Another update from AJ Strata:
Statistics Is Not Simple Math, And Opinion Polls Are Statistically Dodgy
**************************EXCERPT INTRO****************************
Published by AJStrata at 8:36 pm Oct 28, 2008
When NASA or the DoD need to track a satellite we use multiple parallel models to account for gravity, solar pressure, the height and drag of the atmosphere (what little there is at orbital altitudes) and a variety of other models. We update the measurements in all these factors and run statistical models, weighting newer data over older data, all the while checking for the random outlier measurement. Even with all this we can only predict solid orbits for about 7 days before we have to update them in many cases.
And that is with a well characterized and physics bases system to model. Oh if opinion polls could be so simple!
Being a bit of a math addict I sometimes forget that pollsters are not really as versed in scientific sampling as some engineers and scientists, though the case could be made they are better than the scientists who modeled ‘global warming’ and CO2 levels. I found two commentaries very interesting and revealing.
One is from Jay Cost at RCP
which illustrates why the polls are a problem, but interestingly Jay cannot find the answer to the conundrum:
On Friday, I noted that the differences among the national polls is large enough to suspect that something other than random variation is causing the disagreements.
…
It is unlikely that random variation would produce these effects. Today’s Rasmussen poll shows McCain significantly higher than the RCP average, and it has consistently been higher than the RCP average for the last three weeks. IBD/TIPP frequently pegs Obama’s number significantly lower than the RCP average, and it has shown him lower than the RCP average every day since it began. The GWU/Battleground poll has shown McCain consistently higher than the RCP average for 10 of the last 10 release dates, frequently at significant levels.
As I noted in my posts on the Bi-Poller world of polls (here and here) if one groups polls by how they treat party affiliation in their final results into two categories we find two distinct families of polls with consistent results. Either pollster use the historical party turnout models which would give Democrats a maximum advantage of 4% over the GOP side (’traditional’ polls to steal the Gallup vernacular), or pollster assume Democrats will experience a never seen in modern times advantage of 8-12% in voter turnout (the ‘extended’ polls).
Once you realize we have to opposing views on turnout expectations the polls are quite consistent within their respective group - and the RCP ‘average’ falls between the two groups (showing either higher or lower expectations for Obama and the Dems). Both assumption cannot be true - one model will be proven right next week, and the odds are stacked against the ‘extended’ turnout models.