Thursday, June 7, 2012

Gallup and Rasmussen look like outliers

I have to wonder when/if the media will ever notice that Gallup and Rasmussen polls in the presidential race are starting to look like outliers.

Yesterday five national vote polls were released.

Pew: Obama +7
PPP: Obama +8
YouGov: Obama +5

And here are the outliers:

Gallup: Obama +1
Rasmussen: Tie

Now remember, the two outliers release tracking polls every day while the others tend to poll once/month or less. So that means that Gallup and Rasmussen have many more entries in the aggregates and thus skew the results.

On a less frequent basis, Rasmussen also polls in individual states. I decided to look at one state that is crucial in the election where they recently polled: Virginia. Here are the latest results there:

NBC: Obama +4
Washington Post: Obama +7
PPP: Obama +8
Quinnipac: Obama +5
Rasmussen: tie

Any questions?


  1. Nate Silver of 538 (NYT) had an interesting set of blog entries on the exact situation you are seeing in Rasmussen.

    Long story short, Ras cheats. The method of sample selection is biased. They are shills for Republicans. The effect is mostly seen in large state and is mitigated in small states (smaller populations to sample).

    Mysteriously, these numbers tend to clean themselves up in the 2 week period prior to an election.

    I wonder why?

    1. Well put, Der Farm! Many have noticed this about Ras polls.


  2. When a poll by FOX News shows the President up several points over Romney, while a Gallup/Rass poll shows a "dead heat", you know that there's some weird stuff going on.

    FOX has no reason to suck up to the President, and they've hammered him nine ways till Sunday. They are the mouthpiece for the GOP. And yet their poll shows President Obama ahead.

    What does FOX know that Gallup/Rass do not?

  3. One difference is that most pollsters poll registered voters, while rasmussen polls "likely voters", a category among whom republicans normally do better.

    The key point, then, is how good is Rasmussen's model for determining what fraction of registered voters are "likely" to actually vote. Since any such model reuires them to estimate what turn-out will be, and since Democrats generally do better the higher the turn-out is, polling likely voters requires them to pre-judge the very thing the poll is supposed to find out.