Hallo again, A couple of days ago I wrote a message concerning the polls for the presidential campaign that are currently being published on a daily basis, and especially the usual +-4% error and how it is computed. I received three replies pointing out some of the difficulties such as the fact that these estimates are based on stratified samples rather than simple random sampling. A very nice discussion of the issues involved can be found at http://slate.msn.com/framegame/entries/00-10-26_92147.asp . An explanation of the methodology of the Gallup tracking poll is at http://www.gallup.com/poll/faq/faq000101.asp , where they say that the +-4% is for a 95% confidence interval. I would like to pose another question, though, namely what a reader should make of all these polls. There is of course the additional problem of the horse race question: if Bush's percentage is estimated at 46%+-4% and Gore gets 44%+-4%, than in fact Bush is 2% ahead of Gore, but this difference has an error of +-8% and so is clearly not significant. Now if a poll would put one of them at 49% and the other at 41%, then the difference would be 8% and that would be borderline statistically significant at the 5% level. But then there are all these other differences between polls (I guess we might call them systematic errors) which probably mean the +-8% error on the difference in percentages is in fact too small. So here is my question: how large a lead of one candidate over the other should really be taken seriously? I am sure there is no definitive answer to this question, but I would like to know your "gut feeling". Wolfgang > Wolfgang Rolke wrote: > > > Today the following polling results were given on cnn.com for the > > Presidential race: > > > > CNN/USA TODAY/GALLUP POLL > > October 20-22 > > Likely Voters' Choice for President > > Bush 46% > > Gore 44% > > Nader 4% > > Buchanan 1% > > Sampling error: +/-4% pts > > Sample size: 769 > > > > For details go to > > http://www.cnn.com/2000/ALLPOLITICS/stories/10/23/tracking.poll/index.html > > > > I am wondering how they find the Sampling error of +/-4% pts. The usual > > estimate for the standard error of a binomial would be (for Bush) > > > > SQRT(0.46*(1-0.46)/769) = 0.01797 > > > > The error in a 95% CI would then be 1.96*0.01797 = 3.5% > > > > and in a 99% CI it would be 2.576*0.01797 = 4.6% > > > > Am I doing something wrong here? > > > > Thanks > > > > Wolfgang > > "Drake R. Bradley" wrote: > Wolfgang, > > They are probably using stratified sampling methods, and that reduces the > margin or error somewhat by eliminating sampling error associated with over- > or under-sampling certain groups (which occurs with simple random samples). I > don't recall the formula used for computer the SE in this case, but I have > seen one in a methods textbook used by sociologists. > > Regards, > > Drake Bradley > Dept. of Psychology > Bates College Donald Burrill wrote > Not for a simple random sample (SRS); but the Gallup sample is not a >SRS, and the sampling error is slightly larger than it would be for a >SRS. (And it's reported only to the nearest integral number of >percentage points, so the correct value could be anywhere between 3.5% >and 4.5%, presumably.) Also, I wouldn't be surprised to find >(although I don't _know_ this) that it was calculated on the basis of a >population proportion of 0.5, and is therefore slightly conservative for >most results and _highly_ conservative for results like 4% and 1%. Bob Hayden wrote >I don't have a complete answer, but it does look like they are >rounding to the nearest whole percentage, in which case their $% might >agree with your 3.5%. In addition, the formula you quote is for >simple random sampling, which is rarely used by polling agencies.
begin:vcard n:Rolke;Wolfgang x-mozilla-html:FALSE org:University of Puerto Rico - Mayaguez adr:;;;;;; version:2.1 email;internet:[EMAIL PROTECTED] title:Associate Professor note:Web Sites: http://home.coqui.net/wolfgang http://math.upr.clu.edu/~wrolke x-mozilla-cpt:;15760 fn:Wolfgang Rolke end:vcard