detail below the fold, as well as a discussion of what makes races more volatile.