Jed was reading crop circles... On Wednesday, November 7, 2012, Jed Rothwell wrote:
> OrionWorks - Steven V Johnson <[email protected]<javascript:_e({}, > 'cvml', '[email protected]');> > > wrote: > > Well, Jed, you predicted final count would be 303. >> >> U wuz right. >> > > > Plus I said 2% of the popular vote, which was right on the nose. > > Obama will probably take FL so I was too conservative. But FL is amazingly > close. You might as well say both sides won it. > > I can't really take credit. I was just picking the pollsters I trust most, > and Nate Silver, who sure knows a lot about statistics and s/n ratios. > > Gallup veered far from the other pollsters, but in the end fell almost > back into line. It was almost right. You have to hand it to those > people. As I suspected, their definition of "likely voter" (LV) was a > little too conservative. See: > > http://www.gallup.com/poll/election.aspx > > Their LV is off by -3%; their registered voter number is off by +1%. The > LV polling ended on Nov. 4. The gap was closing. If they had continued to > Nov. 6 the gap might have been within 1% instead of 3%. > > I think their main mistake was to underestimate the youth vote and > the Hispanic vote. > > It is uncanny how good modern polling has become. It is scary. The truth > is, there were practically no surprises in this election for people who > understand polling, statistics, margin of error, sample size and other > issues. > > You should try to understand these issues. They are important in cold > fusion and other experimental science. > > Rasmussen is controversial. I don't think they are so bad. Their number > are reliable if you add +3% to the Democratic side. In other words, they > have a fixed bias. This means they have good methodology for collecting > data, with a proper random set of respondents, a good set of questions, and > a large enough sample. But they introduce a bias in post-interview > processing. This seems clear to me when I read the the actual questions > they asked during the telephone interviews, and the processing methodology > they describe in their literature. The questions they asked during > telephone interviews and procedures seem well-designed. > > Every poster has to have some degree of postprocessing or the answers will > be meaningless. For one thing, you have to adjust your responses to fit the > population. For example, if you poll people at random and reach only ~5% > Hispanics, you have to weigh their responses to represent ~10% of the > likely voter (LV) population. Overly conservative posters put them at 8% > instead of 10%, because they assumed Hispanic turnover would be lower than > it was. This was a judgement call. This -- plus random variation -- is why > there are differences between poll estimates. > > - Jed > >
