Monday, October 15, 2012

Comment on Singiser

Steve Singiser has a big post on likely voter polls vs registered voter polls up at dailykos (one of his favorite topics and one of mine).  I go wild with enthusiasm in comments.

I know you are busy, but I'd really like the exact words that the PPP machine says along with "unlikely" "vote" "hang" "up2 and "phone."

Given how everyone aggregates polls now, the magnitude of the error doesn't matter all that much.  I think the 50-50 favor Dems-Reps result suggests that it makes sense to look at likely voter polls.  But polls are aggregated to the average not the mean so what is the average LV poll error (note you don't need RV polls to calculate this).

LV minus RV Obama-Romney vs LV minus RV Mondale v Reagan.  Look a whole lot of the difference between voters and non voting adult is that voters are older.  I just can't doubt that the new huge LV-RV gap is due to the new huge age gradient of voting intentions with older people now much more likely to vote Republican (before a smaller difference and occasionally the elderly tilted Democratic).

LV filters might oversample the old compared to the young.  RV polls might over sample the young compared to the old.  It should be possible to compare LV filters to actual voting by looking at exit polls (maybe even actual data on votes compared to census data on the population).  This is a way in which things have changed.  It matters because pollsters have checked LV filters by looking at past forecasts minus outcomes.

One of the bees in my bonnet is the Gallup LV screen (total flop in 2010 after decades of working fine).  One question is "do you know where to go to vote".  The no answer has a very different significance on November 5  and on October 5 no ?  If it's Monday and someone claims they will go and vote on Tuesday but still doesn't know exactly where uh then maybe Gallup is right to take the claim with a grain of salt (3 of which grains gets one tossed out of the LV sample).  Another is have you voted there before which is certain to drop people who really will vote for the first time and also people who moved so strongly weeds out the young.  Notably the Gallup LV screen worked very well for the Congressional Generic ballot for decades then caused a total utterly humiliating forecast error of 9 (nine !) percent in 2010.

Average error of LV polls can be calculated for LV polls long before election day.  Yes there is the real shift in public opinion so one wants to average over many elections, but what we want to know *now* is whether polls released now are biased not whether the last poll before the election will be biased.

No comments: