Likely Voter Polls, Registered Voter Polls and the Power of Averaging

I have quoted this post by Steve Singiser from vague recollection so many times recently that I finally looked it up. Singiser looked up polls taken within one month of election day in 2006 and 2008 which reported results both for registered voter samples and likely voter subsamples. He found that the registered voter polls performed better

"THE FINAL SCORE: RV 32, LV 21, Ties 3"

This is a stunning fact. Basically everyone but Singiser ignores registered voter polls if likely voter polls are available. I post, because I think I have a marginally useful suggestion. If results from the registered voter sample and the likely voter subsample work about equally well, then the average of the results for registered voters and likely voters will perform better than either, probably quite a bit better than either.

So I average if I can. For example, this means that I consider the latest CNN poll of Nevada to show a dead heat Angle up 4% in the likely voter sample Reid up 4% in the registered voter sample.