More Drama from the Polls: What Have We Really Learned?

Remember the 2012 election? Remember how the Presidential election was going to be so close it was supposed to go all night? Remember how the Republicans were poised to pick up Senate seats or even take the Senate? Except, remember how that didn't happen? In spite of all the drama leading to the 2012 election, Obama won easily, the Republicans lost seats in Congress and most of us went to bed early.

So, how did we get fooled? Well, a lot of polls said a lot of races would be really close. Except they weren't. True, certain polling gurus like Nate Silver were never fooled by all those polling errors, but many others (including Mitt Romney) were caught completely off guard.

Now, two years later, what can we learn from that mess? Could someone who isn't Nate Silver use past polling errors to make predictions for this next election? Well, let's start by looking at the old polling data to see what might have gone wrong.

A good place to go for all national polling data, past and present, is the website Real Clear Politics. This site posts polls for every major National and Statewide election going back to 2004. Rather than trying to weigh or analyze polls the way 538 does, It calculates a simple numerical average of all the recent polls for a given race (the RCP average). This average worked pretty well for most of the Senate elections in 2006 and 2008, and very well for the 2008 Presidential election.

But in 2010, something different happened. In the few weeks immediately before the election, the polling data indicated seven senate races where the RCP average showed that neither candidate had more than a five point lead over the other. Races this close are considered "toss ups" by Real Clear Politics. Come election day, the Democrats carried five of those elections. The was somewhat surprising, considering the Republicans were slightly favored to win four of them, but it wasn't ridiculously unlikely.

But here's the really strange part: The Democrat outperformed the polling average in every single race -- seven out of seven times! Really outperformed them -- in "toss up" races in 2010, Democrats beat the polls by an average of 4.4 percent. And in the one Senate race that was in the "leans Dem" category (where the Democrat is favored by between 5 and 10 percent) the Democrat also outperformed the polling prediction by just over 3 percent.

Still, let's not jump to conclusions. There were a lot of races in 2010 and trying to draw any conclusion from just seven of these races is what statisticians would call drawing a bulls eye around a cluster of arrows. Now let's look at what happened in 2012.

In 2012, the year we were expecting all the drama including Republican gains in the Senate, things really fizzled on election night. How did the polls get it so wrong that year? Using Real Clear Politics you can see that just prior to the 2012 election there were again seven Senate races in the "toss up" category. And once again, the Democrat in each race outperformed the polling average for their race, this time by an average of 4 percent.

That night the Democrats won six of the seven Senate elections that were considered a statistical tie by polling averages. What's more, in the five races considered "Leaning Democrat" (where the Democrat polled an average of 5-10 percent better than the Republican in the race) Democrats outperformed the polling averages by an average of 5.5 percent.

In 14 "toss up" races, you'd expect Republicans to do a little better than predicted in about half (7). You might believe 5 or even 4, but the number was zero! That's pushing pretty far past the laws of statistical likelihood. Now throw in the "leans Democratic" races, and we have 20 races where the Republicans outperformed polls in zero of them. The Democrats beat the point spread in 20 of 20 statistically close senate races. And they beat it by a pretty significant average of 4.5 percent.

And we're not done. There were four Governor's races that were considered either toss ups or leans Dem in 2012. In all four races the Democrats outperformed the polling averages for their race by an average of 4.4 percent. Now, there was no clear bias in the 2010 polling for Governor's races but, when added to the Senate polling data, this does look suspicious.

And then there's the Presidential election. In 2012 the polling average predicted Obama would only win by only 0.7 percent. Instead, he won by a very comfortable margin of 3.9 percent (about 5 million votes). Guesses are supposed to get better when the numbers get bigger, so missing the "point spread" by 4 million votes is a big goof.

That's 25 out of 25 elections in which Democrats outperformed the polls. In every case it was for a race that was considered a statistical tie or the Democrat was slightly favored (I found no clear bias for races in the other categories). The odds of this happening purely by chance is close to a million to one!

I can't say why this is happening, but the overwhelming numbers make it look a lot like it's built in to the system somehow. Maybe these errors are due to political bias. Maybe the media encourages polls that make an election cycle appear more exciting than it really is. Or, maybe it's just polling incompetence.

Whatever the reason, this year we're again being told that the Republicans are about to retake the Senate. Also, as I'm writing this blog, there are nine Senate races listed as toss ups on Real Clear Politics. Now, I don't pretend to know exactly what's going to happen in November, but if history is any guide (and it often is), expect Democrats to again do a lot better than people are predicting, especially in close Senate races.