Princeton Election Consortium

A first draft of electoral history. Since 2004

High accuracy of aggregated Democratic polls

March 28th, 2016, 3:00pm by Sam Wang


Are Democratic primary polls accurate? In individual states they are good, with a few exceptions. At an aggregated level, they are remarkably accurate. The delegate-weighted polling margin has a total error of 3.1%. This is better accuracy than one would expect from the reputation of polling these days. The overall error is also not nearly enough to change my calculation from yesterday.

Above are Election Day results so far this season, plotted against poll medians taken shortly before the election. The correlation coefficient is +0.85, pretty good. Red symbols indicate caucus states.

The notable exceptions are Minnesota, Utah, and Idaho. These are all caucus states. However, they also have a high fraction of non-Hispanic whites (see Will Jordan’s graphs showing this here and here). Perhaps the anomalies are accounted for by some aspect of the combination: highly committed Sanders supporters and deliberating in a crammed room. As an aside, it seems to me that the caucus mechanism requires either cultural homogeneity or a willingness to cross cultural lines to gather in groups. If caucuses date from a previous time, is segregation the historical antecedent that gave rise to a caucusing/race correlation?

The most important quantity for prediction is the accuracy of all the polls aggregated. The delegate-weighted sum of polls is Clinton +17.9%. The delegate-weighted voting result is Clinton +14.8%. The difference, 3 percentage points in Sanders’s direction, is greater than the ~2% that one would expect from sampling error alone. However, it does not come anywhere close to the 22-point swing that I calculated is necessary for Sanders to get an overall majority of pledged delegates by June 7th.

Tags: 2016 Election · President

No Comments so far ↓

Like gas stations in rural Texas after 10 pm, comments are closed.