Don’t Trust Those Early Polls

An Arkansas poll shows incumbent Democratic Senator Mark Pryor with a 51-40 lead over his Republican challenger, Representative Tom Cotton.

Be very careful about overinterpreting these results.

First, it’s still very early in the cycle. Senate forecasting modelers agree: Polls are weak signals early in the election year, but by October a good polling average is all that’s needed to predict winners. Indeed, some of the models don’t even include polling yet. That’s in part because, at this early stage, it’s very possible that recent news (or recent ad buys) might have a lopsided effect in favor of one candidate, and that could dramatically affect what people tell pollsters. For now, only a tiny fraction of Senate voters have focused on the contest, making them likely to over-respond to recent information that may not be similar to what they will hear in the weeks just before they vote.

Second, we don’t have many polls this early in the cycle, and even a professionally conducted survey will be off a few points, and some will be off by more. The cure is polling averaging, but that doesn’t work as well with just a few surveys. Still, HuffPollster has Pryor with a more modest 4 percentage-point lead. The odds are that’s more accurate than the 11-point margin in the Marist/NBC poll.

It’s also possible that the new numbers are accurate, and that Marist/NBC picked up a shift that gets missed when older polls are included.

In any case, don’t overreact to single polls; use the polling averages. Don’t put too much stock in surveys months before an election. On the flip side, don’t waste a lot of effort taking apart this or that polling result; just toss it all into the polling averages.

This is a very good number for Pryor, but it tells us nothing conclusive about what will happen in November, or even what is happening now.

[Cross-posted at Bloomberg View]

Jonathan Bernstein

Jonathan Bernstein is a political scientist who writes about American politics, especially the presidency, Congress, parties, and elections.