Saturday, September 22, 2012

Kevin Drum Link Bait




Yesterday Kevin Drum asked whether congressional generic ballot overstate the Democrats performance on average (he had the sense that they do by about 4 or 5%).  I did some arithmetic in a comment.  It seems he didn't read the comment and did some math of his own and quoted some from Sam Wang.  There is pretty general agreement that this used to be true, but ceased roughly in the past two elections.

Having averaged a total of about 12 numbers by the sweat of my brow I post my calculations here hoping for more attention (I deleted some mild churlishness).

The evidence is available on the web.
For comparison Real Clear Politics current generic  is Democrats lead by  2.2 %

Real Clear Politics 2010 outcome R + 6.8
Last RCP average (late October)     R+ 9.4 difference  R -2.6 
Polls including Sept 21 2010 (narrow window as I am doing it in my head and lazy) R+5
difference R+1.8
(latest RCP corrected with past RCP forecast  error 2.2 - 1.8  = 0.4 so fits your conclusion that Democrats not really ahead).

Just eyeballing  Sept 2010 seems to average to a strong Republican lead.

2008 Final D + 10.7 Last RCP average D + 9 Difference R  -1.7 
polls with samples including Sept 21 2008 D+  9.8 Difference R -0.9 
2006 actual outcome D + 7.9
RCP last average        D +  11.5 difference R + 3.6 
ohhhhh that's it.  The rule worked for last polls in 2006 

There no congressional generic ballot polls included Sept 21 2006, but 4 closest average to D + 11. Again fits Kevin Drums vague recollection.  None of the 2006 polls I averaged from Rasmussen.   
2004 actual result R + 2.6 final RCP average (late October) tied.  

update: RCP average of just 3 polls with 9/21/04 in their sample D + 0.33 so difference R + 2.9 
The reason I went to the huge effort of averaging 3 numbers is that maybe things are different in presidential years and mid terms.  Hmmm average for last two presidential years is R + 1.
2002 result R 4.6 last RCP average R 1.7
I'm too lazy to do hand averaging for September 2002 and 2004.It seems Kevin Drum's rule of thumb used to work, but didn't work so well in the past two elections.  My guess is that this is due to Rasmussen singlepollsterly cancelling the bias in the average of other polls (they had a  - 3.8% bias in 2010 and published Lotsssss of polls).my average of average forecast errors of polls taken around Sept 21 is R + 1.2 % (but obviously it would be better to average more polls its just arithmetic is boring).
(here R+1.2% is outcome minus poll. 

update comment:  More generally it could be the cell phone only household issue.  Many pollsters, including Rasmussen, are automated robopollsters (phone rings and you are talking to a computer). They can't call cell phones (by law).   This has added a bias which seems to have roughly cancelled the old bias.  In update above I averaged three numbers related to 2004.  The idea is that forecasting who will actually vote is very different for presidential years and midterms.

No comments: