Thursday, August 25, 2016

These polls are closer to the truth: Gravis, PPP, Rasmussen, YouGov, UPI/CVOTER, LA Times/USC, and Zogby

The Huffington Post has a useful interactive website where you can compare poll results by several factors. I wanted to look a little closer to see if polls tend to be biased against Trump because some supporters are shy about admitting that they plan to vote for him, and that the methods fail to fully capture the enthusiasm for him.

What you see below is the average Hillary lead by poll characteristics:

Partisan/Registered Voters/Live Phone  11.0

Non-Partisan/Registered Voters/Live Phone  6.2
Non-Partisan/Likely Voters/Live Phone  7.3
Non-Partisan/Registered Voters/Non-Live Methods  5.9
Non-Partisan/Likely Voters/Non-Live Methods  4.8

You can see that partisan polls REALLY favor Hillary. Non-partisan polls that are otherwise the same produce a gap that is 4.8 points smaller (11.0 vs. 6.2 points).

Likely voter polls do not necessarily yield smaller gaps than registered polls, so there is no clear evidence here that registered voter polls are failing to capture the greater support for Trump. On the other hand, it is possible that there are more Hillary supporters who say they are likely to vote but won't actually do it because they can't get excited about her.

Finally, there is clear evidence that the gap shrinks when private methods are used so participants don't have to tell real humans their preferences. The gap shrinks .3 points for registered voter polls and 2.5 points for likely voter polls.

So, the most reliable polls use non-live methods. On the other hand, registered voters polls are not necessarily inferior at this point in the race. Still, it makes sense to give preference to likely voter polls since many registered voters will not actually vote. Based on this, the surveys to place the most trust in (at least the ones that come out frequently) are: Gravis, PPP, Rasmussen, YouGov, UPI/CVOTER, LA Times/USC, and Zogby.  Reuters should also be on the list, but my confidence in them has been shaken. The results are sometimes really weird, and of course, we've seen reporting that suggests they monkey with the data. The good polls indicate Trump is down 4 or 5 points.

No comments:

Post a Comment

Is skin tone correlated with job prestige?

The General Social Survey rated the skin darkness of a sample of black Americans, ranging from "very dark brown" to "very lig...