In recent weeks people have been debating whether or not presidential match up polls are biased in terms of the proportion of their samples which identifies as democrats, republicans, and independents. This article shows that the polls are in fact biased, that correcting for this bias cuts Clinton’s lead over Trump in half, and that looking at likely voter polling, as opposed to registered voter polling, or correcting for general pro democrat bias in polls, suggests that the race is currently too way too close to call and will come down to turnout.
The Polls are Biased Against Trump
Determining whether or not polls are biased is fairly simple. All you need to do is compare the party ID proportions of the general population, or recent voters in presidential elections, to the party ID proportions of match-up polls.
I did this with 5 recent polls by Reuters/Ipsos, Fox, Yougov, Quinnipiac, and Morning Consult. Comparing their party ID proportions to Party ID polls done by Pew and Gallup, as well as 2012 presidential NYT exit poll party ID proportions, it was clear that recent polls over sample democrats and under-sample republicans and independents.
The Reuters poll, which had Clinton winning by 10 points, is the most extreme example of this.
These 5 recent polls had Clinton winning by an average of 5.6 points. Re-weighting their samples based on 2012 voter turnout shrank that lead to 2.72 points, a reduction of more than 50%.
The Polls are Usually Biased Against Republicans
Moreover, polls are generally biased in favor of democrats, especially this far out from the election. Specifically, according to an analysis of American polling in every presidential election between 1952 and 2012 polling done around this time period normally over-predicts democrat performance by about 2 points.
Now, this might be because polls at this time always have a party ID bias that goes away later on, but I doubt it. There is no reason to think pollsters magically get better at sampling as we get closer to elections.
If this is right, then we can reasonably predict that Clinton will do about 2 points worse than polls are currently saying she will which, after correcting for party ID bias, implies she would win by 0.7 points. In other words, we might reasonably say that the polls are way to close too call and the election will come down to turnout.
It’s worth mentioning that not all poll exhibit anti republican bias. Specifically, an analysis of American presidential polls from the years 1992 to 2012 found that the median registered voter poll was biased in favor of democrats (by 2 points) but the median likely voter poll exhibited no bias at all.
Currently, the only major pollster doing likely voter polls on a regular basis is Rasmussen. Likely voter polls in general have a history of no bias, but Rasmussen in particular over-predicted Romney’s performance in 2012 by 3.7 points. So, with that being said, Trump is ahead by 4 points in the most recent Rasmussen poll. Thus, if we adjusted for their 2012 pro republican bias, we might reasonable say, once again, that the polls are way to close to call and the election will come down to turnout