Sunday, September 18, 2016

Strangeness in the polls

I'm about to break with my standard advice.  Do as I say, not as I do.  I always advise everyone to look at polling averages, not individual polls.  Why?  Individual polls have strange biases, and we don't know what they are.  In aggregate, they tend to cancel out.  Every once in a while, though, it is useful to look underneath the average, and point out the strangeness.

So let's do that.  RealClearPolitics shows a Clinton nationwide lead right now of .9 percentage points.  Basically, a tie.  Now, look under the surface.  What the hell is going on with the LA Times?  They are the ones showing Trump leads.  Why?  We don't know.  Here is where polling methodologies matter.

1)  Sampling.  The LA Times could have a different sampling methodology that is more likely to pick up Trump supporters for whatever reason.

2)  Likely voter screens.  They might have a different method of figuring out who is more likely to vote, and that method may be more likely to include Trump supporters, for whatever reason.

3)  Sample weights.  Nobody ever gets a truly random sample, and organizations correct for that with sample weights.  The LA Times' method might be giving more weight to the Trump supporters, for whatever reason.

Those are the main reasons.  Are they right?  Who knows?  That's why we look at polling averages.  My point is that it is weird.  Note, though, that Trump has also been gaining in swing states, and those polls are not being done by the LA Times, so we can't write this whole thing off as just a fluke of the LA Times and their weird methodology, but it is worth noting that the situation is not clear.

Betting markets are still favoring Clinton, but the odds are narrowing, as well they should.  Just because the LA Times is the outlier here doesn't mean they are wrong, and the state-by-state polls by other organizations back them up.  See, for example, these numbers in Ohio.

State-by-state polls are harder to do, but we really should be looking at them...


  1. The LATimes/USC Dornsife poll has been odd the whole year.

    It's an online tracking panel poll. The conventional wisdom is that they got a sample that skews a few points R. (Where R stands for either Republican or Racist or both).

    Still good for watching trends, but I'd expect election results to be on the left of them.

    1. There is no a priori reason to expect an on-line poll to skew Republican. The fact that their sample is consistently more Republican than everyone else by a country mile almost certainly means that the real number is more Clinton-friendly, but my point is that the state-by-state polls look more like the LA Times than the rest of the national polls, and that's a mystery. You can't pin that on a weird LA Time on-line sample.