The New York Times’ Upshot published an article on their latest New York Times/CBS News/ YouGov poll which highlighted work that I have done with Justin Wolfers of the University of Michigan on expectation polling. The expectation question asks, “Regardless of how you are voting, which candidate do you think is most likely to be elected?” Historically, this question has been extremely effective at pointing towards the eventual winner of the election and, even identifying the vote share. (For more background, you can read the paper or this Q&A with Justin in 2012.)

When it comes to the 2014 senatorial elections the most interesting differentials between the expectation polling with traditional intention polling is the sizable lead in the expectation polling for the Republican incumbent Roberts in Kansas and Republican Perdue in the open Georgia seat. Both of these are dead-heats with the litany of quantitative forecasters (Upshot, FiveThirtyEight, Huffington Posts’ Pollster, etc.) whose forecasts are mainly driven from traditional polling. And, the traditional intention question in this YouGov poll also has them within the margin of error. But, the Republicans have commanding leads in the expectation question.

The below chart shows the percent of Democratic and Republican supporters that expect the candidate from the other party to win the election. I count just the supporters that expected one of the two major candidates to win the election (i.e., I disregard those people who respond that they do not know who will win). I plotted them from left to right depending on the expected vote share of the Republican candidate on PredictWise.

Note: there is no Democratic candidate in Kansas, but rather an independent running against a Republican.

Despite millions of dollars of polls (each poll in Pollster and Real Clear Politics’ lists costs tens of thousands of dollars) and millions of lines of historical data feeding into PredictWise’s algorithms, this single poll’s expectation polling breaks perfectly on the 50% line. For every election that PredictWise expects the Republican to win, a higher percentage of Democrats cross over to the Republican candidate. Similarly, for every election that PredictWise expects the Democrat to win, a higher percentage of Republicans cross over to the Democratic candidate.

There is strong upward slope depending on how convincing PredictWise expects the election to be, but all of the identification comes from the opposite party support. The percentage of Republicans that expect the Republican candidate to win is not too interesting once the Republican has a commanding lead (right side of the chart). But, the left side of the chart has a lot of identification about the expected vote share. While it is not a strictly monotonic relationship, generally, the higher percentage of the vote share the Democratic candidate is likely to receive, the more Republicans cross over and think the Democratic candidate will win.

We know that there is a lot of information going into people’s expectations: their own voting intention, the voting intention of their social network, and what they are hearing from the media. Further, this interplay of information varies by demographics. Yet, what matters for forecasting is not the exact information that goes into each forecast, but that it is meaningful information and that its relationship with the outcome is stable.

The data from expectation polling is so powerful, that it is the equivalent of the respondent going out and asking 10 random likely voters who they will vote for, including their own vote, and then telling the pollster who won his/her private poll (of 10 random likely voters and themselves). We do not think that that actually is what people do when they answer the poll, but that is how powerful the poll is for forecasting.

We have no problem that 30-50 percent of partisans think their candidate is going to win in landslide losses; actually it makes perfect sense! In our model that percentage perfectly mirrors the impact of the supporter including him/herself in their poll. A Republican supporter starts his/her personal poll with one definite Republican supporter. We know in reality it is also driven by some wishful thinking, but that is ok, because, this relationship between expectations and outcomes is stable through dozens of election cycles, with varying degrees of media coverage.

So, PredictWise, with its millions of dollars of inputs and millions of lines of data is saying a toss-up in Georgia and Kansas. But, the expectation of the local voters, from just one poll, is a mighty powerful data point. Thus, I expect Georgia and Kansas will likely go Republican.

Updating Predictions: senatorial, senatorial balance of power, and gubernatorial.